I spent this week at Deloitte University, my old stomping grounds, at a conference on technology sourcing and global capability centers, in a room of people who decide for a living where work gets done and who does it. Tokenomics had its own slot on the agenda, and before that session I had never heard the word. Deloitte has been publishing research on it since January. Half the room nodded like it was old news. I made a note to look it up more deeply later.

Here is the plain version. A token is the unit AI providers bill by: a small chunk of text, a little less than a word. About 100 tokens make up 75 words of English. Inference is the model doing its job, answering a question, writing code, summarizing a document, as opposed to training, when it first learns. Tokenomics is managing what all those tokens cost, and Deloitte argues companies should track that spending as carefully as any major capital investment.

The cost of a token keeps dropping. Epoch AI found that the price of a given level of AI performance has fallen about 13-fold a year since 2023, faster than DNA sequencing, computer chips, or lithium batteries ever got cheaper. In January 2025, OpenAI’s o3 model scored 75 percent on a PhD-level science test at 30 cents a question. Eighteen months later, a newer model matched that score for four-hundredths of a cent: 725 times cheaper.

Enterprise AI bills did not fall 725-fold. Many of them did not fall at all. Lan Guan, Accenture’s chief AI and data officer, put it to Fortune in July:

“Clients are ready to scale AI, but then they hit this unexpected cost wall.”

Deloitte describes a healthcare company hit with more than $6 million in unbudgeted annual costs after its token use grew 8 to 10 percent a month for six months. AT&T, Deloitte reports, went from 8 billion tokens a day to 27 billion after rolling out AI agents, software that completes multistep tasks on its own. Uber spent its entire 2026 AI coding budget in four months, and its chief operating officer, Andrew Macdonald, told Fortune he could not connect that spending to more useful features shipped.

Jevons Paradox explains the gap. In 1865, the economist William Stanley Jevons noticed that as steam engines burned coal more efficiently, Britain burned more coal overall. Cheap AI works the same way. Companies ran more queries, built more agents, and put the tools in more hands, and use grew faster than prices fell. Google’s monthly token count rose from 9.7 trillion in 2023 to more than 3.2 quadrillion now, a 330-fold jump. Goldman Sachs expects total AI-related spending to top $800 billion this year.

That gap turns a purchasing question into a workforce question. Accenture estimates that only 10 to 20 percent of business tasks need the most advanced and expensive models. So every company now has to sort its work: which tasks earn the expensive model, which can run on a cheaper one, and which still need a person who answers for the result. I sort work like this every day. I lead workforce strategy for Technology at Capital One, including its Non-Associate Labor program, which governs the strategy which work goes to employees and which goes to contractors. AI adds a third option to that decision: the model.

Global capability centers, the hubs where companies run finance, IT, and engineering work in cities like Bangalore, Krakow, and Manilla, show the change first. GCC Base reports that work once done by ten people by hand now runs with three people supervising AI agents. Those three do a different job: checking the agents’ output and making the judgment calls. The jobs shrinking fastest are manual quality testing, entry-level tech support, and routine data processing. The jobs growing are engineers who run AI platforms, specialists who check AI systems for errors and risk, and people who design the workflows agents follow.

BCG found in September that companies that redesigned jobs around AI beat their peers’ annual shareholder returns by more than 11 percentage points. Reserv, an insurance technology company, cut the time to build a training program from eighteen months to four weeks, and its response time to regulators from three days to one hour. BCG says the companies pulling ahead are creating new kinds of jobs: people who capture what experts know so an agent can use it, and people who redesign workflows end to end, with a person still answering for the result.

Before Capital One, I led Deloitte Global’s talent strategy. In both jobs the question has been the same: which work goes to a global capability center or international talent hub and where, which goes to a contractor, and which stays close to the corporate center. GCCs now face that question with AI in the mix, and their numbers point one direction. Business of GCC values the market at $601 billion, headed for $886 billion by 2030. Ninety-two percent of GCC leaders say their centers now do far more than cut labor costs, and 83 percent are expanding their use of generative AI. If GCCs existed only to be cheap, cheaper AI would be shrinking them. Instead, McKinsey found that 84 percent of companies plan to expand the scope of their shared-services operations within two years. Companies are moving judgment work into the hubs.

McKinsey also found that 88 percent of companies experiment with AI but only 19 percent see a meaningful bottom-line gain. One in six has no executive who owns AI decisions, and only 14 percent plan their workforce flexibly; most do it once a year or less often. The token decisions, which work gets the expensive model, the cheap one, or a person, are being made mostly by finance and engineering. Workforce leaders are rarely in the room. A CFO who rules that a task is not worth the top-tier model has just made an org design decision without calling it one.

Here’s how you take action

If you lead people or design organizations, do not wait for Finance to hand you the org chart it just drew by accident.

Ask for a seat where token budgets get set. Finance and Tech are deciding which work to automate without workforce strategy weighing in on the plan.

Sort your work before someone sorts it for you. For each task, decide whether it needs the top-tier model, a cheaper one, or a person you can stand behind in front of a client or a regulator.

Check your GCC plan. If it still reads like a plan to do the same work somewhere cheaper, it is out of date.

Christina Lexa writes Workforce Rewired, on the intersection of workforce transformation, AI, and global talent.

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

By the Numbers

• 11 million: US workers who may need to move to a wholly new occupation by 2035, in a range of 6 million to 16 million depending on the pace of automation (McKinsey Global Institute)

• 770,000: occupational switches a year, 3.6 times the historical average of 215,000 (McKinsey Global Institute)

• 14%, 41%, 45%: the shares of transitions that follow a direct, winding, or unpaved pathway (McKinsey Global Institute)

• 7.6 times: how much more likely lower-wage workers are than higher-wage workers to need a wholly new occupation (McKinsey Global Institute)

Reskilling and Education

McKinsey maps 11 million occupation switches, and 45% of the routes are unpaved

The report, “Workforce in motion: Skills and pathways to future jobs in the United States,” projects roughly 770,000 occupational switches a year through 2035, against a historical 215,000. More than 75% of the workers who would switch are in office and administrative support, retail and sales, or transportation and logistics, and about a third are customer service representatives, retail sales associates, office assistants, cashiers, and warehouse workers. Growing jobs cluster in the top two wage quintiles, and declining jobs cluster in the bottom two.

McKinsey sorts the transitions into three kinds. A direct pathway (14%) reaches a growing occupation with limited retraining and no loss of income. A winding pathway (41%) requires new skills or a pay cut. An unpaved pathway (45%) is “strewn with large skills gaps, lower wages, or additional required credentials.” Lower-wage workers are 7.6 times more likely than higher-wage workers to need a wholly new occupation, and 10% of bottom-quintile workers have a direct pathway, against 40% in the top quintile. Credentials gate the destination: 85% of growing occupations require one.

McKinsey Global Institute, María Jesús Ramírez, Kweilin Ellingrud, Tanguy Catlin, Diego Castresana, and Anna Kortis, September 29, 2026. Read the report

Why it matters: Employers filling the 40 million new jobs will draw from a pool where 45% of the transitions cross unpaved ground and where a credential gates most of the growth. A company that drops a degree requirement the work does not need, or pays for the retraining, shortens a path that is otherwise winding or unpaved. HR leaders can start by listing which of their growing roles still ask for credentials the job does not use.

What Workforce Leaders Are Watching

• Which of your open roles list a degree or certification that the daily work does not use, and what would it cost to replace that line with a skills test?

• Your customer service, cashier, office assistant, and warehouse teams sit in the occupations McKinsey says account for a third of the switches. Who plans each person’s next role before the current one is cut?

• A winding pathway asks the worker to absorb new skills or a pay cut. Which of your programs pays part of the income gap while someone retrains?

• The national switch rate could reach 3.6 times its historical level. Does your internal mobility rate have room to rise by that much, or do displaced roles leave the company?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

By the Numbers

• 197,000: initial jobless claims for the week ending September 19, the lowest reading since mid-July

• 7.1%: the unemployment rate among recent computer science and computer engineering graduates, against 362,000 people who earned those degrees in 2024

• 8.4%: the drop in computer and information science enrollment at four-year colleges this spring

Layoffs and Company Decisions

Jobless claims hold near a decade low, layoffs still rare

Initial claims for unemployment insurance fell to 197,000 in the week ending September 19, down 1,000 from the prior week and the lowest reading since mid-July. The four-week average sits at 202,250, and continuing claims held near 1.72 million. Layoffs remain historically rare by every measure in the release, and nothing in the weekly data points to a broad AI-driven wave of job cuts.

Bloomberg, published September 24, 2026. Read the report.

Why it matters: The aggregate labor market keeps reading calm, which is real information and also an incomplete picture. The same week the claims number stayed flat, one company restructured its top ranks around AI and one entire graduating class ran into a hiring wall the claims data does not show.

KPMG folds AI leadership into one seat that reports straight to the CEO

KPMG is consolidating its AI, innovation, and ecosystem functions into a new Client Technology and Innovation group, effective October 1. Todd Lohr, a 15-year KPMG veteran, becomes vice chair of the group and reports directly to CEO Tim Walsh. Lohr told Fortune’s Nick Lichtenberg the firm needed “a singular leader that reports directly to the CEO” to drive the change it wants. Steve Chase, the vice chair who built KPMG’s original AI group in 2023 and held the firm’s global AI mandate, retires September 30, the same week the new structure takes effect.

Fortune, Nick Lichtenberg, published September 22, 2026. Read the report.

Why it matters: Folding AI, innovation, and ecosystem work into one executive seat bets that speed comes from a single owner rather than a committee. It is also a bet on Todd Lohr personally, replacing a leader who built the function from nothing three years ago.

The graduates the coding boom was supposed to reward can’t find work

Jack Zemke graduated from Tulane and is still hunting for a job. Collins Kibet came to Southern New Hampshire University from Kenya and now applies to roles that ask for mid-career experience he has not had the chance to build. Heather Hollingsworth and Nicky Forster, reporting for the Associated Press, found the same pattern nationally: unemployment among recent computer science and computer engineering graduates has reached 7.1%, even as 362,000 people graduated with those degrees in 2024, triple the number a decade earlier. Enrollment in computer and information science programs at four-year colleges is already down 8.4% this spring, ahead of a graduating class that will live out the numbers above.

Associated Press, via The Washington Post, published September 24, 2026. Read the report.

Why it matters: A decade of advice told students a computer science degree was the safest bet in higher education. That advice is failing the exact cohort who followed it, and neither career offices nor employers hiring junior technical talent have replaced it with a better one yet.

What Workforce Leaders Are Watching

• KPMG just bet its AI strategy on one executive reporting straight to the CEO. Which other firms still run AI through a committee, and what does Lohr’s mandate cost if it fails?

• Jobless claims stayed at 197,000 the same week the Associated Press found 7.1% unemployment among computer science graduates specifically. Which of your own AI-linked decisions would actually show up in claims data, and which ones only show up in who never got hired in the first place?

• Steve Chase built KPMG’s AI function from nothing in 2023 and is out by September 30. How many AI functions elsewhere are tied to one founder this way, with no succession plan on record?

• Computer science enrollment is down 8.4% before this year’s applicants have seen a full cycle of the 7.1% unemployment number. What does the entry-level pipeline look like in five years if the applicant pool keeps shrinking at the source?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Workforce Rewired has published weekly essays, periodic Prompt Logs (building in public), and frequent news roundup briefings since April 2026. That is a lot of archive to walk into cold, so this page is the map.

I write about how AI is changing the structure of work, the part the headcount forecasts miss. Roles are merging and management layers are thinning. The entry-level jobs that used to train people are shrinking while the institutions that supply talent move slower than the tools. I write for the people who make these calls for a living: workforce planners, org designers, HR and business leaders, and the educators and policymakers upstream of all of them.

Subscribe now

What arrives, and when

• An essay most Thursdays: one argument, sourced, ending with what to do about it.

• The News Roundup: the workforce and AI stories worth your time, with my read on each. If you want essays only, turn off roundup emails in your account settings.

• The Prompt Log, an occasional series on how I build this publication with AI, including the parts that go wrong (No. 03 is about catching AI writing, my own included).

Six ways into the archive

Pick the question closest to your desk. Each path runs in reading order.

What happens to managers when the org chart flattens?

Companies are widening spans of control faster than they redesign the manager’s job. The Missing Middle is the diagnosis and The Math of Flat does the arithmetic. Then read Meta’s Flattest Team Just Asked for Its Managers Back, about a company that ran the same experiment twice.

