Companies have slowed entry-level hiring in the roles most exposed to AI, and senior engineers now do the work those hires used to do. IT absorbs the extra work, HR controls the hiring budget, and neither department measures both.
Companies track AI’s effect on jobs with two tools: total headcount and executive judgement. Both work well when a change spreads evenly across the business. Neither one catches a change confined to a single group.
Three datasets published within six months of each other look like they contradict each other. All three are accurate. They agree once you see what companies are actually doing. They have not cut large numbers of current staff. They have sharply slowed hiring for entry-level workers in the roles most exposed to AI.
The work moved rather than disappeared. Junior staff used to write routine code and handle first-pass tasks. AI does that work now, and senior staff review and correct it. Senior engineers spend more of the week checking output and less of it building, and the junior colleagues who used to take on that checking are not being hired.
Two departments hold the two halves of the problem. Engineering feels the extra review load. HR controls the hiring budget that would relieve it. Seven in ten CEOs leave HR out of AI governance entirely, so nobody holds a number that shows both halves at once.
AI is not producing large company-wide layoffs. It is removing entry-level positions in exposed roles while adding work to senior staff, and standard workforce metrics register neither movement.
Two large studies looked for AI’s effect on employment across whole companies in 2026. Both found almost nothing. Every workforce plan built on headcount and executive input is reading this result.
Nine in ten executives said AI had changed neither productivity nor headcount at their company over three years. Their companies do use AI: 69% of them have it in place. The executives answering the question used AI themselves for about an hour and a half a week.
Employers and employees do not even agree on which way it goes. Employers in the sample predicted AI would cut employment by 0.7%. Employees in the matched sample predicted it would raise employment by 0.5%.
Only 18% of US companies use AI at all. Among those that do, 57% run it in three or fewer departments and 65% apply it to three or fewer worker tasks. Two-thirds use it to help staff do existing work rather than to replace them.
Two percent of companies have cut any jobs because of AI. Putting AI into a worker’s daily tasks shows no measurable link to that company shedding headcount.
Read together, these two studies describe an economy where AI use is real, shallow, and not costing jobs. That answer is correct for the question both studies asked, which was about whole companies. This brief asks what happens one level down.
Stanford looked at the same labour market and asked a narrower question: what is happening to young workers in the jobs AI touches most?
Brynjolfsson, Chandar and Chen published the August 2026 revision of “Canaries in the Coal Mine?” on 12 August. They used ADP payroll records covering workers aged 22 to 70 through June 2026, sorted by how exposed each occupation is to AI.
Companies now employ 19% fewer workers aged 22 to 25 in the most AI-exposed jobs than the trend predicts. Eleven months earlier that number was 15%. The gap is still growing.
The August revision adds one more finding: young women are more exposed than young men.
A change with those four features cannot reach any report a company already runs. Nothing happens that anyone has to file. The only place it shows up is the hiring rate for one age group in one set of jobs, compared against what that rate used to be.
PwC read more than one billion job advertisements across six continents for its 2026 Global AI Jobs Barometer, published 16 June. Postings that require specific AI skills grew 69% while the wider job market grew 9%. Jobs built around human judgement grew twice as fast as average and pay rose 42% faster in them.
One PwC finding connects directly to Stanford. Employers now ask for senior-level skills in AI-exposed junior roles seven times more often than before. A company that wants to fill an entry-level position with someone who will grow into the work no longer has that option in these roles.
While companies hire fewer juniors, the work those juniors handled is growing.
Hanover Research surveyed 200 US technology decision-makers at manager level or above for New Relic’s 2026 State of AI Coding Report, published 10 June. Every respondent already uses generative and agentic AI in software engineering. New Relic funded the report, so weight the direction rather than the exact numbers. The pattern inside it is hard to explain away.
94% of these managers say AI-written code looks good when they review it. 78% also say production incidents have gone up. 74% say at least a quarter of AI-written code needs substantial rework, and 62% say their teams ship it without anyone reading it line by line. 82% had at least one live failure traced to AI code in the previous six months.
One number carries the workforce consequence. 86% say their senior engineers now spend more time fixing code.
