The AI productivity paradox is the gap between the productivity gain people report from AI and the gain that can actually be measured. On March 25, 2026, economists at the Federal Reserve Banks of Atlanta and Richmond and Duke University documented it in a working paper drawn from a survey of nearly 750 corporate executives: perceived productivity gains are larger than measured productivity gains.
That sentence should stop a finance leader. It means the number in the board deck and the number in the results are not the same number, and almost no organization can say by how much.
What is the AI productivity paradox?
The Atlanta Fed paper is not an argument that AI does not work. It reports the opposite in several places. Labor productivity gains are positive, vary across sectors and are expected to strengthen in 2026, with the largest effects concentrated in high-skill services and finance. More than half of surveyed firms had already invested. The gains are not primarily explained by firms buying more capital, but by increases in revenue-based total factor productivity tied to innovation and demand.
The paradox sits alongside all of that. Perceived gains run ahead of measured ones. The authors offer a benign explanation: revenue realizations lag the work.
Why the explanation matters less than the gap
That explanation is plausible. It is also not the only one available, and an executive survey cannot distinguish between them. If perception runs ahead of measurement, the cause could be timing, in which case the gains arrive later. It could be that gains are real but captured somewhere the firm is not looking. Or the perception could be optimistic. All three produce the same survey result.
Read the status of the source before you carry the number
This is a Federal Reserve working paper, not peer-reviewed published research, and working papers are revised. The perceived side of the comparison comes from executives describing their own firms. Treat the direction as credible and well-documented. Treat any specific magnitude as provisional.
Why do the 2026 numbers disagree with each other?
Four of the most-cited 2026 datasets on AI and work point in different directions, and the differences are usually reported as contradiction. They are better read as instrument mismatch.
Pew Research Center surveyed 3,488 US adults between June 22 and 28, 2026, publishing on August 18. Seventy-one percent think AI will lead to fewer jobs in the United States over the next 20 years, up from 64 percent in 2024. Fifty-two percent say increased AI use makes them more concerned than excited, against 37 percent in 2021.
Gallup surveyed 23,717 employed US adults between February 4 and 19, 2026. Eighteen percent said it was very or somewhat likely their own job would be eliminated within five years due to AI or automation, rising to 23 percent among employees at organizations that had adopted AI.
These two numbers cannot be subtracted
Seventy-one and eighteen get placed side by side often, usually to argue that people fear for everyone else's job but not their own. The comparison does not hold. The populations differ, all US adults against employed US adults. The horizons differ by 15 years. And the object differs: one asks about the national labor market, the other about the respondent's own position. Two surveys that were never designed to be compared will not become comparable because both mention AI and jobs.
What can be said is narrower and more useful. Public expectation of labor market disruption is rising, on the one measure that has a clean trend behind it: Pew's 64 percent in 2024 to 71 percent in 2026. Whether it is rising faster than firm-level reality is a question nobody currently has the data to answer, because there is no comparable time series on the firm side.
The Atlanta Fed paper does find little evidence of near-term aggregate employment decline from AI, alongside compositional reallocation, with routine clerical roles declining and relative demand for skilled technical roles increasing. That finding comes with a split the reassuring summaries tend to drop: larger companies anticipate AI-driven workforce reductions, while smaller firms expect modest gains. Composition is moving, and it is moving in different directions depending on where you sit.
Does AI change how work gets done, or how it feels?
Gallup's Q1 2026 data contains the cleanest version of the problem. Among employees at organizations that had implemented AI, 65 percent said AI improved their productivity and efficiency. In the same population, 8 percent strongly agreed that AI had transformed how work gets done in their organization, and 34 percent strongly disagreed.
Gallup's own conclusion is that benefits are concentrated at the level of individual tasks rather than broader systems, and that many organizations have not redesigned workflows, roles or processes around AI.
That sits awkwardly next to the Atlanta Fed's compositional finding, and the tension is worth naming rather than smoothing. Gallup's employees report that roles and processes have not been redesigned. The Fed's executives report reallocation happening both within and across firms. Those can both be accurate if the reallocation is visible from the top of the organization and not from inside a job. They can also both be accurate if one group is wrong. Two self-report instruments, pointed at two populations, cannot settle it.
Both figures are still perception. One asks what employees believe about their own output, the other what they believe about their organization. Neither observes the work.
The eight-fold swing worth sitting with
On August 16, 2026, Gallup published a finding that deserves more attention than it received. Employees who strongly agree their manager champions AI were far more likely to say AI has transformed how work gets done in their organization than those who do not strongly agree: 33 percent against 4 percent.