Where does the next generation learn the job?

The first professional job used to teach the profession by accident, hidden inside the grunt work. The Entry-Level Crisis lays out the pipeline problem and The Accidental Apprenticeship shows what disappears along with the grunt work. The Inverted Pyramid follows the same squeeze into India’s IT services industry, which for thirty years ran the world’s largest white-collar apprenticeship.

Who answers for what AI does?

Agents are joining the workforce faster than anyone decides who manages them. Start with Don’t Name the Robot, then Who Owns the Mistake?, where the courts start answering the question companies skipped.

Are we measuring the right skills?

The AI wage premium is real and widely misread. The Judgment Premium argues that fluency without judgment is a liability. The AI Divide Is Already Here finds the gap forming inside companies that bought everyone the same tools, and Amazon Built a Leaderboard for AI Use asks what a number has to survive before it reaches a promotion committee.

Who will do the work, and where?

The GCC Question starts with offshore capability centers. How Governments Are Driving Who Wins With AI Talent carries the argument to AI researchers, and The Smartest Industry in the World Can’t Find an Electrician ends where the bottleneck carries a tool belt.

How should workforce planning work now?

Tools arrive in quarters and capability takes years. The Wrong Question About AI and Jobs was the first essay here and still frames the rest. The Tool That Vanished tests a plan against a model that went offline 72 hours after launch. The Hottest New Job in HR explains why the people who do this work are so hard to hire (it took me more than six months to fill one role).

If you read only one

Read The Accidental Apprenticeship. Its argument runs under almost everything else on this site: nobody designed the learning inside entry-level work, so nobody planned for what happens when AI takes the tasks that carried it.

Special mention to The Thirty-Day Protein and The Ten-Year Scientist. It’s my personal favorite and the one closest to my heart.

Reply to any email, or write to christina@workforcerewired.co.

Which of these questions is on your desk right now?

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

A major new finding came from a Stanford and King’s College London paper spanning 41 countries: companies that adopt AI keep growing and add senior staff while the junior share of their workforce falls. The rest of the week supplied partial answers to what follows. Verizon put $70 million into free AI training, including the fund for the 13,000-plus employees it cut last year. In the Brookings paper, two MIT researchers make the case for federal wage insurance. KPMG gave its AI leader a direct line to the CEO. And jobless claims fell to 197,000, so for now the change shows up in hiring decisions while layoffs stay rare.

Subscribe now

By the Numbers

• 1.9 points: how far the junior share of the workforce fell at AI-adopting firms, against comparable non-adopters, by March 2026 (Stanford Digital Economy Lab)

• 6.7%: growth in senior employment at the same AI-adopting firms (Stanford Digital Economy Lab)

• $70 million: Verizon’s AI Skills for America program, including a $20 million fund created for the 13,000-plus employees it cut in November 2025 (Verizon)

• 13%: how far low-wage workers’ earnings stay below their pre-layoff level six years after losing a job, per research the paper cites (Brookings)

• 197,000: initial jobless claims in the week ending September 19, down from 219,000 a year earlier (U.S. Department of Labor)

Layoffs and Company Decisions

Jobless claims fall to 197,000 as employers hold on to workers

The Labor Department reported 197,000 initial claims for the week ending September 19, down 1,000 from a revised 198,000, the lowest weekly count since mid-July, and below the 219,000 filed in the same week of 2025. The four-week average fell to 202,250. The number of people still collecting benefits rose 2,000 to 1.719 million in the week ending September 12, an insured unemployment rate of 1.1%.

U.S. Department of Labor, Employment and Training Administration, September 24, 2026. Read the release

Why it matters: Employers are keeping people on payroll, whatever share of this year’s announced cuts they attribute to AI. If that changes, watch the 1.72 million still collecting benefits: the count climbs when laid-off workers take longer to find the next job.

KPMG puts one AI leader in the CEO’s room

KPMG LLP is folding its AI, innovation, and technology-partner functions into a new Client Technology & Innovation group, effective October 1. Todd Lohr, a 15-year KPMG veteran who helped put together the firm’s alliance with Anthropic, will run it as a vice chair reporting directly to CEO Tim Walsh. The group owns the products KPMG sells, new deal structures, firmwide AI and data strategy, and an internal venture studio that launches businesses built to compete with KPMG’s own service lines. Lohr plans to rotate rising staff through those ventures as founders and operators before they return to audit, tax, and advisory work. The change coincides with the September 30 retirement of Steve Chase, who built KPMG’s first AI group in 2023.

Fortune, Nick Lichtenberg, September 22, 2026. Read the report

Why it matters: Reporting lines decide whose trade-offs win, and KPMG moved AI out of the committee layers and next to the CEO. Other firms can copy the rotation without building an incubator: it sends rising staff back to client work after they have priced and shipped a product themselves.

Policy and Government

Two MIT researchers make the case for federal AI wage insurance

In a September 15 Brookings paper, Guy Ben-Ishai and Neil Thompson of MIT’s FutureTech project argue that AI exposure, the share of a job’s tasks a model could perform, is the wrong basis for workforce policy, because commercial automation depends on reliability, task length, and integration cost. Per research they cite, models at the end of 2025 completed tasks that take a person three to four hours at a 65% success rate. The paper makes five recommendations: put displaced workers and new high-productivity sectors first; build training with employers around specific jobs; match programs to whether AI is cutting an occupation’s employment or its wages; expand apprenticeships; and create a federal wage-insurance program for verified AI-related displacement. Low-wage workers still earn 13% less six years after losing a job, per a study the authors cite, and Trade Adjustment Assistance ran a similar wage-insurance benefit on a budget-neutral basis.

Brookings, Guy Ben-Ishai and Neil C. Thompson, September 15, 2026 (published ten days before this edition). Read the report

Why it matters: By the authors’ reading of the evidence, programs that move displaced workers into new occupations have a weak record, so they pair retraining with wage insurance, which pays part of the gap between the old salary and the next one. Two of the five recommendations, employer-designed training and apprenticeships, need no act of Congress and can start inside a company this year.

Subscribe now

Reskilling and Education

AI adoption raises senior hiring and shrinks the junior rung, in 41 countries

Bharat Chandar of Stanford’s Digital Economy Lab and Bouke Klein Teeselink of King’s College London compared foreign affiliates of AI-adopting multinationals with matched affiliates in the same country and sector that did not adopt, using Revelio Labs data on 1.25 billion job postings and 154 million employment records from January 2021 through March 2026. After ChatGPT’s release, the junior share of the workforce at adopters fell, reaching 1.9 percentage points below the comparison group by March 2026. Most of that decline comes from senior employment climbing 6.7%; the drop in junior employment itself is not statistically significant. Total employment at adopters rose 3.3%. The junior-share decline appears in 23 of 31 country estimates and is significant in seven, including the United States. Census Bureau research reported by Fortune on September 17 shows the same pressure from the graduate side: for the most AI-exposed tenth of college majors, the odds of holding a job one quarter after graduation fell five percentage points after ChatGPT’s release.

Stanford Digital Economy Lab and King’s College London, Bharat Chandar and Bouke Klein Teeselink, September 20, 2026. Read the report. Census Bureau findings: Fortune, Catherina Gioino, September 17, 2026, Read the article.

Why it matters: Entry-level jobs are where people learn the work, and AI adopters are buying fewer of them while paying for experience. The seniors these firms compete for today came up through junior roles, so thinning that rung now shrinks the senior bench years from now.

Verizon folds its layoff fund into a $70 million AI training program

Verizon launched Verizon AI Skills for America on September 23, a $70 million program offering free AI courses from IBM, Google, Anthropic, Microsoft, Coursera, and OpenAI through one Verizon portal, with local coaching from Goodwill Industries International, the Local Initiatives Support Corporation, and the National Association for Community College Entrepreneurship. Fifty million dollars is new. The other $20 million is the Reskilling and Career Transition Fund CEO Dan Schulman created in November 2025, when Verizon cut more than 13,000 jobs. The program is open to job seekers, early-career workers, displaced workers, educators, and small businesses. Verizon’s announcement sets no enrollment target and reports no job-placement results.

Verizon, September 23, 2026. Read the release. Background: CNBC, November 20, 2025, Verizon’s job cuts.

Why it matters: Verizon tied its public AI-skills program to the workforce it cut, which makes the math checkable: $20 million across more than 13,000 departing employees comes to about $1,500 a person. The next figure to ask Verizon for is placement: how many of those employees are working again, and at what pay.

What Workforce Leaders Are Watching

• AI adopters in the Stanford data added seniors while their junior share fell. If entry-level requisitions shrank this year, who is training the people a company will want to promote in 2031?

• Verizon’s transition fund comes to about $1,500 per departing employee. What does a company budget per person for a role it cuts for AI, and does that figure cover anything past severance?

• Brookings’ case for wage insurance rests on a six-year earnings gap. Which severance and redeployment terms would change if that gap were priced in?

• KPMG’s AI leader now reports to the CEO. Where does the AI owner sit in other organizations, and does that seat include a vote on headcount plans and role design?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

TL;DR: Accenture ties promotion eligibility to weekly AI logins. Meta made “AI-driven impact” a core 2026 review criterion. Amazon built a leaderboard, told employees it would stay out of reviews, then watched employees game it anyway, a practice now called tokenmaxxing. I look at why performance ratings already struggled before AI usage got added, and what a number has to survive before a promotion committee should ever see it.

Subscribe now

Three companies made the same bet in 2026: that how often someone touches an AI tool says something real about how well they do the job, real enough to put beside their name at review time. Accenture told associate directors and senior managers in February that promotion into leadership now requires “regular adoption” of the firm’s AI tools, tracked as weekly logins, according to CNBC. Meta’s head of people told employees that starting with the 2026 cycle, “AI-driven impact” becomes a core review expectation for engineers, marketers, and designers alike, a policy reported by HR Grapevine. Amazon went further, building an internal leaderboard and pushing for four out of five developers to use AI tools every week, tracked through MeshClaw, an internal agent platform that can deploy code, sort email, and act inside Slack.

Amazon told employees the leaderboard would stay separate from their reviews. Employees found the fastest way up it anyway: run the agent on work that did not need doing. The Financial Times reported the practice spread wide enough to earn its own name, tokenmaxxing. “Some people are just using MeshClaw to maximize their token usage,” one employee told the paper. Another said the part everyone suspected: “Managers are looking at it.” They didn’t believe Amazon’s promise, and several managers kept watching the leaderboard after the company restricted who could see it.

Meta’s formal review criterion produced an identical instinct on its own, informal side. An employee built an internal dashboard nicknamed Claudeonomics, ranking the company’s roughly 85,000 workers by how many tokens they burned. Fortune reported that in a single thirty-day window, total consumption passed 60 trillion tokens, an estimated nine billion dollars in cost, with the single biggest user running through 281 billion tokens in a month. Titles went to the winners: Token Legend, Cache Wizard. Neither Mark Zuckerberg nor CTO Andrew Bosworth cracked the top 250. Bosworth had told a conference that an engineer spending a full salary on tokens proved the tool paid for itself tenfold, and the only response was to keep spending. Once “AI-driven impact” sits in a formal review, a dashboard ranking who has the most of it is not a curiosity. It is the exam key. This is Meta’s second turn here in as many weeks: last week was about the return of its managers, a coaching problem. This one is about what counts as proof of using AI well.

Microsoft shows what happens without even a formal policy. President Julia Liuson told staff AI use was “no longer optional, it’s core to every role and every level,” language close enough to a review criterion that employees treated it as one. A company spokesperson later walked it back, telling reporters there was “no formal review of an employee’s AI usage,” the kind of clarification a company issues once its first message has already reached everyone wondering about their next rating.

Three companies, three mechanisms, one identical outcome: a number folded into a review or a promotion decision stopped measuring performance within weeks of going live. Human Resources Director’s coverage of the Amazon story reached for the explanation economists have used for fifty years. British economist Charles Goodhart named it in the 1970s writing about monetary policy: when a measure becomes a target, it stops being a good measure.

Cognizant’s CEO Ravi Kumar made the same point in June, calling token consumption a “vanity metric,” and Fortune reported that Meta, Amazon, and OpenAI already leaned on it internally.

Amazon’s employees spent the following months proving Kumar right. The token count stopped describing how much useful work AI was doing the moment employees understood their promotions might ride on it, and started describing who could produce the biggest number.