Code that passes review and then breaks in production has to be found and repaired, and finding it takes the judgement that comes with experience. Junior engineers used to do the first pass of that checking. Companies are hiring fewer of them, so senior engineers do it instead.
Block told shareholders on 7 May that each engineer now ships more than 2.5 times as many code changes as a year ago, and that incidents after those changes fell by more than 70%. Every employee was using AI tools by early April 2026. Block’s own tooling reviewed over 90% of production code change requests, and AI made 15% of production code changes with little human involvement.
Block also cut its headcount from over 10,000 to under 6,000 in February 2026. The figures are self-reported and unaudited, and the productivity gain cannot be separated from that cut. A board will raise this example, and any workforce argument that cannot answer it will fail.
| Tool | What it counts | What it catches | What it misses |
|---|---|---|---|
| Headcount forecast | Total positions at the company | Layoffs, resignations, requests to backfill a roleCensus CES-26-25: cuts at 2% of firms | Hiring for one age group falling below trend while total headcount holds steady |
| Executive surveys | What leaders believe is happening | Changes big enough for an executive to noticeNBER 34836: ~9 in 10 report no effect | Changes in one group that never reach an executive’s desk |
| Payroll records by occupation | Age group against job exposure | Hiring rates falling away from trend, by age and by job typeStanford: 19% fewer 22 to 25s in exposed jobs | Where the work went inside the company, and who is doing it now |
| Engineering delivery logs | Incidents, rework, review hours | How much repair work there is and who is doing itNew Relic: 86% report seniors spending more time on fixes | That the junior engineers who used to do that repair work were never hired |
| AI tool inventories | Approved software only | What staff run inside sanctioned systemsGartner: 37% can see what AI tools are in use | The personal tools where staff report the largest time savings |
Gartner surveyed more than 12,000 enterprise employees and managers in the first quarter of 2026 and published the results on 13 May. Only 37% of companies know which AI tools are running in their own organization. Staff who run their own tools alongside the approved ones save 1.7 times more time than staff who stick to approved tools.
88% of employees have access to a company AI tool, but the heaviest and most productive users are managers and executives, who make up 73% of that group.
That produces two problems. First, when a company measures what AI is doing for productivity, it is mostly measuring its senior people, the same people an entry-level hiring freeze does not touch. Second, the staff getting the best results are getting them on tools the company cannot see, so there is no record of how they did it for HR or IT to teach anyone else.
Security researchers find the same behaviour. BlackFog surveyed 2,000 workers at companies with 500 or more employees and reported on 29 January 2026 that 49% use AI tools nobody approved, and 51% have connected an AI tool to a work system without asking IT. Saviynt asked 235 security leaders at companies with 5,000 or more employees and found that 75% have already caught unapproved AI running in production.
If AI’s effect on jobs is concentrated in one age group, shows up as slower hiring rather than layoffs, and creates extra repair work in a department that does no hiring, which leader is holding a number that would show all three at once?
Three research groups worked separately, used different data and asked different questions. Read together, their findings describe one thing happening inside companies that none of them describes on its own.
Census and NBER counted whole companies. Stanford counted one age group inside one set of jobs. When a change is severe for a narrow group and absent everywhere else, those two methods produce exactly this result: a strong signal from Stanford and almost nothing from the other two.
Both answers are correct. A company reading only its headcount total is asking a question that cannot return this finding.
New Relic shows senior engineers spending more time fixing AI-written code. Stanford shows companies hiring far fewer of the young workers who used to do the first pass of that checking. PwC shows why a company cannot simply hire a junior and train them: employers now ask for senior skills in those same junior roles seven times more often.
Each study describes one side of a handover. None of them describes the handover.
BCG asked 152 chief executives at companies with at least $500 million in revenue who sits on their AI governance body. 82% include technology leaders. 30% include HR. That is not a question of seniority or interest. It decides which department ever sees the number.
Gartner found the same split written into the plans themselves. Making AI operational is the second priority on the CIO list, and 67% of CIOs plan to spend on AI and machine learning. Their published spending breakdown covers technology categories only, with no line for talent, training or change management.