Gallup reads this as manager support driving real adoption, and their broader research on manager effects supports that reading well. A second reading is available from the same numbers: a perception measure that swings roughly eight-fold on manager sentiment is, in part, measuring manager sentiment.
Two things are true at once here. Supportive managers almost certainly do produce better adoption. And a survey cannot tell you how much of that 33 percent is changed work and how much is changed mood.
The same Gallup article, drawing on a separate survey of 102 global Fortune 500 CHROs fielded between February 10 and March 16, 2026, found that 50 percent were not confident in their managers' ability to guide employees on AI use, while 99 percent called AI important to strategy. That CHRO sample is a self-selected roundtable rather than a probability panel, so treat it as directional. The pattern still stands: the people being asked to drive the transformation are the same people whose enthusiasm colors the main instrument measuring it.
What each instrument can and cannot see
| Instrument | What it captures | What it cannot see |
|---|---|---|
| Public opinion polling | Expectation and sentiment across a national population | Anything about a specific firm or team |
| Employee self-report surveys | What employees believe about their own work and their organization | Whether output, quality or cycle time changed |
| Executive self-report surveys | Leadership's read on firm performance and investment | The gap between that read and the financials |
| Vendor usage telemetry | Seats, logins, prompts and features touched inside one vendor's tool | Work done in every other tool, and outcomes anywhere |
| Observed behavioral signals | What changed in the artifacts of work across the stack | Intent, motivation and anything happening off-system |
No single row is sufficient. The failure mode of 2026 is that most organizations are working from the first four rows and reporting the result as though it came from the fifth.
What measuring the paradox actually requires
The Atlanta Fed's finding is a measurement problem stated in economic language. Closing it does not require better surveys. It requires a second instrument that does not ask anyone anything.
Four elements are non-negotiable:
- A baseline. How the work ran before the capability arrived, captured in the same terms you intend to measure afterward.
- A comparison group. Similar teams that did not adopt, so that a movement can be distinguished from a trend everyone was already on.
- Controls. Tenure, role mix, work type and tool stack, because unmatched cohorts produce confident nonsense.
- A disclosed confidence level. A number without a stated confidence is a claim, not a measurement.
Levos calls this controlled cohort analysis with confidence scoring. It compares adopting teams to non-adopting teams while controlling for tenure, role and tool stack, and it does not claim full attribution. The limits are published rather than buried, because a method that overclaims fails the first time a CFO tests it.
Where the signal comes from
The inputs are behavioral rather than declarative. Levos pulls behavioral signals from Microsoft 365, Google Workspace, GitHub, Salesforce and the AI tools themselves, through customer-authorized OAuth connections rather than software installed on anyone's device. Individual data flows only to the employee and their direct manager, which is the access that manager already has. Aggregated team views require a minimum of five people, and every employee can see what Levos sees about them.
That last constraint is not a limitation to apologize for. A measurement system that can be used to single out an individual will be resisted, and a resisted system produces exactly the kind of distorted data this article is about.
Frequently asked questions
What is the AI productivity paradox? It is the gap between reported and measured productivity gains from AI, documented in a March 2026 Federal Reserve working paper covering nearly 750 executives. The authors attribute it to lagging revenue realization. That is plausible, but an executive survey cannot rule out the alternatives.
Why do 2026 AI and jobs surveys disagree? They are different instruments, aimed at different populations, over different horizons. Pew's 71 percent expecting fewer US jobs over 20 years and Gallup's 18 percent expecting their own job to go within five years are not two answers to one question.
Does AI change how work gets done? By employee report, it changes tasks more than systems. Sixty-five percent of employees at AI-adopting organizations reported productivity gains, while 8 percent strongly agreed AI had transformed how work gets done.
Can manager enthusiasm distort the measurement? Possibly. Gallup found employees whose managers champion AI were roughly eight times more likely to say AI transformed their organization's work. Genuine adoption and sentiment contamination both fit that result, and no survey can separate them.
How do you measure AI productivity without self-report? Baseline, comparison group, controls for tenure and role and tool stack, and a disclosed confidence level, applied to observed behavioral signals rather than survey responses.
The next step
If your AI program is being evaluated on adoption dashboards and sentiment pulses, you are holding the same instrument the Atlanta Fed flagged as running ahead of reality. That is a defensible place to have been in 2025. It is a harder position to defend to a board in 2027.
The related question of how much of that reported gain sits with a small minority of your workforce is covered in our analysis of the AI productivity gap.
Request a Demo to see how controlled cohort analysis with confidence scoring works against your own stack. Levos is accepting design partner applications from US organizations of 150 to 2,500 employees with an active AI rollout.