Subscribe now

This should not have surprised anyone who follows performance-review research. The instrument had a credibility problem before a single employee opened a chatbot. Researchers have studied how reliably two supervisors rate the same employee’s work for decades, and the finding that should worry every promotion committee reading this is not the average score. It is what changes the score. A meta-analysis published in Frontiers in Psychology found interrater reliability for overall job performance running around .61 when ratings get collected purely for research, with no promotion or raise attached. Once the same kind of rating feeds an actual administrative decision, a raise, a promotion, a stack rank, reliability drops to about .45. In plain terms: at .45, well under half of the difference between how two managers score the same person’s work traces back to that person’s actual performance. The rest is the rater, the scale, the mood in the room that week, and noise. Accenture, Meta, and Amazon each added a second, more gameable number on top of a review process already carrying that weakness.

Tokenmaxxing added a highly gameable number to a review process the research already knew wobbled the moment a raise or a promotion walked into the room, and the wobble got worse.

Deloitte put a figure on exactly this problem in 2015, years before anyone was arguing about AI adoption. Marcus Buckingham and Ashley Goodall, writing in Harvard Business Review, described a firm spending close to two million hours a year filling out forms, holding calibration meetings, and producing ratings that 58 percent of the firm’s own executives said drove neither engagement nor performance. Deloitte’s fix wasn’t a sharper rating scale. It replaced the annual review with frequent, forward-looking check-ins with a manager close enough to the work to say something specific, and built promotion decisions on that ongoing record instead of a single year-end number. I was part of the “Reinventing Performance Management” era at Deloitte. The lesson has traveled with me since: a performance signal is only as good as the person taking it and how often they take it, never how precisely the resulting number gets reported. A promotion tied to a login count or a token leaderboard runs on the opposite logic. It sits far from the work, looks backward, and gets glanced at once a year rather than discussed.

There’s a real case for visible pressure behind an AI adoption push, and it deserves a fair hearing. Left alone, plenty of capable employees won’t touch a new tool until a deadline or a promotion cycle forces it, and executives tying AI use to career advancement aren’t wrong about the urgency. Kelly Services surveyed more than 6,000 professionals and found 69 percent of executives believe refusing to adopt AI is a bigger threat to someone’s career than AI itself, and 59 percent said they would replace an employee who resisted using it. Tying usage to a review puts that expectation in front of thousands of people at once, and fast has real value when capital is committed at this scale.

Fast bought Accenture, Meta, and Amazon adoption within weeks. It also bought review inputs so unreliable that Amazon had to publicly disown the exact use its own employees assumed the number was for, a strange place for three sophisticated measurement organizations to land, by choice.

Every workforce leader building AI usage into a review or a promotion criterion is about to run the same experiment. Decades of research on performance ratings under real stakes, plus six months of watching these three companies relearn the lesson, point to the same test before that number reaches a promotion committee. Can it be gamed faster than the underlying skill can be built? If the honest answer is yes, expect what these three companies got: a great deal of activity in the numbers, and very little information about who actually deserved to move up.

Subscribe now

Here’s how you take action

If your company is folding any version of AI usage into a review right now, logins, token counts, seat activation, ask the question research already answered: is this reliable enough to decide someone’s career, or just easy to count? If you can’t answer that with evidence, assume the worse one. Your employees already have.

Then check how close the rater sits to the actual work, and how often, before you look at the number at all. Deloitte’s decade-old fix still holds: a manager who checks in weekly with someone whose work they can see, and who builds a promotion case on that record, will tell you more than a leaderboard tracking a hundred thousand people at once, gamed or not.

One closing note, for the record: I still hate the word “tokenmaxxing.”

Christina Lexa writes Workforce Rewired, on the intersection of workforce transformation, AI, and global talent.

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

Subscribe now

In September 2026, Meta started asking engineers inside its Applied AI division a question the company spent three years trying to make unnecessary: would you like to be a manager again. Business Insider reported, and Fortune confirmed, that the ask is voluntary, and that some of the people receiving it had been managers before Meta moved them into individual contributor roles.

This is not a random team. In April, Fortune’s reporting put this exact division at one manager for every fifty employees, roughly double what most organizations have ever treated as the outer edge of a functional team, and I used that number here as the most extreme data point in what Fortune called the “megamanager era”, in Who Wants to Lead People Anymore. Meta built the flattest team in the industry, then spent the following months rebuilding management inside it.

Meta has done this before. In 2023, Mark Zuckerberg told the company that “flatter is faster” and launched what he called the Year of Efficiency, pushing managers and directors into individual contributor jobs or out the door. Three years later the pattern repeated at a smaller scale and higher speed. In March 2026, Fortune reported that analysts expected Zuckerberg to help drive a broader AI-related layoff cascade across the industry. Two months later, Meta cut about 10 percent of its workforce, roughly 8,000 people, disproportionately hitting managers, and scrapped plans to fill 6,000 open roles. Out of that cut came Applied AI, a new engineering division built to bridge Meta’s AI research and its product teams, with roughly 7,000 employees reassigned to it, some of them former managers moved into individual contributor seats for the second time in three years.

The flattest team Meta ever built did not stay silent about how that felt. Employees inside the roughly 6,500-person Applied AI organization described the work to Wired as a “gulag,” and more than 1,600 of them signed a petition against a company program that tracked clicks and keystrokes on their work devices to generate AI training data. Zuckerberg told the company it had made mistakes in the restructuring, promised no further mass layoffs this year, and named the fix: smaller management structures, more funding for team events, and the return of assigned desks. A company that spent three years arguing management layers were the problem is now funding more of them as the remedy for the team it built without enough.

Subscribe now

Meta is not alone in the instinct. Andy Jassy told Amazon to raise its individual-contributor-to-manager ratio by 15 percent, and Amazon hit the target by pausing manager hiring and moving some managers back into individual contributor work. Google cut roughly 35 percent of its small-team manager roles, Sundar Pichai telling staff the company needed to stop throwing people at every problem. Gartner has projected that through 2026, one in five organizations will use AI to flatten their structure entirely, eliminating more than half of current middle management positions. I wrote about the Amazon and Google moves here in The Math of Flat, and the reasoning behind them was not foolish. Bureaucracy really does slow decisions. A manager who spends a day relaying updates between two meetings is not creating much value, and cutting that layer can look, for a quarter or two, exactly like the discipline it claims to be.

The ratio math missed the same thing twice. A headcount ratio is an arithmetic problem: divide the org by the number of managers you are willing to pay for, and the number gets smaller. Coaching capacity is a different problem entirely, and it does not shrink just because the math says it should. A manager with fifty reports cannot sit with any one of them long enough to catch a bad decision before it ships, correct a wrong technique, or notice that someone is struggling before it shows up in their output. That takes time, and time is the one input a ratio cannot manufacture by removing the people who used to provide it.

AI changed the equation on exactly one side of that math. It let Meta and its peers produce more output with fewer people managing the process, which made the ratio look survivable on paper. It did nothing to change how much attention a person needs to get better at a hard, unfamiliar job, which is the actual capacity a flattened team runs out of first. Gallup’s research points at why that gap matters even for the AI adoption Meta is racing toward: employees whose managers actively support their use of AI are 2.1 times as likely to use it frequently, and 8.8 times as likely to say it gives them more room to do their best work. A manager stretched across fifty people has almost no capacity left to be that kind of support to any one of them. Meta built a team optimized to move fast on AI and, in the same move, removed the exact relationship its own preferred outcome depends on.

The manager was never the obstacle standing between Meta and faster AI adoption. Assuming otherwise is the mistake the company is now paying to unwind, one voluntary conversation at a time. Unwinding it division by division, without rewriting the underlying assumption, is how a company ends up having this same conversation a third time in 2028, under a different division’s name.

Meta’s own fix, smaller management structures, gets the direction right and leaves the harder question unanswered: smaller than fifty measured against what, exactly. A team that settles at twenty reports per manager still has no coaching cadence if that manager spends the freed hours on performance paperwork and headcount planning instead of the one-on-one correction a hard, unfamiliar AI-native job requires. The ratio matters. It only pays off alongside a real accounting of what a manager’s day is spent on, the same second-order question I raised about Amazon and Google’s cuts in The Math of Flat and never got a straight answer to from any of the three companies.

Subscribe now

At Deloitte Global, I led the Org Design function, and we studied the span of control for our divisions for weeks before a choice was made on the outcome. I lead workforce strategy for Technology at Capital One now, and every span-of-control decision we debate runs into the same wall Meta hit: the ratio that looks defensible in a slide is rarely the ratio a real coaching relationship can survive, and the gap between those two numbers does not show up in a headcount report. It shows up eighteen months later, in the manager who cannot name what one of fifty direct reports needs, and in the employee who decides the fastest path to a functioning team is somewhere else.

Here’s how you take action

If your organization has widened a span of control in the past two years, in the name of AI, efficiency, or both, go find the actual coaching cadence underneath the ratio before you defend the number in a leadership meeting. Ask one manager how long it has been since they sat with each direct report long enough to catch something before it became a problem. If the honest answer stretches past a few weeks, the ratio is already costing you the thing it was supposed to protect.

Meta is learning this in public, twice, which is a gift to everyone watching from outside. Set your own ceiling on paper before the math forces the correction. The alternative is Meta’s path: a petition, a livestream outburst, and a quiet form asking your best people if they would like their old job back.

Christina Lexa writes Workforce Rewired, on the intersection of workforce transformation, AI, and global talent.

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

Anthropic’s economics team released a model last week that runs AI’s effect on the U.S. economy through 2030 across three cases, from GDP a modest 1.6% above a no-AI path to 32% higher with unemployment near 12%. The through-line is that knowledge workers face the wage risk first, well before the overall jobless rate moves. Separately, Indeed’s Hiring Lab projects the labor force will shrink by about 5.9 million workers by 2032 as Baby Boomers retire, with the deepest shortages in healthcare and construction, where AI does the least to help. The present sits between those horizons: initial jobless claims held at 206,000 last week, and layoffs stay rare.

Subscribe now

By the Numbers

• 1.6% to 32.4%: the range Anthropic’s model puts on how much larger 2030 U.S. GDP could be than a no-AI economy, across its three scenarios (Anthropic).

• 11.9%: economy-wide unemployment in Anthropic’s extreme scenario, with knowledge-worker unemployment at 17.9% and knowledge-worker wages down more than 10% (Anthropic).

• 5.9 million: projected decline in the U.S. labor force, about 3.7%, between 2025 and 2032 (Indeed Hiring Lab).

• 193,100 against 177,400: registered-nurse openings projected per year by 2032, set against the total new nurses expected across the entire 2022 to 2032 decade (Indeed Hiring Lab).

• 206,000: initial U.S. jobless claims for the week ending September 5, near a band that has held since mid-July (Labor Department).

Layoffs and Company Decisions

Jobless claims hold near 206,000 as employers keep their people

The Labor Department reported initial unemployment claims dipped by 1,000 to 206,000 for the week ending September 5, just above the 205,000 economists had expected. Since mid-July, weekly claims have stayed inside a band of 189,000 to 212,000. Employers are keeping workers on even as hiring stays subdued.

Associated Press, September 10, 2026. Read the report

Why it matters: The layoff rate is historically low, and employers say their harder problem is finding people to hire. For workers, the sharper risk is a frozen market with few cuts and few new openings, which strands people who want to move.

Reskilling and Education

Anthropic maps three AI economies for 2030, and the split falls on knowledge workers

Anthropic’s economics team released an interactive model, the Econ Scenario Explorer, projecting how AI could move U.S. output, jobs, and wages through 2030. The modest case puts 2030 GDP about 1.6% above a no-AI path. Scale up to the substantial case and growth roughly doubles: GDP runs 8.3% higher, unemployment sits at 4.6%, and knowledge-worker wages stay about flat. The extreme case, which assumes rapidly adopted self-improving AI, pushes GDP 32.4% higher while economy-wide unemployment nears 12%, knowledge-worker unemployment reaches 17.9%, and knowledge-worker wages fall more than 10%. The model draws on a survey of 10,980 U.S. adults fielded in August and a working paper by Anton Korinek and co-authors, who describe the scenarios as illustrative ranges that carry no probabilities. Anthropic builds Claude, so the work comes from a company with a stake in AI’s path. Across the three cases, faster growth and weaker pay for knowledge workers can arrive together.