On the HR side, AI strategy climbed to the second priority and Change Management and Workforce Resiliency moved from fifth to third. HR is planning the people work. Technology is holding the money.
Two further surveys measured the disagreement instead of inferring it. CompTIA asked 1,049 HR and learning professionals in April 2026 and found 31% believe the training budget sits inside HR, against 19% of IT leaders who think so. Grant Thornton asked 950 business leaders between 23 February and 18 March 2026 and found 58% of CIOs and CTOs call the workforce fully ready for AI, against 12% of COOs.
Neither pair is arguing with the other. They have never compared numbers.
When work moves from one department to another, and each department only counts its own half, nobody sees the move.
None of these is settled practice yet. Each one appears in 2026 primary material, and each is a different answer to the question of who owns the problem.
Rather than making technology and HR settle ownership between them, the CEO hires a Chief AI Officer and gives the job to that person.
Evidence76% of organizations now have a Chief AI Officer, up from 26% a year earlier. 77% say talent and technology leadership are merging, and 59% of CEOs expect HR influence to rise. IBM Institute for Business Value with Oxford Economics, 4 May 2026, n=2,000 CEOs across 33 geographies.
Every AI agent gets a named human owner and a directory record, the same as an employee. When that person leaves, someone has to be assigned their agents, which makes supervising agents part of a real job description.
EvidenceMicrosoft Entra Agent ID (generally available, documentation dated 1 May 2026) issues directory identities with owners, sponsors and managers, plus workflows for when a sponsor changes role. Okta shipped agent-to-agent connection policies and combined human-and-agent access reviews on 22 July 2026. KPMG India, 30 June 2026, names four workforce tiers: builders, agent managers, system managers, front-liners.
Technology and HR run workforce readiness together from the first day of a deployment, against shared numbers, instead of technology installing the software and handing training to HR afterwards.
Evidence35% report HR and IT making AI decisions jointly and 48% report shared goals. Companies with joint ownership act on workforce changes 13% faster and are 67% more likely to say they can decide at the pace AI demands. Cornerstone OnDemand with Vanson Bourne, 29 July 2026, n=2,000 across eight countries. Vendor-published; read the direction, not the magnitude.
Staff run their own AI tools alongside the approved ones and get better results. The company cannot see the tools, so it cannot copy what works or govern what does not.
EvidenceOnly 37% of companies can see which AI tools staff use, and staff mixing personal with approved tools save 1.7× more time (Gartner, 13 May 2026, 12,000+ employees and managers). 49% of workers use unapproved AI tools (BlackFog, 29 January 2026, n=2,000). 75% of large enterprises have caught unapproved AI running in production (Saviynt CISO AI Risk Report 2026, n=235).
The company tells the market it is cutting staff because of AI, in a regulatory filing, before it has published any new job structure describing what the remaining people will do.
Evidencemonday.com Ltd., Form 6-K filed 22 July 2026: roughly 20% of the workforce cut to align the company around its AI Work Platform, with a $45M to $55M restructuring charge. Verizon CEO Dan Schulman, Bloomberg Tech Conference, 4 June 2026, cited three months of agent trials with a 1,280 basis point rise in customer satisfaction.
The budgets being written now decide how many experienced engineers a company will have at the end of the decade.
A company that hires fewer graduates in the 2027 budget notices nothing that year. The consequence arrives two to four years later, when those people would have become the engineers who review work and handle the difficult cases.
By 2029, when the shortage of mid-level staff finally shows up in a report, the decision that caused it is several budget cycles old. No company can undo it inside the same planning horizon.
There is money in the 2027 budget. 82% of technology leaders expect a larger budget next year, which makes this the cycle to fund a shared capacity line.
Forrester, 15 Jul 2026, 2,600+ decision-makers
CEOs expect HR to matter more. 59% say the HR chief’s influence will grow as talent and technology decisions merge.
IBM IBV with Oxford Economics, 4 May 2026, n=2,000
You can now put a name against every AI agent. Microsoft and Okta both shipped directory tools that assign agents a human owner, which creates a workforce record that did not exist a year ago.
