Last year, 32 percent of respondents told McKinsey they expected AI to reduce their organization's workforce. This year, asked what actually happened, 14 percent of respondents at AI-using organizations reported that AI contributed to a decline in workforce size โ "less than half the 32 percent who, in last year's survey, expected workforce reductions over the same period" (McKinsey, 2026). Two-thirds reported little or no AI-related change in total employment at all.
That is the first time this survey has been able to grade its own AI workforce forecast against a realized outcome. The forecast missed by roughly a factor of two.
Then the panel raised it. Thirty-nine percent now expect AI-related declines in total employment over the coming year, up from last year's 32 percent. The estimator that overshot by 2x has been revised upward, not downward, and a large number of 2027 operating plans are quietly resting on it.
What the Survey Actually Measured
The 2026 edition of The State of AI surveyed 1,719 participants across 97 nations, fielded 4 May to 8 June 2026. Thirty-six percent work at organizations above $1 billion in annual revenue, which means roughly two-thirds sit below it โ unusually relevant for a mid-market reader, though the panel still skews larger than a 200-FTE operation.
The context matters because adoption is not the variable that failed. Eighty-nine percent use AI regularly in at least one business function. Forty-four percent say it is scaling enterprise-wide, up from 38 percent. Eighty percent report improved individual productivity (McKinsey, 2026).
So the sequence is not "AI didn't arrive, so headcount didn't fall." AI arrived, productivity was reported, and headcount still didn't fall. Meanwhile the share attributing at least some EBIT impact to AI sat flat year over year at 37 percent, and the population of genuine high performers held at about 6 percent of all respondents.
Diffusion moved. Realized workforce reduction did not follow it.
The Realized Data Agrees With the 14 Percent
One survey grading itself is interesting. Two independent datasets pointing the same direction is a planning input.
The National Bureau of Economic Research surveyed nearly 750 corporate executives through The CFO Survey panel and found "little evidence of near-term aggregate employment declines due to AI, though larger companies anticipate AI-driven workforce reductions, while smaller firms expect modest gains" (NBER, 2026). The aggregate effect is close to zero, and the dispersion is the finding: large firms expect AI to cut their employment by around 0.8 percent in 2026, while smaller firms expect small increases. What does move is composition โ routine clerical roles declining, relative demand for skilled technical roles rising, within firms and across them.
Announced-layoff data agrees from a third angle. Through September 2026, U.S. employers announced 573,195 job cuts, down 39 percent from the 946,426 announced in the first nine months of 2025 โ the lowest January-to-September total since 2022. AI was cited in 120,136 announcements year to date, approximately 21 percent of all cuts, which does make it the leading stated reason for the year (Challenger, Gray & Christmas, 2026).
Read those two numbers together, because the second is the one that gets quoted alone. AI leads the reason table inside a shrinking total. In September specifically, AI was the fifth-most cited reason, at 3,961 cuts or about 9 percent of the month. Third-quarter cuts fell 43 percent from the second quarter.
A rising share of a falling denominator is not a wave. It is a composition shift inside a cooling market.
Why the AI Workforce Forecast Errs in One Direction
Forecast error that consistently overshoots is usually structural, not random. Forrester's Predictions 2026: The Future of Work names the mechanism: cuts made against AI capabilities that do not exist yet. Fifty-five percent of employers report regretting layoffs made for AI, and Forrester predicts half of AI-attributed layoffs will be quietly rehired โ offshore or at materially lower salaries (HR Executive, 2026).
Note the asymmetry in that last clause. The reversal does not appear as a correction in anyone's plan. It appears as offshore spend, contractor lines, and re-recruiting cost, none of which reconcile against the original headcount saving. The forecast stays unfalsified in the internal record even when it has already failed.
Two further details from the same research sharpen the cost. The share of disengaged employees Forrester calls "coasters" was 27 percent in 2024, 25 percent in 2025, and is expected to reach 28 percent in 2026 โ a figure that moves when survivors watch premature cuts. And only 16 percent of workers scored high on Forrester's AI-readiness measure in 2025, with the highest readiness concentrated in the cohort whose entry-level roles are being eliminated first.
The Planning Error, Stated Plainly
The error is not optimism about AI. It is booking anticipated workforce reduction as a realized baseline, one year early, and then building the backfill freeze on top of it.
Anticipated change is a forecast with a now-documented one-year error rate. Realized change is a measurement. Only one of them belongs in a plan that commits real money.
The Honest Counter
Three limits, stated before anyone else states them.
This is self-reported survey data, not payroll. Nobody audited these organizations' headcount ledgers. Respondent-level attribution of a headcount change to AI specifically is judgement, and attribution is exactly where self-report is weakest. The same caveat applies to the NBER figures, which are executive expectations rather than realized payroll outcomes.
The panel is not identical year over year. McKinsey states it checked respondents who completed both the 2025 and 2026 waves and that the findings reflected the overall sample, but a changing panel is a real limit on reading 32 percent and 14 percent as the same people contradicting themselves.
The expectation may simply be early rather than wrong. Capability lags announcement, and deployment lags capability. A reduction forecast for 2025 that materializes in 2027 was mistimed, not mistaken โ and the NBER composition data, with routine clerical share expected to decline further by 2028, is consistent with a slower, more selective version of the same direction. The defensible claim is about timing confidence, not about direction.
What survives all three is narrower and still actionable: the distribution of realized outcomes is heavily concentrated on "little or no change," and a plan that treats the tail as the base case is mispriced.
What to Re-Baseline Before the Q4 Plan Closes
Four moves, none of which requires reversing a strategic position on AI.
1. Separate anticipated from realized in the plan document. If your 2027 headcount model contains a line justified by AI-driven reduction, mark whether it rests on measured change in your own organization or on expected change. If it is expectation, label it as a forecast carrying the error rate above. This is a fifteen-minute edit that changes how the number gets challenged.
2. Re-underwrite the AI business case on throughput, not FTE. Cycle time, revenue per employee, error rate, cost per transaction. The NBER productivity finding points the same way: gains showed up as revenue-based total factor productivity tied to innovation and demand channels, not as capital deepening that substitutes for labour. If the case only closes on headcount reduction, the case has not closed.
3. Gate any AI-attributed cut behind demonstrated, not promised, capability. The test is a named workflow running at target quality with a measured error rate over a defined period โ not a vendor roadmap. This single gate is what separates the 14 percent from the 55 percent who regret.
4. Plan for the team you will still have. This is the move most 2027 plans are missing. If two-thirds of organizations saw little or no AI-related headcount change, the base case is that your current roster is roughly your 2027 roster, with composition shifting underneath it. That makes the binding constraint a capability question rather than a cost question: which competencies sit in which seats, where succession is thin, and which routine-clerical capacity can be redeployed into the skilled-technical demand the NBER data says is rising. A backfill freeze decided without that map is not a saving; it is an unmeasured capability loss. Competency and succession measurement is the instrumentation that makes that call defensible โ 9-box position computed from per-competency data rather than manager recall, and succession readiness rated per critical seat. Scovai's performance and succession module documents how that mapping is structured if you want a worked reference.
The uncomfortable part of this data is not that AI underdelivered on headcount. It is that an organization can run a successful AI programme โ 89 percent adoption, 80 percent reporting productivity gains โ and still have the same number of people at the end of it, doing meaningfully different work.
That is a workforce development problem wearing an AI workforce forecast.
Before the Q4 plan closes, find the headcount line that assumes AI will shrink the team, and ask it one question: are we measuring that, or still expecting it? If the answer is "expecting," the 2x error rate is now on the record โ and the burden of proof has moved to the forecast.