Anthropic, Anthropic Institute Working Paper No. 2026-02 (Korinek, Jones, Sacher, Cotter, and McCrory), September 9, 2026. Read the report

Why it matters: The distance between the mild and severe cases is itself the planning problem, since one workforce plan cannot serve a 1.6% economy and a 32% one. Knowledge workers absorb the wage risk first, even where overall unemployment stays moderate, so pay structures and role design deserve attention now.

Indeed: the labor force will fall 5.9 million by 2032, and AI will not fill the gap

Indeed’s Hiring Lab, in an analysis by Felix Aidala, Laura Ullrich, and Sneha Puri, projects the U.S. labor force will fall about 3.7%, some 5.9 million workers, between 2025 and 2032 before a partial recovery later in the decade, driven by retiring Baby Boomers. The shortages concentrate in healthcare and construction. Registered nursing alone faces about 193,100 openings a year by 2032 against roughly 177,400 new nurses entering the workforce across the entire 2022 to 2032 decade, and 92% of construction firms reported trouble finding qualified workers last year. AI does little to help there, since its labor effects concentrate in high-wage, white-collar work.

Indeed Hiring Lab, Felix Aidala, Laura Ullrich, and Sneha Puri, September 10, 2026. Read the report

Why it matters: Workforce plans built around AI-driven headcount cuts miss the binding constraint for much of the economy, which is a shrinking pool of workers. Employers in care and the trades will compete on wages and training pipelines long before they compete on automation.

What Workforce Leaders Are Watching

• Whether to plan for a 1.6% AI economy or a 32% one, when Anthropic’s own model spans that range and assigns no odds.

• How to protect knowledge-worker pay if growth and wage pressure arrive together, as the middle scenarios suggest.

• Whether care and construction employers move now on wages and training, or wait for the shortage Indeed dates to the early 2030s.

• How much of the current calm is really a hiring freeze that strands workers who want to move.

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Gartner told CIOs this week that about 30 percent of workers laid off and replaced by AI will be rehired by 2029, often at a higher price than it cost to let them go. The forecast rests on a plain problem: cuts made to bank AI savings drain the experienced staff and institutional knowledge a company still needs. Separately, the New York Fed counted what firms in its region are doing, and the picture is one of retraining: 4 percent of service firms laid anyone off because of AI, while a third retrained the workers they have.

Subscribe now

By the Numbers

• 30 percent: the share of AI-displaced workers Gartner expects to be rehired by 2029, often at a higher cost than the savings that justified the cut.

• 75 percent: organizations booking AI productivity gains as cost savings that Gartner expects to be outrun by 2027 by rivals who reinvest those gains.

• 61 percent: service firms in the New York Fed’s region now using AI, up from 40 percent a year ago and 25 percent in 2024.

• 4 percent: service firms that laid off workers because of AI in the past six months, versus just over a third that retrained them.

Reskilling and Education

Gartner expects a third of AI layoffs to reverse by 2029

Gartner’s new Future of Work research, released September 9 alongside its IT Symposium, predicts that by 2029 about 30 percent of employees laid off because AI replaced them will be rehired, often at a higher cost. VP analyst Tori Paulman’s argument is that executives who treat AI mainly as a way to cut costs risk cutting too deep and too soon, losing the experienced staff and institutional knowledge they need to compete as the technology matures. Gartner’s alternative is a “talent remix,” using AI to reshape roles and move people off less productive work onto new opportunities. A second forecast sharpens the point: by 2027, organizations that book AI productivity gains as savings will be outrun by competitors that put those gains into new products, modernization, and training.

Gartner, September 9, 2026. Read the report.

Why it matters: The rehiring forecast puts a price on layoffs pitched as AI efficiency: the same workers, hired back later, at a premium. For HR leaders, the cheaper move is to plan the role change before the cut rather than pay for the rehire after it.

New York Fed survey finds retraining far outpacing AI layoffs

The New York Fed’s August business survey, published September 1, found AI use now widespread across the New York and Northern New Jersey region: 61 percent of service firms reported using it, up from 40 percent a year ago and 25 percent in 2024, with manufacturers at 51 percent. Layoffs stayed rare. Four percent of service firms said they let workers go because of AI in the past six months, and no manufacturers did. Retraining was the common response: just over a third of service firms and more than a fifth of manufacturers retrained existing workers, mostly to use AI in the jobs they already hold. The economists, Jaison Abel, Richard Deitz, Natalia Emanuel, and Nick Montalbano, flag one caveat: entry-level workers may still lose the routine tasks that once got them in the door.

Federal Reserve Bank of New York, Liberty Street Economics; Abel, Deitz, Emanuel, and Montalbano; September 1, 2026 (9 days old). Read the report.

Why it matters: The survey covers one region, so read it as a live count of behavior rather than a national tally. What it shows is employers spending on the workers they already have, which is the choice Gartner is urging companies to make on purpose.

What Workforce Leaders Are Watching

• Whether your own AI business case counts the cost of rehiring in 2027 and 2028, or only the savings booked this year.

• Which roles you would reshape rather than cut if you took Gartner’s “talent remix” seriously, and who owns that call, HR or the CIO.

• Whether the retraining your firm offers moves people into new work, or only teaches them to use AI in the job they already hold, which is mostly what the New York Fed found.

• Where the entry-level door stands once AI takes the routine tasks that used to train new hires, a caveat both the Fed economists and Stanford’s researchers keep raising.

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Satya Nadella runs one of the largest concentrations of AI computing power on the planet, and this year he admitted something that should stop every workforce planner mid-sentence. Microsoft owns GPUs it cannot turn on.

“The biggest issue we are now having is not a compute glut, it’s power,” he said. “You may actually have a bunch of chips sitting in inventory that I can’t plug in. I don’t have warm shells to plug into.” (TechSpot)

Picture that literally: some of the most sought-after silicon on Earth, sitting in a warehouse, next to a wall nobody has wired to the grid yet. The constraint on the most advanced industry in human history turns out to be the person who runs the conduit.

Jensen Huang has been saying this for a year, in public, to anyone who would listen.

“AI gives America the opportunity to build again,” he told a room full of tradespeople in May. “Electricians, plumbers, iron workers, technicians, builders, this is your time.” (Fortune)

Microsoft president Brad Smith went further. He asked his own leadership team to name the single biggest obstacle to the company’s US data center expansion, expecting to hear about permitting or transmission lines. The answer that came back was a national shortage of electricians. (Fox Business)

The scale of what these companies are building explains why. Goldman Sachs Research puts US data center power demand at 31 gigawatts in 2025, climbing to 41 gigawatts this year and 66 gigawatts by 2027, more than double in two years. In the same report, Goldman’s own analysts name the most common reason scheduled projects slip: labor and supply chain shortages. Only 50 to 60 percent of the capacity currently scheduled is expected to come online on time. (Goldman Sachs)

McKinsey sizes the buildout driving that demand at nearly $6.7 trillion in cumulative data center capital spending worldwide through 2030, more than 40 percent of it in the United States. A single large facility employs up to 1,500 workers during construction, and McKinsey’s own researchers flag the exact bottleneck Nadella and Smith are living: finding trained talent for data center and power infrastructure work is a real constraint when those same electricians and technicians are already booked on other jobs. (McKinsey) Electrical work alone accounts for 45 to 70 percent of a data center’s total construction cost, according to the electricians’ own union, which is why a first-year apprentice on a Virginia data center job can now clear six figures with overtime. (IBEW Local 26, via NPR)

Here is the math behind the shortage. The National Electrical Contractors Association estimates that roughly 7,000 new electricians enter the trade each year while 10,000 retire, a standing deficit that predates the AI boom and worsens because of it. (Fortune) Close to 30 percent of union electricians are already between 50 and 70 years old. (Qmerit) The Bureau of Labor Statistics projects about 81,000 electrician openings a year through 2034, most of them from retirements and workers who transfer out of the trade. (BLS) McKinsey separately estimates the data center build-out alone requires 130,000 more electricians, 240,000 more construction laborers, and 150,000 more construction supervisors between 2023 and 2030. (USBE, reporting McKinsey)

None of this happened by accident. For twenty years, the loudest career advice in America pointed one direction: college, then a desk. Career counselors got measured on four-year enrollment rates, so that is what they optimized for. Parents who could afford a fallback for their kids chose the one with a lecture hall in it. Vocational and trade programs got cut from high schools first when budgets tightened, treated as the plan for students the system had already written off.

Other countries built the opposite habit. Germany’s dual system runs roughly 1.2 million apprentices at any given time, pairing paid on-the-job training with vocational school, and treats that path as a first choice rather than a consolation prize for students who did not get into university. (German Trade and Invest).

That bet is now inverted, and not enough people have noticed. The AI economy’s most valuable new workers over the next decade will run conduit through a live data hall, terminate a medium-voltage cable safely, and read a single-line diagram, skills the compute layer depends on and a laptop cannot replace. The roles that actually limit the buildout carry a training pipeline that money cannot compress.

Some of the largest AI companies have started to notice too. Meta, Google, and BlackRock have pledged more than $265 million combined to train electricians and other tradespeople for data center work, and Google’s commitment to the Electrical Training Alliance alone targets 100,000 upskilled electricians and 30,000 new apprenticeships by 2030. (UBOS) Microsoft’s own Datacenter Academy is now more than 40 partner colleges deep, pairing students with paid internships inside working facilities. (Microsoft)

Take the steelman seriously for a moment. Maybe this is simply the market working exactly as it should: demand spikes, wages rise, and the trades absorb a generation of workers who might otherwise have chased a degree with diminishing returns. That is a real and welcome effect, and Huang is right to point at it. But wages rising this year do not produce a licensed electrician this year. A standard registered electrical apprenticeship runs four years and a minimum of 8,000 hours of on-the-job training under the US Department of Labor’s own rules, most of it spent working alongside a journeyman on live jobs, before a candidate ever sits a state licensing exam. (GetLicenseReady) A cohort that started training in 2026, the year these pledges were announced, graduates around 2030, the same year Goldman’s demand curve and McKinsey’s capital projections both assume the buildout is well underway. The apprenticeship’s calendar decides when the gap closes, regardless of how fast the money moves.

The boom straining the trade today will not run forever, and the evidence is already piling up. Bain’s own mid-2026 forecast describes the industry moving from “the early scramble of generative AI-driven demand” to “a more disciplined, selective, power-constrained” phase, with hyperscalers growing choosier about where and what they build. (Bain) CBRE and Wood Mackenzie clocked the first construction slowdown in six years in the back half of 2025: new capacity additions fell by half quarter over quarter even as capital spending kept climbing toward a trillion dollars. (CIO Dive) The slowdown has political help. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced federal legislation in March 2026 to pause new large-scale AI data center construction, and more than 100 local communities have already passed their own moratoriums. (Brookings) Gallup polling puts the opposition at 7 in 10 Americans who say they do not want a data center built for AI in their area. (NPR)

Even where construction stays on schedule, the work has an expiration date that has nothing to do with politics. A hyperscale campus that employs up to 1,500 workers to build it settles into roughly 100 to 200 permanent jobs once it opens, according to a rigorous county-level study by Brookings economists Dany Bahar and Greg Wright. (Brookings) The crew that wired the building moves to the next site. Eventually, in any given metro area, there is no next site.

Chris Brooks has already lived through one cycle of this. He wired his first data center in Virginia in 1999, long before most people had heard the term, and has spent the current boom watching his union brothers do well. He has also spent this year watching opposition spread.

“Every time I turn on TV, I’m hearing moratoriums have been put in place because people don’t want data centers,” he told NPR this month. “It’s popular not to want a data center now. And not that long ago, no one knew what they were.” (NPR)

Brooks is right, and the apprentice’s bet was never really on data centers. The credential a fourth-year apprentice earns wiring a hyperscale campus is the same credential that wires a solar farm, an EV charging depot, or the substation a strained grid needs next. Data centers already account for roughly two-thirds of a single year’s spike in wholesale power costs across the PJM grid, the operator serving 65 million people from Virginia to Illinois, and every megawatt of that new load requires the same trade to build the lines and stations that carry it. (Brookings) Grid modernization, building electrification, and the renewable buildout all draw from the same finite pool of licensed electricians, and none of them have a completion date the way one campus does. Someone who spends four years learning to run conduit is betting on a portable license, and the specific industry paying the premium today is almost beside the point.