Microsoft Entra Agent ID; Okta, 22 Jul 2026
Disclosure rules are starting. From 1 October 2026, employers covered by WARN in Connecticut must say whether a mass layoff relates to their use of AI.
Connecticut SB 5, signed 27 May 2026. Enrolled text not yet read; confirm the exact wording before relying on it.
A total headcount figure cannot show you this. Track how many people you hire into each age band, in the roles where AI does the task rather than assisting with it, and compare that against what you used to hire. Stanford’s method is published and runs against your own applicant and payroll records.
This is a reporting change, not a policy change. No company can decide what to do about a hiring freeze it has not yet found in its own numbers.
Measure the hours senior engineers spend reviewing and repairing AI-written work, and carry that as a cost line rather than absorbing it. With 86% of technology leaders reporting that time rising, and 74% reporting at least a quarter of AI code needing substantial rework, there is enough of it to budget for.
Give HR a seat and a budget line on the AI governance body, and have both leaders sign off on one set of workforce numbers before the budget closes. Companies that own this jointly act on workforce changes 13% faster. That figure comes from a vendor, so use the direction and treat the size as unconfirmed.
Cutting graduate intake can be the right call. Cutting it by default across every exposed job family is not. The question to answer out loud is whether you intend to buy experienced staff on the open market in 2029 at three to four times average pay, or grow them yourself starting now.
Both are legitimate answers. Only one of them is currently being chosen deliberately.
Absorption capacity is the maximum amount of operational, cognitive and cultural change an organization can absorb at one time without causing fragmentation, burnout or workflow collapse. It is assessed across five components: People, Processes, Governance, Technology and Data.
Score your company on each of the five questions below, from 0 to 4. A 0 means nobody measures it. A 4 means it is measured, reported, owned by a named person, and reviewed by technology and HR together. Each question maps to a capability band of the Intelligence Method™.
| Component | Band | The question to answer | Where to look | Score 0 to 4 |
|---|---|---|---|---|
| 1. People | Band 2 — People & Purpose | Can you say how many 22 to 25 year olds you hired into AI-exposed roles this year, and how that compares with three years ago? | Applicant and payroll records, last eight quarters | ☐ |
| 2. Processes | Band 3 — Operational Integration | Does anyone count the hours senior staff spend reviewing and fixing AI output, and does that number reach a capacity plan? | Code review logs, ticket rework data, delivery telemetry | ☐ |
| 3. Governance | Band 4 — Adaptive Governance | Does HR hold a vote and a budget line on AI deployment decisions, or only get told afterwards? | Governance charter, committee membership, budget sign-off | ☐ |
| 4. Technology | Band 1 — Right-Fit Technology | Can you list the AI tools your staff actually use, including the ones nobody approved? | Software inventory, measured against the 37% who can | ☐ |
| 5. Data | Cross-band | Do technology, HR and operations plan from the same workforce readiness numbers? | The three readiness reports, placed side by side | ☐ |
If you scored 0 or 1 on People or Processes, your company cannot currently see this change in its own numbers. That is a measurement problem before it is a workforce problem, and you can fix it inside one reporting cycle.
If you scored well on the individual components but low on Data, each department is measuring carefully and separately. That is the pattern that produces one leader saying 58% ready and another saying 12%, and it will not resolve by either department measuring harder.
A focused 90-minute executive working session applies this diagnostic to your own hiring records, delivery telemetry and governance documents. The output is a single capacity metric that IT, HR and Operations can plan against together.
Schedule a ConversationThis research is the foundation for our AI and workforce planning executive workshop series.
This brief synthesizes 27 sources across five categories.
Methodology and source treatment. Vendor-funded and advisory-funded sources are named as such at the point of use, so the reader can weight them. Where a figure is directional rather than measured, this brief says so. Two items in the horizon and external factors sections rest on analyst guidance with no published sample, and are marked. The Connecticut SB 5 timeline comes from legal analysis; the enrolled statutory text has not been read and should be verified before any formal compliance decision.