A March 2026 Atlanta Fed working paper, drawing on a survey of nearly 750 corporate executives, just put a number on something most mid-market operations leaders have been quietly noticing in their own pilots: AI productivity gains are real, but they are smaller measured than perceived, and the gap is wide enough that the authors named it a productivity paradox in plain text (Atlanta Fed, 2026). That paradox is not a measurement artifact. It is the empirical fingerprint of a deeper deployment problem — one that, on a curve presented by economist Scott Cunningham to the Federal Reserve Board on March 27, 2026, has a specific name: the danger zone where AI-augmented output falls below the pre-AI baseline (Forbes, 2026).
For a Head of Operations at a 50–500-FTE company deciding which AI pilots to scale this quarter, the danger-zone framing reframes the planning question from how much time can we save to where does AI compress time without eroding the judgment that produces accuracy. The mid-market AI deployment that scales cleanly in 2027 is the one that runs that calibration explicitly. Most pilots currently in flight do not.
The Atlanta Fed's Productivity Paradox, Quantified
The Atlanta Fed paper (Working Paper 2026-4) is the cleanest cross-firm read on AI's productivity impact currently in the public record. Lead authors Salomé Baslandze and colleagues fielded the survey in late 2025 and early 2026, pulling responses primarily from CFOs through the Duke/Federal Reserve CFO Survey panel, supplemented by Financial Executives International members (Atlanta Fed, 2026). Three findings matter for an operations function.
First, adoption is broad but uneven. More than half of surveyed firms have invested in AI, with the largest measured productivity gains concentrated in high-skill services and finance. The mid-market segment is precisely the cohort still ramping investment — meaning the deployment decisions being made this quarter are setting the productivity trajectory for the rest of 2026.
Second, the gains are positive but modest. Labor productivity improvements show up clearly in the data but vary substantially across sectors. The authors expect gains to strengthen through 2026, but the headline level — at the time of fieldwork — is well below the figures circulating in vendor decks and consulting forecasts.
Third — and this is the load-bearing finding — perceived gains run ahead of measured gains. Executives systematically report larger AI-driven productivity improvements than the underlying data confirms. The authors interpret this as a delay in revenue realization. It is also, more uncomfortably, the empirical signature of pilots that feel productive on the inside while producing output that has not yet shown up as durable, measurable gains on the outside.
A perceived-versus-measured gap of this size is the precondition for the danger zone Cunningham named.
Cunningham's Curve and the Mechanism Behind the Drop
On March 27, 2026, Scott Cunningham — an economist at Baylor — presented to the Federal Reserve Board of Governors and did something most economists do not: he used AI live during the talk to replicate a landmark immigration-sentiment study, downloading 305,000 congressional speeches via an AI agent for eleven dollars (Forbes, 2026). The substance of the presentation, beyond the demonstration, was a production function — a formal economist's curve mapping human time investment to cognitive output, plotted both before AI and after.
The curve has two important features. First, the post-AI curve sits above the pre-AI curve at every level of human engagement — AI raises potential output across the board. Second, when human time investment falls below a critical threshold, the post-AI output curve crosses below the pre-AI baseline. Cunningham calls this the danger zone: the region where the technology that was supposed to make the worker more productive has, in practice, made the worker less productive than they would have been with no AI at all.
The mechanism is straightforward. Before AI, human time and machine time were complements — both were required to produce cognitive work, the way a kitchen requires both a cook and an oven. As machine capability rises, the inputs increasingly become substitutes. The economics push toward a corner solution: all machine, no human. But cognitive output requires judgment — the quiet, hard-to-instrument layer that catches the AI's plausible-sounding mistake, that knows which of three drafts is the one a client will actually act on, that frames a problem accurately enough for the AI to be useful in the first place. Cut that layer too aggressively and the output stops being usable. The pilot still produces deliverables; the deliverables just no longer do the work they did before.
The danger zone is not a hypothetical. It is the operational explanation for the Atlanta Fed's perceived-versus-measured gap. Pilots inside the danger zone produce output that feels faster — because it is — and that measures worse, because the judgment layer has been thinned past the point where the output still holds.
The 13% Wall: When Outsiders Borrow Insiders' Domains via AI
A separate September 2025 Harvard Business School working paper from Iavor Bojinov, Edward McFowland III, and collaborators puts a number on a specific version of this drop. In a controlled study at IG Group, a global derivatives trader, the researchers asked three groups — 12 web analysts (the occupational insiders who normally write the company's investing content), 26 marketing specialists (adjacent outsiders), and 40 software developers (distant outsiders) — to produce investing articles, with AI access standardized across all three groups (HBS, 2025).
The marketing specialists, with AI, produced articles nearly as good as the web analysts'. The software developers — equally capable of operating the AI tool — produced articles that lagged the web analysts by 13% on clarity and competence, even with the AI's full assistance. The researchers named the effect the GenAI Wall: a ceiling on horizontal expertise transfer that AI does not dissolve.
For an operations function, the finding is more uncomfortable than the headline reads. The implicit promise of most workflow-level AI deployments is that AI flattens the difference between specialists and generalists — that a generalist with a good model can do specialist work. The HBS data says the flattening is partial. AI compresses the gap between insiders and adjacent outsiders. It does not close the gap between insiders and distant outsiders. The 13% accuracy delta is what shows up in the final product when a team uses AI to staff work outside its actual domain.