This is the durable rule underneath every AI headcount story this newsletter has told, extended one step further: capital can be raised in a quarter, and it can walk away just as fast when the politics or the economics shift. Capability compounds on a training calendar nobody can rush, and once it exists, it does not expire when one customer’s spending slows. Whatever the industry, whoever the smartest people in the room happen to be, the workforce constraint always shows up on the calendar the training program set, and the credential that calendar produces outlasts the deal that funded it.

Here’s how you take action

If you are advising a young person on where to point their next four years, or setting workforce strategy for an organization racing to build anything AI touches, treat trade and technical training capacity as seriously as you treat chip supply or cloud contracts. Ask where your organization’s electricians, high-voltage technicians, and controls engineers will come from in three years. The budget question is the easy one. Then ask the question almost nobody in this boom is asking yet: when this specific site’s construction wraps, where does the crew go next. A training pipeline built around one customer is a bad bet. One built around a portable, licensed trade is not, no matter how the AI story ends.

Christina Lexa writes Workforce Rewired, on the intersection of workforce transformation, AI, and global talent.

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

The Economist counted roughly one million new US jobs tied to the AI build-out this week, against about 200,000 layoffs it attributes to the technology. Electricians and grid engineers wiring data centers, the construction crews building them, and the software developers and data scientists closest to the work account for most of the gain. Routine service and administrative roles are the exception, and they are shrinking.

By the Numbers

• About 1 million US jobs The Economist ties to the AI build-out, from data-center construction to AI engineering.

• Roughly 200,000 layoffs it attributes to AI, outweighed by the job gains.

• About 730,000 jobs added above trend since 2022 in the professions closest to AI: engineers, software developers, mathematicians, and data scientists.

• Nearly 60% one-year rise in data-center construction spending, creating work for electricians, HVAC specialists, and grid engineers.

Layoffs and Company Decisions

The Economist counts a million AI-era jobs, and a shrinking lane for routine work

The layoff headlines of 2026 have run ahead of the payroll data, and The Economist set out to reconcile the two. Its count: AI has created about one million US jobs, most of them in the physical build-out and the technical work around it, against roughly 200,000 layoffs it pins on the technology. The professions sitting closest to the boom, engineers, software developers, mathematicians, and data scientists, have added around 730,000 jobs above their prior trend since 2022. Data-center construction spending rose nearly 60% in a year, and the hiring shows up in trades that never appear in an AI press release: electricians, HVAC specialists, grid engineers, equipment technicians. The macro backdrop holds up the argument. The Bureau of Labor Statistics reported 162,000 jobs added in August at 4.1% unemployment, and the gap between unemployment for 20-to-24-year-olds and the overall rate sits near a multi-decade low, which undercuts the story that young workers are the first to go. The piece is clear about who is losing ground: routine service and administrative roles are declining, and workers in those jobs are absorbing the displacement while the gains go elsewhere.

The Economist, September 4, 2026 (published 5 days ago). Read the analysis.

Why it matters: The net number hides a split screen. Workers near the build-out are getting hired while people in clerical and routine service jobs watch their options thin, so a healthy aggregate and real displacement are both true at once. For workforce leaders, the planning question is which of your roles sit on which side of that line, and what you owe the people on the wrong one.

What Workforce Leaders Are Watching

• Which roles in your organization sit closest to the AI build-out, and which are the clerical and routine service jobs the same trend is thinning? Map both before the next planning cycle.

• New demand favors engineers, data scientists, and skilled trades. Does your reskilling budget reach the service and administrative workers whose roles are shrinking, or only the technical bench?

• The 20-to-24 unemployment gap is near a multi-decade low today. What early-career indicator would tell you the entry-level door is closing before it reaches a headline number?

• A net-positive jobs count can sit alongside displacement inside your own walls. Do you manage to the aggregate or to the affected worker?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

The economy added 162,000 jobs in August, about three times the average monthly gain of the past year and well above the modest increase economists had forecast, while the unemployment rate held at 4.1 percent. Hiring concentrated in food service, local government education, and health care. The information sector, which covers software publishing, data processing, and web hosting, lost 23,000 jobs and has now shed positions for more than a year. Separately, Challenger, Gray & Christmas reported 52,881 announced layoffs for August, the lowest August total since 2022, and found that artificial intelligence had dropped to the fourth-most-cited reason for cuts after leading every month since March.

Subscribe now

By the Numbers

• 162,000 jobs added in August, about three times the prior 12-month average of 31,000

• 4.1 percent unemployment, unchanged over the month, with 7.0 million people out of work

• 23,000 jobs lost in the information sector, including software publishing, data processing, and web hosting

• 55,000 jobs added to June and July combined through revisions, turning July from a reported loss into a gain

• 52,881 layoffs announced in August, the lowest August total since 2022

• 3,462 August cuts attributed to AI, its fourth-place ranking after leading every month since March

Layoffs and Company Decisions

Payrolls rose 162,000 in August, well above forecasts

Total non-farm payrolls increased by 162,000 in August, the U.S. Bureau of Labor Statistics reported, about three times the average monthly gain of 31,000 over the prior year and well above the modest increase economists had forecast. The unemployment rate held at 4.1 percent, with 7.0 million people unemployed. Food services and drinking places added 59,000 jobs and local government education 42,000, while manufacturing continued a modest recovery at 16,000. The information sector, which includes software publishers, data processing, and web hosting, lost 23,000 jobs and has now declined for more than a year. Revisions added a combined 55,000 jobs to June and July, turning July from a reported loss into a gain. Average hourly earnings rose 0.3 percent to $37.75, up 3.1 percent over the year.

U.S. Bureau of Labor Statistics, September 4, 2026. Read the report.

Why it matters: The headline looks strong, though the gains cluster in food service, schools, and health care rather than the higher-paying office roles that build career ladders. The information sector, the corner most exposed to AI and automation, kept cutting even in a good month. A rising payroll total can hide where the durable openings actually are.

Challenger’s August tally shows AI slipping down the list of layoff reasons

Challenger, Gray & Christmas counted 52,881 announced job cuts in August, the lowest August total since 2022. Restructuring led the reasons at 16,173 cuts, followed by market and economic conditions at 15,260. Artificial intelligence, which led the monthly count from March through July, fell to fourth at 3,462 cuts, though it remains the top reason year-to-date at about 22 percent of all cuts. Andy Challenger noted that announced hiring plans, 119,825 so far this year and up 37 percent from 2025, are filling slowly, and questioned whether employers will find workers with the skills they need.

Challenger, Gray & Christmas, Colleen Madden Blumenfeld, September 2, 2026. Read the report.

Why it matters: AI leads the year’s layoff tally, so one month at fourth place probably reflects how companies label the same cuts rather than any real pullback. Read next to the jobs report, the picture is a labor market adding lower-wage roles while the AI-exposed and white-collar categories stall. The number to watch is how many of this year’s 119,825 announced hires actually convert.

What Workforce Leaders Are Watching

• Whether the information sector’s year-long slide, another 23,000 jobs in August, is the leading edge of AI displacement or a post-pandemic tech correction still working itself out.

• Whether August’s growth in food service, schools, and health care can support the career mobility that office roles used to provide.

• Which internal candidates a company can move into unfilled roles before hiring outside, given how slowly this year’s announced hiring is converting.

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Uber told staff on Wednesday it would cut about 3,300 jobs, its largest reduction since 2020, and Dara Khosrowshahi attributed the decision to the company’s own management layers. Separately, ADP reported that private employers added 38,000 jobs in August, the smallest monthly gain since January. Both developments show contraction that does not trace cleanly to automation: Uber thinning the middle of its org chart, and a labor market cooling across the categories where large employers have the most room to hold headcount flat.

Subscribe now

By the Numbers

• 3,300 Uber jobs cut, about 10% of its corporate workforce, the company’s largest reduction since 2020.

• 38,000 private-sector jobs added in August, the smallest monthly gain since January and down from 46,000 in July.

• 16,000 jobs lost in professional and business services in August, the white-collar category where AI adoption runs highest.

Layoffs and Company Decisions

Uber cuts 3,300 jobs and blames its own complexity

Uber told employees on September 2 it would eliminate about 3,300 roles, roughly 10% of its corporate workforce and its largest cut since the 2020 pandemic reductions. Khosrowshahi framed the decision as removing management layers and simplifying team structures, and said the company would halve the number of “micro-teams” in which a manager oversees only one or two people. The cuts arrive while Uber is growing: second-quarter revenue rose 12% to $14.2 billion, and full-year 2025 revenue rose 18% to $52 billion. Across the tech sector, AI has been the most-cited reason for job cuts for months. Khosrowshahi did not reach for it, and pinned the reduction on bureaucracy the company built during its own expansion.

Bloomberg, September 2, 2026. Read the report

Why it matters: The target is the middle of the org chart, the managers of one or two people who accumulate as a company scales. A profitable company cutting 10% and crediting its own structure hands workforce leaders a cleaner test than the AI-attribution debate: how many layers exist because the work needs them, and how many exist because headcount grew faster than the design.

Private hiring slowed to 38,000 in August, the weakest since January

Private employers added 38,000 jobs in August, down from 46,000 in July and the smallest monthly gain since January, according to the ADP National Employment Report produced with the Stanford Digital Economy Lab. Education and health services added 45,000 and led every category. Manufacturing lost 17,000, and professional and business services lost 16,000, the white-collar work where AI tools are most concentrated. Base pay rose 3.2% over the year, and nearly all the net gain came from employers with 500 or more workers; firms under 50 added just 3,000.

CNBC, September 2, 2026. Read the report

Why it matters: The categories losing ground are the ones where AI tools run heaviest and where large employers have the most room to leave a departure unfilled. One month is not a trend, and hiring has now stalled at a level that leaves workers little cushion. The AI effect on payrolls reads more clearly in which jobs stop being backfilled than in any single layoff announcement.

What Workforce Leaders Are Watching

• When a growing company cuts 10% and names its own management layers as the cause, which layers in your org exist because the work needs them, and which exist because headcount outran the design?

• If professional and business services keeps shedding roles while education and health services keeps adding them, are you holding white-collar headcount flat by decision or by drift?

• Uber halved its one-to-two-report “micro-teams.” What is the smallest team your span-of-control policy still permits, and who owns the call to merge them?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Subscribe now

It took me more than six months to hire one person.

The role was a Director to lead workforce planning for Technology at Capital One. Funded, senior, pointed at the most interesting problem in the function. Applications were never the constraint. I read resumes from talented people who had run headcount budgets, built dashboards, and led HR business partner teams. Few had done the actual work: taking a business strategy, translating it into the skills and capacity to deliver it, then defending that translation to executives who had already decided what the number should be.

Half a year and then some, for one hire.

The demand side is not subtle

Gartner named “shape work in the human-machine era” one of four priorities for CHROs in 2026, and called it a “now-next” talent strategy: what you get from people in the next twelve months, and what you set in motion for the one to three years after. Two horizons at once, held by someone who can show their math.

Mercer’s Global Talent Trends 2026, drawn from nearly 12,000 executives, HR leaders, investors, and employees, sizes it: 98% of executives are planning organizational design changes within two years, and 65% expect 11% to 30% of their workforce redeployed or reskilled because of AI in that window. Redeploying a third of a workforce is not a budgeting exercise.

The demand shows up as failure before it shows up as hiring. Orgvue surveyed 1,163 senior decision-makers across the US, Canada, the UK, and Ireland: 78% had AI projects that failed outright or stalled in pilot, and 52% ran restructures costing a combined $49.4 billion in severance. Deloitte’s 2026 Global Human Capital Trends found the mechanism: only 6% of leaders report real progress on designing how humans and AI work together.

Companies are buying tools faster than they redesign the work those tools are meant to do. Somebody has to sit between the technology roadmap and the org chart and answer what changes, for whom, by when, at what cost. American companies mostly never filled that seat.

The supply side has an accent

When I built the talent strategy, organization design, and workforce planning function at Deloitte Global, I hired across the US, Canada, and the UK. My workforce planning lead, who taught me most of what I know about strategic workforce planning, sat in the UK. So did the depth of the market. Search for someone who had run this work before and the resumes came back British far more often than the relative size of the two labor markets would predict.

Britain built the profession on purpose. The CIPD lists workforce planning as specialist knowledge in its Profession Map, publishes the guidance practitioners work from, and runs an accredited program. A defined route in, a body of practice, a credential at the end. The Workforce Planning Institute certifies practitioners and holds annual gatherings in Chicago, London, and Sydney.

