Managers AI Evaluation Study

Every manager in the study uses AI daily. So do their teams.

This is not an early-adopter story. Copilot, Claude Code and Cursor are standard issue in the organizations I studied, usually rolled out top-down and often with board-level conviction. One company gave every engineer a premium AI subscription and wrote the cost off as table stakes. Another reorganized its engineering function around AI entirely. The managers in this study are not speculating about a future workflow. They evaluate people inside it.

35 conversations with the people who judge engineers

Semi-structured interviews, 25 to 45 minutes each, with managers running teams of 4 to 20 engineers across fintech, SaaS, cloud infrastructure, health tech and ten other sectors. Everyone in the sample carries real evaluative authority: reviews, compensation, promotion. Each conversation covered how they evaluate today, where AI has entered the process, what they do when their judgment disagrees with the data, and who answers when an evaluation goes wrong.

“Everybody can churn out code now.” Output stopped being the signal.

When AI produces the baseline, raw volume stops telling you anything. 26 of 35 managers said their definition of performance changed materially within the last 18 months. What rose: architectural judgment, risk handling, and what one manager called the multiplier factor, the ability to make everyone around you better. What fell: lines shipped, tickets closed, velocity for its own sake.

The metrics also miss the work that holds teams together. Several managers described engineers who look weak in raw numbers while quietly carrying coordination, mentoring and risk-handling for everyone else. And more than one pointed out that the real bottleneck has moved: AI shifted it from writing code to reviewing and integrating it, which is exactly the work the dashboards weigh least.

AI didn't shrink the work of evaluation. It moved it.

The efficiency story goes like this: AI gathers the evidence, the manager saves time. The managers describe something else. 24 of 35 report that evaluation takes more effort since AI entered the workflow, because the hours saved collecting signals now go into validating them. One manager built a fully automated pipeline that pulls delivery metrics, analyzes code and drafts performance summaries. He also described a standing Monday ritual of repairing his agents when they break. His verdict on his own system:

Input, not outcome

31 of 35 managers drew the same line without being asked: AI can inform an evaluation, it cannot make one. They cross-check it, override it, and treat scientific-looking output as the start of a conversation rather than the end of one. Nearly all of them described some routine for verifying AI claims before acting on them. "Trash in, trash out," as one put it.

The line gets sharper as decisions get more political. One manager split salary from promotion: salary is quantitative enough that AI could plausibly help, but promotions run on company interests that are written down nowhere a model can read them. The most consequential decisions are the least legible ones.

The hardest decisions never touched AI

Every manager could recall their hardest evaluation. An engineer unravelling because of troubles at home. A senior hire who could not land a role change. A strong performer who quietly stopped trying. Not one of them reached for an AI tool in those moments, including the heaviest adopters in the study. The decisive inputs were timing, trust and reading a person in the room.

When it goes wrong, who answers? Three camps.

Then I asked the question this study was built around. When an AI-shaped evaluation harms someone (a missed promotion, an unfair rating, a termination), who answers for it? Thirty-five managers, three camps. And when asked who answers for a negative review in their org today, the most common reply was one word: "I am."

Formal accountability is holding. The conditions for it are eroding.

Put the findings together and a structural picture emerges. Managers remain fully, formally accountable for evaluations. They insist on it. But the three conditions that make accountability meaningful (understanding the basis of a decision, controlling the process, being able to explain it) weaken as AI mediates more of the evidence. Responsibility stays put while control drains out of it. Researchers have a name for where this ends: the moral crumple zone, the point where the nearest human absorbs blame for outcomes an entire system produced.

Every manager in this study is signing off on documents whose evidence base they can audit less and less. They know it. They sign anyway, because the alternative is to stop managing.

What the extreme cases look like

Two managers in the study sit at the far end of AI adoption, and what they describe sounds less like the future of performance reviews than their replacement. A third group is worried about a different destination entirely.

A few things worth sitting with

None of this comes with instructions. But a few patterns from the conversations seem worth passing on, depending on where you sit.

Where I think this is heading

The study ends where the data ends, but the direction feels legible. Within a few years the unit being evaluated will not be the engineer. It will be the engineer plus their agents. Judging the human in isolation already feels artificial to the managers furthest ahead, and the rest of the industry tends to follow them with a lag.

The management role splits in two. One half becomes an editor: cross-examining AI evidence, auditing what the dashboards claim, deciding what deserves trust. The other half is the part that never changed, reading people, timing hard conversations, noticing when a number is hiding a person in trouble. Managers who are only output supervisors sit directly in the path of automation. Managers who own both halves become more valuable, not less.

Accountability will get formalized one way or another. Either organizations write deployment accountability into how they adopt these tools, with named owners, real override rights and time budgeted for validation, or the gap keeps widening quietly until a lawsuit, a works council or a regulator closes it for them. The crumple zone is comfortable for everyone except the person inside it.

Most companies will not delete their managers. But every company will eventually re-answer the question this study kept running into: when the evidence is machine-made and the consequences are human, who signs? The managers who come out ahead will be the ones who keep insisting that the answer is a name. Starting with their own.

What this is, and what it isn't

Findings draw on roughly 35 semi-structured interviews of 25 to 45 minutes with software engineering managers, conducted in 2025 and 2026 across 14 sectors and analyzed with reflexive thematic analysis. Participants are anonymized and quotes lightly edited for clarity. The percentages summarize qualitative coding of a purposive sample. Read them as directional findings from structured conversations, not survey-grade estimates. The underlying academic study was conducted at Stockholm Business School.