Across roughly 10,000 earnings calls between 2021 and 2025, executives talking about AI were uniformly optimistic โ and that optimism explained none of the variation in their firms' productivity. What did track productivity was the thing nobody puts on a board slide: employee AI sentiment, extracted from millions of Glassdoor reviews at the same firms over the same years (Ding, Ma, Wu & Yang, SSRN, 2026).
One signal is loud, free, and worthless. The other is quiet, unmeasured in most mid-market companies, and the only one with a demonstrated relationship to output.
If you are forecasting your AI return off the confidence of the people who approved the budget, you are reading the wrong instrument.
What the Pittsburgh panel actually measured
The paper โ Tracking Artificial Intelligence Sentiment in the U.S. Labor Market, a 116-page working paper from researchers at Pittsburgh's Katz school and Michigan State โ builds a firm-level panel from four data sources over 2021โ2025: millions of Glassdoor employee reviews, about 10,000 earnings-call transcripts, individual LinkedIn career profiles, and firm-level job postings (Ding, Ma, Wu & Yang, SSRN, 2026).
Three results matter for anyone running operations.
First, employees are measurably more negative about AI than they are about their employer overall. The gap is specific to AI. The same reviewer who rates the company favorably rates the AI program below that baseline โ so this is not general dissatisfaction leaking into an AI question.
Second, employee AI sentiment shows a strong positive association with firm productivity. Higher sentiment, higher output.
Third, and most usefully for a Head of Operations deciding what to instrument: management AI sentiment does not explain productivity. It is uniformly optimistic across firms that went on to perform well and firms that did not, which is precisely what makes it useless as a signal. A variable with no variance cannot predict anything.
Why executive optimism has no variance to give
This is not a claim that executives are lying. It is a claim about what earnings-call language is for.
An earnings call is a positioning exercise with a legal team attached. The CEO who says "our AI investment is not yet returning" pays an immediate market price for the sentence. The result is a corpus in which nearly everyone sounds confident, which strips out the differences you would need to distinguish a working program from a stalled one.
The survey data shows the same compression from a different angle. BambooHR's State of the Workforce 2026, covering 1,200-plus employees and business leaders across six industries, found 81% of leaders reporting a productivity increase from AI while 49% of respondents said AI had not delivered tangible value and was overhyped (BambooHR, 2026). Those two numbers describe the same companies. The leadership layer is not observing a different reality โ it is reporting a different one.
There is a structural reason the executive read runs high, and it is not vanity. Leaders see the pilot demo, the vendor benchmark, and the aggregated dashboard. Employees see the rework, the verification step, and the workflow that was never redesigned around the tool. Both are looking at the AI program. Only one is looking at the part where the productivity would have to come from.
If your AI business case is built on how confident your leadership team feels, you have built it on the one variable in the dataset with no explanatory power.
Job security is the top driver โ and layoff attribution is the amplifier
The paper ranks what pushes employee AI sentiment down. Job-security concern is the strongest negative driver, ahead of poor training, thin reskilling paths, weak AI leadership, and plain doubt that the tools work (Ding, Ma, Wu & Yang, SSRN, 2026).
Then comes the finding with the most direct operational consequence: AI-attributed layoff announcements sharply depress employee AI sentiment at the announcing firm โ offsetting the adoption gains the cuts were meant to bank. And the market reaction to those announcements was negative or approximately zero for more than half of the events studied.
Read that as a pricing error. When a company frames a reduction as "AI made these roles unnecessary," it is choosing a narrative it believes is free. It is not free. It buys a measurable decline in the sentiment variable that correlates with productivity, in exchange for a market reaction that, more than half the time, does not arrive.
The Forrester data on the other side of that decision is unflattering. Fifty-five percent of employers say they regret laying off workers for AI, and roughly half of AI-attributed layoffs are expected to be quietly reversed in 2026 โ rehired offshore or at materially lower wages (HR Executive on Forrester Predictions 2026). So the sequence for a meaningful share of firms is: cut, attribute the cut to AI, absorb the sentiment hit, get no market credit, then rehire.