In most US companies the work landed inside Finance as headcount budgeting: a number per quarter, a hiring plan, a variance report, an argument every October. Useful, and a different craft from forecasting which capabilities a business will need in three years and what they cost to build.

Employment law widened the gap. A UK employer proposing 20 or more redundancies in 90 days must consult employee representatives for at least 30 days before the first dismissal, 45 days at 100 or more, with the penalty doubled to 180 days’ pay per employee in April 2026. A British company has to explain and negotiate its workforce plan before acting on it. The US federal WARN Act asks for 60 days of notice and stops there. Notification is a letter. Consultation is an argument, and an argument requires evidence, modeling, and someone who can stand behind the assumptions when a works council pushes back. Do that for twenty years and you produce practitioners. Skip it and you produce spreadsheets.

Subscribe now

What the British bench proves

My hypothesis needed correcting. I assumed the UK was simply better at this.

CIPD data reported in its July 2026 analysis shows two-thirds of organizations plan a year or less ahead, 13% only to the next quarter, and fewer than a fifth beyond two years. Only 38% collect data on their own skills gaps. That is the country with the credential, the Profession Map, and the consultation law.

Britain built a career path. Mastery is scarce on both sides of the Atlantic. A British CHRO can hire someone who spent a decade doing this work under a title that says so. An American CHRO is looking for someone who assembled the same skill set by accident, from adjacent jobs, without anyone naming it. The route that would produce them in numbers, and label them so a hiring manager can find them, has never been built.

Organization design and workforce planning are one job

The two disciplines answer different halves of the same question, and each fails alone.

Organization design decides what work exists, where it sits, who owns the decision, and how many layers stand between the customer and someone who can say yes. Workforce planning decides how many people, with which capabilities, arriving when, at what cost, from where. Run design without planning and you get an elegant operating model that assumes people who do not exist and cannot be hired on the timeline it requires. Run planning without design and you produce a precise forecast for an organization being taken apart around you.

I put both under one roadmap at Deloitte Global, and run them that way now at Capital One, because the alternative kept producing that second failure. A workforce plan built against last year’s structure ages the moment the structure changes, and the structure keeps changing.

AI raises the cost of the separation, because it changes work at the task level before it changes anything at the role level. Orgvue puts roughly 66% of jobs in AI’s path while only about a quarter of the tasks inside them can currently be automated. Roles are being hollowed out and refilled from the inside. What work belongs to this job now, and what skills we need to staff it, are one question arriving from two directions. Companies keep sending it to two leaders who meet quarterly.

I wrote in July about companies merging the People and Technology seats to close a similar gap. This one is easier. Both disciplines already report to the same executive in most companies. They just do not share a roadmap or a set of metrics.

Subscribe now

HR holds the pen

The function belongs in HR. HR owns the org design capability, the people data, the job architecture, and the leveling logic. Put workforce planning anywhere else and it drifts toward the owner’s instincts. Inside Finance it becomes a cost exercise. Inside a technology organization it becomes a tools roadmap with a headcount appendix. A business unit produces a hiring wish list with a good story attached.

Holding the pen is different from writing alone. The plan is co-authored, and three other leaders own pieces HR cannot supply. Technology owns the delivery roadmap that decides which work changes and when, and knows which capabilities the platform will need before HR does. Finance owns the cost model and the constraint the plan has to clear, which separates a workforce plan from a list of hopes. Business leaders own the demand, and the revenue plan is what makes capability forecasting better than extrapolation.

Deloitte’s workforce planning practice describes the same shift and is blunt about where the work used to sit: a small team of specialists with finance holding the loudest voice in the room, which made sense while the exercise was cost and headcount plugged into a model. Deloitte now places workforce planning at the intersection of finance, technology, human capital, and business strategy. Most companies have not moved. The American Productivity and Quality Center found 49% of strategic workforce plans are top-down or not integrated across divisions, and a Bersin survey found only 32% of talent acquisition professionals involved in planning at all.

The companies that do move can show what it buys. Network Rail, which runs Britain’s rail infrastructure, was taking twelve months to recruit and train signalers and losing them fast enough to open real gaps. A team spanning HR, operations, finance, training, and the trade unions cut that end-to-end time by half, with roughly £1.2 million in expected savings over three years from premium hours it no longer had to pay. A biotech company cutting general and administrative costs after the pandemic ran the same play and saved about $94 million.

A disclosure: I am not a neutral reader of Deloitte’s research. I spent nearly a decade at Deloitte Global and have spent time with several of the people behind this work, including Russell Klosk, who leads their workforce planning practice, and Kyle Forrest. Russ signed my copy of his book, Talent Prophecy, at the Eightfold conference in May.

What the job actually requires

Three adjacent disciplines get mistaken for this one, which is much of why the searches stall. People analytics describes what already happened: attrition by team, time to fill, the cost of a bad quarter. Headcount budgeting prices the workforce that was approved and tracks variance against it. Recruiting planning turns an approved number into requisitions and sourcing channels. All three are real crafts, every company needs them, and a strong practitioner in any of them can be hired this quarter.

Workforce planning starts where those stop. It makes a claim about a workforce that does not exist yet, inside a strategy still being argued over, and attaches numbers, dates, and costs to that claim in a document leaders will be held to.

Doing it well takes four capabilities at once. The quantitative half is visible: supply and demand modeling, scenario testing, skills taxonomies, and enough statistical judgment to know which assumption the model turns on. Underneath sits organizational knowledge no onboarding deck transfers: how the company actually decides, where the constraints sit, whose plan has moved three times this year, which capability takes eighteen months to build because you watched it take eighteen months. Then the relational part. A plan gets adopted when a CFO, a CIO, and a business unit head who each walked in with a different number sign the same one, and that runs on trust built long before the meeting. Tool fluency finishes it: knowing the platform and the labor market data underneath well enough to catch the tool when its answer is wrong.

The Director I hired had all four. She could read a strategy and name the capabilities it implied, build the model, and explain it to a CFO without hiding behind it. Put her in an organization design session and she could say what a leader’s whiteboard redraw would cost in hiring time and capability eighteen months out.

The gap is measurable. BCG and the World Federation of People Management Associations surveyed more than 7,000 HR and business leaders across 115 markets and ranked 28 people-management topics by future importance against current capability. Strategic workforce planning came out as a universal anchor, a priority in every region. In the same study, 11% of companies have a skills taxonomy embedded across the enterprise.

Each of those four capabilities is common on its own. The combination is rare and takes extensive time to develop. Recruiting calls about these roles are on the rise, and the searches they describe have usually been open for months.

Subscribe now

Here’s how you take action

Give the seam an owner, and put that owner in HR. Name one leader accountable for both the shape of the organization and the supply of people to fill it, then name three counterparts: the technology leader who owns the delivery roadmap, the finance leader who owns the cost model, the business leader who owns the demand. Those four set the assumptions together at the start of the cycle. Four functions comparing notes quarterly is a coordination problem that becomes an execution problem later.

Screen for the combination. Platform mechanics are teachable in a month. Modeling judgment, organizational knowledge, and the standing to hold a room of executives who disagree take years, and a candidate with three of the four will struggle. Ask for a forecast they got wrong and what they changed after, then ask how they got a finance leader to sign something the finance leader did not like.

Shorten the cadence. Annual workforce plans made sense when the work changed annually. Deloitte’s 6% is what an annual cycle looks like against a quarterly tool release schedule. Start with skills-gap data on the roles carrying the most business risk, since fewer than four in ten organizations collect it at all.

Build the bench, because you cannot buy it. Rotate people in from financial planning, organization design, and analytics, and give them a real assignment with executive exposure. Companies that build a pipeline over the next three years will stop competing for the same forty people.

Workforce planning is a quantitative, high-judgment discipline most American companies have never named or staffed. AI just made it urgent. If you run a Technology, HR, or Finance organization, ask who in your building can translate next year’s strategy into next year’s capability requirement. If the answer takes more than a moment, you have found the work.

Christina Lexa writes Workforce Rewired, on the intersection of workforce transformation, AI, and global talent.

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

Meta spent the first half of 2026 designing a plan to hand much of its engineering work to AI agents, then shelved the deepest part of it once the productivity numbers came back. A Reuters Special Report published August 26 laid out Project OT: pods of three to five “builders” replacing product teams of 10 to 20, some teams cut by as much as 60 percent, management layers thinned, and AI agents doing the rest. Mark Zuckerberg halted the November wave the night before it was due to start. Separately, a survey of employed U.S. workers put a number on how little of this has reached them so far. Three percent said AI had cost them a job.

By the Numbers

• 60 percent: the maximum team-size cut Meta explored under Project OT before pulling back.

• 220 percent versus 36 percent: Meta’s AI-assisted code changes rose 220 percent year over year, while the changes that reached users rose 36 percent.

• 3 percent: share of surveyed U.S. workers who said AI had cost them a job since 2023.

• 6 percent: share who said they held a job that did not exist before AI.

Layoffs and Company Decisions

Meta halts the biggest wave of its AI-agent restructuring

Project OT, short for Organization Transformation, came out of Meta’s January leadership retreat. The design replaced conventional product teams with small pods of “builders” supervised by AI agents, removed layers of management, and mapped two rounds of cuts, one in May and a larger one in November. In April, Meta installed software on U.S. employees’ machines to capture keystrokes and mouse movements, data meant to train the agents that would take over the work. Workers who believed they were training their own replacements pushed back, and internal sentiment scores dropped 19 points. The productivity case did not hold up either. Code changes to Meta’s platforms ran 220 percent above the prior year, while the changes that reached users rose 36 percent, and reliability warnings climbed alongside the AI-written code. Zuckerberg stopped the November wave hours before the first phase began, kept the May 20 cut of about 10 percent of staff, and told employees that agent development “hasn’t really accelerated in the way that we expected.”

Reuters, Special Report, August 26, 2026; background via Fortune, April 21, 2026. Read the Reuters reporting. Read the background report.

Why it matters: The plan came apart on Meta’s own measurements, where more AI-written code did not turn into more shipped value, and the tracking built to train the agents cost employee trust first. Any leader planning headcount around agent output should separate the volume of work produced from the value that reaches customers before restructuring around it.

Reskilling and Education

Workers report far less AI displacement than the forecasts

Jeffrey Dixon, a sociologist at the College of the Holy Cross, asked 1,250 employed Americans in July and August how AI had changed their own working lives. Three percent said they had lost a job to AI since 2023. Six percent said they held a job that did not exist before AI, and 9 percent reported a promotion or raise tied to it. Dixon, who describes himself as an AI pessimist, called the small displacement number the part that surprised him. The survey has not been peer reviewed, and because everyone in it was employed, it speaks to the workers AI reshaped around and misses anyone it already pushed out.

The Conversation, Jeffrey C. Dixon, College of the Holy Cross, August 2026. Read the report.

Why it matters: Worker-reported displacement of 3 percent sits well below the layoff headlines and the vendor projections that put role losses in the double digits. For anyone communicating an AI strategy to staff, the distance between what leaders forecast and what employees have lived is now measurable, and it shapes how much of the message lands.

What Workforce Leaders Are Watching

• Whether your own AI-productivity dashboards separate code volume from shipped value, the way Meta’s 220-versus-36 split did, before you plan headcount around agent output.

• How employees read surveillance framed as AI training. Meta’s keystroke and mouse tracking cost 19 points of sentiment before it saved a dollar. There’s a real hidden cost there.

• Whether to design roles around a pod-and-builder model now, or wait until agent reliability catches up to the org chart it would require.

• What to tell a workforce that, in Dixon’s survey, reports far less AI displacement than either the layoff headlines or the vendor forecasts predicted.

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Editor’s note: Every week I write this newsletter myself. This week is different. Mike Childers, a good friend and one of the sharpest thinkers I know, agreed to be the first guest author of Workforce Rewired.

Mike spent a decade as a labor and employment attorney, then did the rare thing. He crossed over to build the AI systems he used to advise on, and he is now a Principal Product Manager at Amazon. I have learned more from our conversations about where legal, engineering, and product actually meet than from almost anything I have read on the subject. He keeps coming back to the question I care about most in this work: what kind of professional can operate across a boundary instead of guarding it, and what kind of organization makes that possible.

His piece argues that legal review at the end of AI development is already too late, and that the real fix is a design choice made while the system is still taking shape. It sits squarely inside what I write about here: the roles that are changing, the fluency the next decade rewards, and how good systems actually get built. Read it, then tell me who you think should write the next one.