In Atlanta Fed terms, the 13% gap is one specific channel through which perceived gains diverge from measured ones. The pilot deck shows the deployment expanded the team's effective scope. The output, scored on competence, shows the scope expansion came at a measurable accuracy cost.
What This Means for a Mid-Market Operations Function
The mid-market AI deployment patterns that drift into the danger zone share a recognizable structure. There is usually a workflow that was time-consuming for a senior, domain-expert team member. The pilot replaces the senior's time with a more junior teammate plus an AI tool. Cycle time drops. Headcount on the workflow drops. The pilot reports a clean efficiency gain.
Two things have happened that the pilot's instrumentation does not capture. First, the senior's judgment layer — the part that caught the small but consequential errors — has been thinned. Second, the junior teammate has been pushed into work for which they are an occupational outsider rather than an adjacent one. The Atlanta Fed paradox and the Bojinov 13% wall are both active in the same pilot. The board pack shows a green metric. The work has quietly moved into the danger zone.
The diagnostic that catches this is not a productivity metric. Throughput will look fine. Cycle time will look fine. The diagnostic that catches it is a quality audit conducted on the output of the AI-augmented workflow, by the senior who used to do the work, on a sampled basis. If the senior consistently flags errors a pre-AI version of the workflow would not have produced, the pilot is in the danger zone — irrespective of what the throughput dashboard says.
Most mid-market pilots do not run that audit. The senior whose judgment was instrumentally critical is, by the design of the pilot, no longer in the loop on the day-to-day output. The error rate is therefore not visible from inside the workflow. It is visible only externally — through customer escalations, downstream rework, or a delayed quality signal that lags the productivity report by one to three quarters.
The Counter-Argument: "Our Pilots Show Gains, Not Losses"
The natural objection from an operations leader running successful AI pilots is that this danger-zone framing is overstated. The pilot metrics are positive. The team reports satisfaction. The customer hasn't complained.
The Atlanta Fed data is precisely the counter to that objection. Across nearly 750 firms, the typical reported pattern is positive perceived productivity gains and a smaller-than-perceived measured gain. The pilot satisfaction signal is not in dispute. The gap between what teams report and what the productivity statistics confirm is what the data calls a paradox. A pilot showing positive perceived gains is consistent with — not evidence against — being inside the danger zone.
The second counter-argument is more substantive: that the danger zone is a function of pilot design rather than of AI itself, and that mature pilots can avoid it. This is the right reading. Cunningham's curve is not a verdict on AI. It is a map of where the productivity gain lives — and where, on the same curve, the gain inverts. The instrumentation question for an operations function is whether each pilot has been designed to land in the productivity region of the curve and stay out of the danger region, not whether the pilot is reporting positive numbers in month two.
A pilot that has not run a competence audit cannot tell which region of the curve it is operating in. A productivity report is necessary but not sufficient.
One Calibration Most Pilots Never Run
The single calibration that separates a danger-zone pilot from a productivity-zone pilot is structurally simple and operationally rare. It has three components, none of which require additional headcount or vendor spend.
Sample the AI-augmented output at a defined frequency and score it against pre-AI baseline output, using the same senior reviewer who would have produced the pre-AI version. The score is not a thumbs-up/thumbs-down. It is a per-dimension competence rating on the elements that matter for the workflow's downstream use — accuracy, completeness, judgment calls, edge-case handling.
Track the perceived-versus-measured gap explicitly, not as a productivity number but as a quality delta. Atlanta Fed-style language: how much of the pilot's reported gain is durable measured improvement, and how much is perceived improvement that has not yet shown up as a measurable outcome (Atlanta Fed, 2026)?
Define a stop-loss threshold before scaling. If competence on sampled output falls below a defined floor — the HBS data suggests 13% is roughly the lower edge of what shows up on careful scoring of distant-outsider AI work (HBS, 2025) — pause the scale and rebuild the human-time investment until the score recovers. This is the part of the calibration that most pilots cannot do, because the team has already committed the headcount savings to the next quarter's plan.
The unglamorous reality is that the calibration costs maybe 2–5% of the workflow's senior time per quarter. The cost of not running it is that the operations function discovers the danger zone through the customer escalation queue or the downstream rework backlog, with a one-to-three-quarter lag.
The Decision This Quarter
The PwC and BCG data that dominated 2026's AI strategy conversations established the importance of business-model posture in AI deployment. The Atlanta Fed, Cunningham, and Bojinov findings now establish the parallel point on the operational side: deployment posture is not just about where AI gets pointed. It is about how thin the human judgment layer can get before the AI-augmented output drops below the pre-AI baseline.
A Head of Operations does not need to redesign the AI portfolio this quarter to act on this. The decision is narrower. For every AI pilot currently in flight, ask one question: is there a competence audit running on the output of this workflow, conducted by the senior who used to do the work, on a sampled basis, with a defined stop-loss? If the answer is no, the pilot is — on the Atlanta Fed/Cunningham/Bojinov composite reading — operating without the only instrument that distinguishes a productivity-zone deployment from a danger-zone one.
The Atlanta Fed productivity paradox is the most disciplined empirical signal currently in the market that perceived AI gains and measured AI gains are not the same thing. The mid-market operations function that audits the difference this quarter is the one that scales the AI portfolio in 2027 without discovering — through escalations and rework — that the gains were already eroding the baseline.
Add the competence audit to the next pilot review. The instrument costs nothing the operations function does not already have. The cost of running without it is the one number the productivity dashboard cannot show.