The macro picture says most of that cutting was never justified by aggregate headcount math in the first place. The Fed's CFO Survey panel โ 603 responses, plus 145 supplemental, fielded November 2025 through January 2026 โ found a negligible AI effect on employment in 2025 and an expected 2026 effect near zero in aggregate. The movement is compositional: an expected 0.76% decline in the routine-clerical share in 2026, 2.19% by 2028, partly offset by growth in skilled technical roles (NBER Working Paper 34984, 2026).
Composition shifts do not require an announcement. They require a hiring plan.
The retention finding nobody prices
One more result from the panel deserves its own line in your workforce model: firms with more AI-skilled employees show higher turnover, a lower hire rate, and slower employment growth. The authors checked this against the obvious confound โ post-COVID over-hirers correcting โ and it holds when those firms are excluded, so it is not mainly an artifact of AI-washing (Ding, Ma, Wu & Yang, SSRN, 2026).
Alongside it: a rising wage premium and faster upward mobility for AI-skilled workers.
Put those together and the mid-market exposure is specific. The people you spent 2025 training into AI fluency became more mobile in a market that now pays them more, inside a firm that is hiring more slowly. Your AI enablement program has a retention bill attached, and it comes due at exactly the seniority band where replacement is hardest.
The counterargument worth taking seriously
This is a working paper, not a peer-reviewed publication, and the central relationships are observational. No one randomized employee sentiment. The association between sentiment and productivity is consistent with reverse causation โ productive firms are pleasant places to work, and people write nicer reviews about them โ and with an omitted third factor, competent management, that produces both.
That objection is correct, and it does not change what you should do.
You are not being asked to believe that improving sentiment causes productivity. You are being asked to notice that one measurable variable moves with firm performance and another does not. A leading indicator does not need to be causal to be useful; it needs to be informative and available early. Employee AI sentiment is both. Executive AI sentiment, on this evidence, is neither.
The defensible position is not "we must fix morale to get AI returns." It is "we are currently forecasting our returns off a signal with zero demonstrated predictive power, and the alternative signal costs almost nothing to collect."
What to instrument: employee AI sentiment as your leading indicator
Measure employee AI sentiment separately from engagement
Your existing engagement survey will not surface this. The whole point of the Pittsburgh finding is that AI sentiment runs below overall firm sentiment โ pooling them hides the gap. Add three or four AI-specific items to the instrument you already run: is the tooling making your work better, do you have the training to use it, do you understand what it means for your role. Track the delta between AI sentiment and overall sentiment as its own number.
Split it by function and by tenure, and read the dispersion
An organization-wide average will look fine and tell you nothing. The signal lives in which teams have gone negative โ and, given the turnover finding, in whether your most AI-fluent people are the ones going negative. Those two cuts are the report.
Audit how this year's restructuring was attributed
Go back through every reduction of the last twelve months and check what employees were told about why. If AI was named as the cause anywhere it was not the actual driver, you paid a sentiment cost for a narrative that bought you nothing. Fix the language going forward: attribute to the business reason, not the technology.
Fix the drivers in rank order, not the ones easiest to fund
Job security ranks first, ahead of training. Most enablement budgets invert this โ heavy on tooling and prompt workshops, silent on what happens to the role when the workflow changes. A clear, specific statement about what AI does and does not mean for a given job costs nothing and addresses the top-ranked driver. Training addresses the second. Buy them in that order.
Then widen the use cases, not the seat count
Gallup's Q2 2026 workplace data found the share of employees reporting a positive productivity effect rises from 45% among those with one or two AI use cases to 90% among those with seven or more (Gallup, 2026). Sentiment improves when the tool actually helps. Breadth of applied use, not licenses issued, is what moves it.
The decision on your desk
You already run an employee survey. You already hold quarterly business reviews where someone presents an AI progress slide built from vendor benchmarks and leadership confidence. The change is not a new budget line โ it is which of those two artifacts you treat as evidence.
Before your next AI investment review, put one number in front of the room: employee AI sentiment, split by function and tenure, with the gap to overall sentiment shown. If it is lower than anyone in the room expected, that is not a morale problem to route to HR. That is your productivity forecast, arriving early enough to act on.