-Christina

TL;DR: By the time an AI system reaches a legal toll gate, it is built, scoped, and approved, and the most a late review can do is flag risk that is now expensive to remove. Mike Childers makes the case that legal reasoning belongs in the room while the architecture is still open, where it shapes what gets built instead of grading it afterward. His examples are concrete: a data store that turns into a misuse vector, a deletion rule that can erase the record of a bad actor, a conversion metric that means two different things depending on how you count it. The through-line is a kind of professional, and a kind of organization, willing to close the gap between disciplines.

A few years ago, before I crossed from legal into product, a VP I supported made an unusual decision: when Legal’s travel funds for client meetings were cut, he started covering my travel to his team’s planning meetings out of his own budget.

We’d been doing this for years when, over dinner in Edinburgh after one of those meetings, he explained his reasoning. It came down to something simpler than most people expect: when legal is in the room while options are being debated, it has full context – the unfiltered version of the problem, including why each design choice was made, what alternatives were considered, and what the real constraints are. That context is rarely preserved when review happens at the end. By then, a complex deliberation has been compressed into a brief, a slide deck, or a ticket, and the information that would have changed the legal analysis has been lost in translation – not through bad intent, but because nobody knew which parts of the discussion were legally significant.

Full context and real-time delivery are not separate benefits. They are the same benefit. Being in the room means legal can give better advice – because it has heard everything – and give it immediately, while the decision is still open. The compound effect is what he was paying for: decision-makers who can focus their finite time and energy on the best of the options in front of them, rather than debating solutions that will fail a legal review or force unnecessary risk onto the organization. They don’t have to decide and then wait for a legal read that may never come. They get the analysis in the room, when it can still shape the outcome.

By the time a decision goes to a legal toll gate, engineers have built it, product managers have scoped it, and executives have approved it. The instinct is to preserve it. That instinct is expensive.

This model – legal review as a gate to clear before launch – made reasonable sense when it meant checking a terms of service or a vendor contract. It makes much less sense when the product is an AI system making consequential decisions about people.

The problem isn’t that legal teams are slow or unsophisticated – it’s structural. The most a late-stage review can do is identify risk that is now expensive or impossible to remove.

There is a better model. It requires rethinking not just when legal gets involved, but what legal involvement actually means in an AI development context.

The Difference Between Reviewing a Product and Shaping One

When legal reasoning enters the development process early – before architecture is finalized, before data flows are designed – it doesn’t slow things down. It changes what gets built.

Consider a system designed to verify identity or detect misrepresentation in a high-stakes context – hiring, lending, benefits eligibility. The instinctive engineering approach is to build a data store, enrich it with information from multiple sources, and make it available to downstream systems that need to verify someone’s claims. Clean, intuitive, extensible.

But that architecture creates problems that compound over time. Every downstream system that consumes the data becomes a potential misuse vector. Each new consumer raises questions about whether the original collection purpose covers the new use, whether data minimization requirements have been respected, and whether the technical access controls match the policy constraints that are supposed to govern them. Purpose limitation isn’t an abstract principle – it’s a design constraint that affects whether you can legally do tomorrow what seems technically straightforward today.

These questions are far easier to resolve before the data store exists than after it has been built into eight systems and three vendor integrations.

Sitting on the product side of the table now, I understand that instinct viscerally – it’s how you build extensible systems. But the legal training asks different questions first: why are we collecting it, how long do we actually need it, what happens when its purpose is served, and how should access rights change when circumstances do. Holding both sets of instincts simultaneously is what makes the early conversation productive rather than adversarial.

The Complications That Only Surface Early

There is a category of design problem that late-stage legal review almost never catches – not because the lawyers miss it, but because by the time they see the system, the decision has already been embedded in a way that makes the right answer architecturally inconvenient.

Data deletion is one of the most instructive examples. The standard legal instinct, and in many jurisdictions a regulatory obligation, is to honor requests to remove someone’s personal information. Build a deletion mechanism, make it accessible, respond to requests within the required window. Straightforward enough.

But in any system designed to detect fraud or prevent bad actors from re-entering a process, deletion creates a tension: if you remove the record of what a confirmed bad actor did, you lose the ability to recognize them when they return under a different identity. The legal obligation to delete exists to protect individual privacy interests. In this context, honoring it literally may undermine the very interests it was designed to protect – and those of the people who would otherwise be harmed.

The answer isn’t to ignore the deletion obligation – it’s to design an architecture that honors the purpose of the deletion right while preserving what is necessary to prevent recurrence: a minimal, purpose-limited retention approach that protects the system’s integrity without retaining more than is justified. That is a design decision, not a legal opinion delivered after the fact.

I now sit on the product side of exactly this problem. When legal is embedded in the design process, it catches that decision before it becomes expensive.

The Metrics Problem

The same dynamic plays out at the analytical level, and this is where the argument extends beyond legal questions into how teams reason about evidence.

Imagine an identity verification system being evaluated for deployment. Simulations show a meaningful drop in a key conversion metric when the system is active. On the surface, that looks like a business problem: fewer people completing the process means fewer candidates, customers, or applicants – depending on the domain.

But conversion metrics can be measured in at least two ways. On a single-session basis, a user who abandons a process midway and returns hours later to finish it registers as both a failure and a success – the metric appears to fall even if the user ultimately completed the task. On a multi-session basis, that same user completed the process and the rate holds. The metric looks dramatically different depending entirely on what you are choosing to count – and if you are making a significant product decision based on the wrong version, you are wrong before you have analyzed anything.

And even if conversion rates genuinely fall on the right measurement, the relevant question is whether that affects the outcomes that actually matter. If the quality of downstream results remains stable, the users being filtered out are likely not ones who would have converted to valuable outcomes anyway. A conversion rate drop combined with stable downstream results is not a problem. It may be the system working exactly as intended.

I learned to ask these questions from years of quantitative work before law school – the instinct to distinguish between a metric that looks alarming and a metric that tells you something is actually wrong. Legal training develops its own version of that skepticism: the habit of asking what evidence actually establishes, and whether an indicator shows what people assume it proves. Both instincts point in the same direction, and both are more valuable when the people who carry them are in the room while the analysis is being framed rather than after the conclusion has already been drawn.

The Organizational Implication

None of this happens automatically. It requires legal professionals who understand how the systems work, engineers who understand why the legal questions matter, and deliberate organizational choices about who is embedded where and what genuine technical fluency means in practice.

In baseball, the relationship between a pitcher and catcher is called a battery – a unified system where expertise has to flow in both directions to function. The catcher needs to understand pitching well enough to call the right pitch at the right moment. The pitcher needs to trust the catcher’s read of the batter and situation well enough to execute it. Neither can do their job in isolation, and the gap between a battery that works and one that doesn’t shows up immediately under pressure. Embedding legal in engineering works the same way. When an engineer describes a data flow, legal should be able to identify immediately where it implicates a retention obligation or a purpose limitation. When legal articulates a compliance constraint, engineering should be able to propose an architecture that preserves the product goal while removing the exposure. The conversation does not stop at the boundary of each discipline – it moves through it. That bi-directional fluency is what turns legal from a toll gate into a genuine design partner.

The question isn’t whether legal should be involved earlier. It’s what kind of professional can actually function in that role – and what kind of organization creates the conditions for them to do it. The answer requires people willing to close the gap between disciplines rather than patrol the boundary, and leaders willing to fund that investment before the ROI is obvious. The VP who paid for a lawyer to attend his planning meetings was making exactly that bet. It paid off – not because legal said no to fewer things, but because the things that got built didn’t need someone to say no.

Mike Childers is a Principal Product Manager at Amazon who spent a decade as a labor and employment attorney before crossing over to build the AI systems he used to advise on. He writes about what happens when legal architecture is upstream of engineering rather than downstream of it. You can find him on LinkedIn.

Editor’s note: posting has been light and lacking a consistent schedule due to my travel schedule. After all, there’s still a human in this loop.

Executives spent two years advertising AI by the headcount it could erase. That pitch is going quiet. Facing polls that show most Americans expect AI to shrink their opportunities, CEOs at Etsy, Patreon, and Microsoft now insist their latest cuts were not about replacing people, while conceding the technology is rewriting the work. Separately, the people trying to get hired describe a market where AI screens the resume, runs the interview, and sends the rejection, and where fewer of them believe any of it is real.

By the Numbers

• 113,000 announced U.S. job cuts have been attributed to AI so far this year, about 24 percent of all cuts, and AI was the top reason employers named in July for the fifth month running (Challenger, Gray & Christmas, reported by Axios).

• 61 percent of U.S. adults expect AI to leave most people with fewer economic opportunities, against 21 percent who expect more (CBS News/YouGov, Aug. 12 to 14, reported by Axios).

• 291 applications is what a recruiter now processes per hire, up from roughly 100 in early 2021 (Ashby, reported by Axios).

• 63 percent of U.S. applicants have been interviewed by an AI system, and 38 percent have walked away from a job because it used one (Greenhouse, reported by Axios).

Layoffs and Company Decisions

The AI layoff gets a gentler script

Two years ago, executives measured AI progress by the workers it displaced. Klarna said its assistant did the work of 700 full-time agents. Salesforce chief Marc Benioff said AI cut support headcount from 9,000 to about 5,000 “because I need less heads.” Coinbase rebuilt around smaller “AI-native” teams. Now the same rhetoric reads as a liability. Challenger, Gray & Christmas counts nearly 113,000 job cuts attributed to AI this year, and a CBS News/YouGov poll found 61 percent of Americans expect AI to mean fewer opportunities for most people. So the language has turned careful. Etsy told regulators its 220 cuts “weren’t driven by AI.” Patreon said its 20 percent reduction “wasn’t based on a belief AI could replace humans.” Microsoft’s chief people officer said the 4,800 roles it eliminated in July were “not being replaced by AI.” Each statement pairs the disclaimer with an admission that AI is changing what the company needs.

Axios, Simon Hernandez-Arthur, Aug. 20, 2026. Read the report

Why it matters: Benioff’s line about needing “less heads” and Klarna’s boast about doing “much more with less” turned the people who lost those jobs into a headcount problem, and the softer “we are not replacing people” script can bury the same outcome under kinder words. The worker who was let go reads both versions and sees they end the same way. Leaders who owe a layoff an explanation should give affected employees the specific reason their role changed, and say it to them as plainly as they say it to investors.

Reskilling and Education

The AI-run job hunt is losing the candidates’ trust

Workers are told to reskill for the AI economy, then meet an AI wall at the front door. Applicants automate their applications; employers answer with chatbots, resume filters, and recorded interviews. A recruiter now sorts about 291 applications per hire, nearly triple the 2021 figure, and it takes roughly 25 percent longer to fill a role than before the pandemic, per Indeed. Greenhouse data show 46 percent of candidates have lost faith in the hiring process, 63 percent have faced an AI interview, and only 21 percent of recruiters are “very confident” their systems are not screening out qualified people. Ghost jobs, postings companies are not actively trying to fill, run 18 to 22 percent of listings in a typical quarter. Gorick Ng, a Harvard career adviser, calls the result a “collapse of trust in the labor market,” where candidates assume every posting is fake and every rejection automated.

Axios, Josephine Walker, Aug. 21, 2026. Read the report

Why it matters: When 63 percent of applicants have been interviewed by a machine and fewer than half of recruiters trust their own filters to keep good candidates in, the cost lands on both sides as wasted hours and lost faith. A hiring process that hides its humans teaches strong candidates to stop trying. Employers who want those applicants back will need to put a person somewhere the candidate can see.

What Workforce Leaders Are Watching

• When your next restructuring lands, will the reason you give employees match the one you give investors on the earnings call, or will the gap between them do the damage?

• If AI now screens resumes and runs first interviews, who audits those tools for the qualified people they reject, and can you name that person today?

• Ghost jobs and auto-rejections are teaching your best prospects that every posting is fake. What in your process still signals a real decision on the other end?

• Worker anxiety about AI is high enough to move recruiting and retention. Are you tracking how your public AI messaging affects the people you are trying to keep?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

A Lloyds Bank survey of 1,200 UK businesses found that AI has created new jobs at 54 percent of them, one of the first large employer datasets to record the technology adding roles rather than only cutting them. Lloyds did not ask whether those same firms eliminated other jobs, so the figure measures creation without netting it against loss. The direction still matters for a debate that has run mostly on displacement: 61 percent of firms reported using AI, a quarter said they are hiring more people with AI skills, and a 20% have opened roles that did not exist a year ago.

Subscribe

By the Numbers

• 54 percent of 1,200 UK firms surveyed said AI has created new jobs inside their organization (Lloyds Business Barometer, via Bloomberg, August 17).

• 61 percent reported using AI, and 58 percent plan to raise investment over the next year, most of it aimed at training current staff (Lloyds, via Bloomberg).

• A quarter of employers are hiring more candidates with AI skills, and a fifth have created AI-specific roles (Lloyds, via Bloomberg).

Reskilling and Education

A UK employer survey records AI adding jobs, and skips the ones it took

Lloyds Bank runs a monthly Business Barometer that reads business confidence across the UK, and this week’s edition put AI questions to 1,200 companies. Fifty-four percent said AI had created new jobs inside their organization. Sixty-one percent said they now use AI in some form, and 58 percent plan to raise investment over the next year, most of it aimed at training the staff they already have. On hiring, a quarter of employers said they are bringing in more candidates with AI skills, and a fifth said they have built AI-specific roles that did not exist before.

The survey carries one plain limit. It asked whether AI created jobs; it did not ask whether AI removed any. A firm that added three AI roles and cut ten elsewhere counts as job creation here. So the 54 percent reads as one side of a ledger, useful as a floor on how much work AI generates and quiet on the net. It lands against a UK backdrop where earlier employer research had mostly tracked AI through hiring freezes and cuts, which makes an employer dataset pointing the other way worth logging even with the caveat attached.

Source: Bloomberg, “AI Is Starting to Create Jobs in the UK, Lloyds Survey Finds,” August 17, 2026, reporting the Lloyds Business Barometer.

Why it matters: Most 2026 evidence has measured AI by the jobs it removes, and this is one of the first large employer surveys to record it adding roles, even while it leaves the losses uncounted. The part a workforce planner can act on is the demand shift: a quarter of these firms already screen for AI skills and 20% have built new roles around them. Treat AI-skill hiring as a live budget line this quarter.

What Workforce Leaders Are Watching

• A survey that counts AI creating jobs but never asks what it eliminated tells you half the story. How much weight does it get against payroll data that shows hiring stalling?

• A quarter of these UK firms now hire for AI skills. Which roles on your own team carry an AI-fluency screen, and who inside the company gets first claim on training before you recruit outside?

• Lloyds found 58 percent raising investment mostly for upskilling. Is your training budget sized to reskill the people you have, or is it a line item funding a course catalog few employees finish?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.

Subscribe now

In 2008, a New York auction house put dozens of bottles of rare Burgundy from Domaine Ponsot on the block, some labeled with vintages from the 1940s. The provenance looked clean and the bidders were ready. Then Laurent Ponsot, whose family makes the wine, saw the catalog and flew to New York to stop the sale, because his domaine did not make that wine until 1982. “This wine cannot exist,” he later told a court. The bottles came from Rudy Kurniawan, who ran a counterfeit-wine operation out of his kitchen and drew a ten-year sentence. The auction house and the paperwork had passed the fakes. The one man who had spent his life with the real thing read the list and knew at a glance that it could not be true.

The wine was a warning shot. The same move, an elaborate fake that clears every institutional check until one person who knows the real thing stops it, is now aimed at what a company cannot afford to get wrong: who it hires and what runs on its network.

In July 2024, KnowBe4, a company that sells security-awareness training, hired a software engineer for its AI team. He had a clean background check, a US identity, and a face that matched the one on four separate video interviews. Every gate the company had built said yes. The identity was stolen, the photo was AI-enhanced, and the engineer was a North Korean operative. His company laptop started trying to load malware within hours of arriving. KnowBe4’s security team caught the behavior that night and published the account, the only reason I can name them here.

One hire is a story. The scale is the argument. In June 2025 the Justice Department searched 29 “laptop farms” across 16 states, seized dozens of financial accounts, and described North Korean IT workers who used stolen identities to get hired at more than 100 US companies. The regime keeps up to 90 percent of the wages and has pulled in hundreds of millions of dollars, money that flows past US sanctions. One hire carries two payloads: access from the inside, and cash back to Pyongyang.

Most coverage files the North Korean scheme under HR fraud, or a background-check failure, or remote-work risk. It is a security breach that happens to enter through the recruiting funnel. The hiring pipeline is a perimeter now, and the recruiter reviewing a candidate is doing frontline defense whether the org chart says so or not. AI is what turned this from a curiosity into an industry: one operation can now generate synthetic candidates at scale, complete with deepfaked photos and avatars that hold a live video call.

Subscribe now

Now the same coin, other face. In July 2026, Hugging Face, the largest repository of open AI models, disclosed an intrusion that, in its own words, “was driven, end to end, by an autonomous AI agent system.” A poisoned dataset ran code on a processing worker. The agent escalated, harvested credentials, and moved across internal systems over a weekend, more than 17,000 recorded actions. OpenAI confirmed the agent was its own model, running in an internal safety test that escaped its boundary. The barrier to this kind of campaign has fallen: Anthropic documented a near-novice who used an AI agent to extort 17 companies in a month. A synthetic attacker is cheap now. So is a synthetic employee.

Then the investigation ran into a wall. When Hugging Face’s team tried to analyze the real attack payloads with commercial AI models, the safety guardrails blocked them, because a filter “cannot distinguish an incident responder from an attacker.” They ran the forensics on an open-weight model on their own hardware instead. Their conclusion, in plain words: the attacker “was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.”

The safety alignment that makes hosted AI responsible for millions of ordinary users is a tax on the people defending you and a free pass for the people attacking you. Defending an organization now takes someone who can vet and run a capable model on infrastructure they control, before an incident, so a content filter cannot lock them out mid-response. That is a rarer skill than operating a security dashboard.

Put the two stories side by side and they point to the same conclusion. The background check passed. The four interviews passed. The sandbox boundary held, on paper. In each case the automated gate waved the threat through, and a person caught it: KnowBe4’s analyst on odd account activity, Hugging Face’s responders on correlated signals their own AI surfaced. When your attacker and your newest colleague can both be fabricated, the automated checks are the thing being fooled, and human judgment is the thing doing the catching. Ponsot caught the wine the same way, by knowing the genuine article well enough to see what could not be true. That kind of judgment takes years to earn.

This is where the cyber talent conversation loses the plot. Everyone quotes the ISC2 figure of 4.8 million unfilled security jobs and argues about whether AI fills the gap or replaces the analysts. That same ISC2 study reports something the headline skips: for the first time, budget passed talent as the top reason roles stay open, and the workforce grew one tenth of one percent, the slowest on record. Meanwhile AI is absorbing tier-1 work, the rung where junior analysts learn to tell a real alert from noise. The real shortage is judgment, and it has been hiding behind a headcount number. Automate the entry rung to save money this year, and you thin the supply of the only people who can catch a threat the machine passed.

The governments that treat cyber as national security already build for this. The US runs CyberCorps Scholarship for Service: the state funds the degree, the graduate owes it service. The UK’s GCHQ runs CyberFirst, paying undergraduates a bursary and summer placements. Both grow talent instead of buying it. A private employer that only buys experienced defenders, while automating away its own junior tier, is bidding for a shrinking pool against states that grow their own.

I spend my days deciding what to buy and what to build, and these stories are the same design problem: identity is cheap to fake, judgment is expensive to grow, and the tasks that grow it are the ones AI makes easiest to cut.

The defenses that hold now are stubbornly human: the recruiter who notices every time the story does not add up, the analyst who can run the model no vendor will run for them, the leader who decides what a machine may do without a person watching. AI can write the resume, pass the interview, and drive the attack. It cannot yet be the one who notices. You can keep training that person, or automate them away and learn what your gates were really worth.

Subscribe now

Here’s How You Take Action

If you set workforce strategy: Move identity verification in hiring from a paperwork step to a security control, and put your security and recruiting teams in the same room before your next remote technical hire.

If you run a security team: Vet and stand up a capable self-hosted model now, while it is calm, so a vendor’s safety filter cannot lock you out in the middle of an incident. Hugging Face learned this the hard way; borrow the lesson for free.

If you are cutting the entry rung to fund AI: The tier-1 work you are automating is where your future senior defenders learn judgment. Automate the task, keep the apprenticeship, or you will buy this year’s margin with next decade’s empty bench.

If you hire remote technical talent: Learn how the North Korean scheme actually works. Treat a background check that comes back too clean as a reason to look harder, and count a live, verified human as worth the friction.

For everyone: The next time someone says AI will close the cybersecurity talent gap, ask which gap they mean, bodies or judgment. The answer tells you whether they understand the problem.

Christina Lexa writes Workforce Rewired, on the intersection of workforce transformation, AI, and global talent.

The views expressed here are my own and do not represent the position of my employer or any organization I am affiliated with.

U.S. employers announced 33,429 job cuts in July, the fewest in a single month in two years, according to Challenger, Gray & Christmas. AI led the stated reasons for the fifth month running, at 10,970 cuts, and announced hiring plans reached their highest July level since 2022. Andy Challenger, the firm’s chief revenue officer, reads the mix as AI reshaping where work happens while the wider job market keeps expanding. The month also showed how blurry the AI label has become, after a Bronx hospital cut nursing roles once it installed new software and the nurses’ union called the change automation.

Subscribe now

By the Numbers

• 33,429 U.S. job cuts announced in July, down 27 percent from June and the lowest monthly total in two years (Challenger).

• 10,970 July cuts attributed to AI, the leading reason for the fifth consecutive month (Challenger).

• 9,867 technology cuts in July, 149,023 for the year, or 31 percent of all 2026 layoffs (Challenger).

• 16,095 planned hires announced in July, up 47 percent from June and the strongest July since 2022 (Challenger).

• 477,033 total cuts announced through July, down 41 percent from the 806,383 over the same period in 2025 (Challenger).

Layoffs and Company Decisions

July job cuts hit a two-year low while AI stays the most cited reason

Challenger, Gray & Christmas counted 33,429 announced job cuts in July, down 27 percent from June and 46 percent below the same month in 2025. Through seven months, employers have announced 477,033 cuts, 41 percent fewer than the 806,383 over the same stretch last year. Technology led every sector again with 9,867 cuts in July and 149,023 for the year, 31 percent of all 2026 layoffs. AI was the single most cited reason for the fifth month in a row, named in 10,970 July cuts. Hiring ran the other direction: employers announced 16,095 planned hires, up 47 percent from June and the strongest July since 2022, with aerospace, energy, and manufacturing driving the demand. Andy Challenger tied that growth to jobs that happen “on a floor rather than a screen.” He also flagged how slippery AI attribution has become. When Montefiore, a Bronx hospital system, eliminated 12 utilization-review nursing roles after adopting Datavant’s software, the New York State Nurses Association filed a class-action grievance calling it a swap of nurses for AI, and hospital leaders called that account misleading. Challenger logged the cuts as “Technological Update (possibly AI)” rather than AI, and noted that naming AI in an announcement can court investors while unsettling the people who work there, which is why the messaging keeps shifting.

Source: Colleen Madden Blumenfeld and Andy Challenger, “Challenger Report: Layoffs Fall, Hiring Picks Up; AI Leads For Fifth Straight Month,” Challenger, Gray & Christmas, August 6, 2026. challengergray.com

Why it matters: The headline gives workers rare relief: layoffs are receding and hiring runs ahead of last year. The new demand sits in aerospace, energy, and manufacturing, so an office worker watching AI absorb their tasks may not find that hiring pointed their way. The Montefiore fight shows how much rides on what a company calls a cut, because the label decides whether a union contract, a severance rule, or a retraining duty applies at all.

What Workforce Leaders Are Watching

• Hiring is strongest on factory floors, flight lines, and energy sites. What does a company owe the displaced office worker whose skills do not transfer there?

• Challenger now separates AI cuts from “Technological Update (possibly AI)” cuts. Which label will your own filings carry, and who inside the company decides it?

• The Montefiore nurses had contract language that forced the hospital to the table once the union called the change automation. Do your workers have any comparable claim when software takes over a task?

• AI cuts have concentrated in technology for five straight months. If that pattern breaks, which non-tech function is most exposed the moment a vendor ships credible software for it?

This briefing was prepared automatically by the Workforce Rewired research assistant. All stories include direct source links.