AI Use Is Now a Line Item in Your Performance Review. Usage Went Up. Confidence Went Down.

It happened quietly, the way the big shifts always do. Sometime in the last two quarters, using AI at work went from encouraged to required, and now it's a thing you get graded on.

JPMorgan built a dashboard that tracks the AI tool usage of roughly 65,000 engineers and buckets each one as a "light user," "heavy user," or "non-user" — and tied it to performance expectations. Microsoft's developer-division chief said using AI on the job is "no longer optional." Meta made "AI-driven impact" a core expectation in every employee's review, regardless of role. KPMG is assessing staff against AI objectives. Block went furthest: leadership warned that people who don't use AI risk being laid off.

And here's the number that should make every one of those decisions feel a little queasy. Across the same period, regular AI use among workers went up 13% — and their confidence in the technology's actual usefulness went down 18%. We mandated the tool right as the people using it trusted it less.

You Can't Mandate the Outcome, So You Mandate the Proxy

A week ago I'd have told you the story was about fear — engineers nervous their Copilot stats would end up in a stack-rank. That fear was real and I wrote about it. This is the next chapter: the fear came true. It's policy now. The dashboards aren't hypothetical, they're in the HR system, and "did you use AI" is a row in the form.

Here's the problem nobody running these programs wants to sit with: you cannot mandate the outcome, so you mandate the proxy, and the proxy is garbage. "Did this engineer use AI" is trivially easy to measure and almost completely disconnected from "did this engineer do better work." Goodhart's Law doesn't care about your good intentions. The moment "heavy user" becomes the label that protects your job, every rational person on the team optimizes for heavy user — more sessions, more accepted suggestions, more agent invocations — none of which is the same as shipping better software. You wanted leverage. You incentivized theater.

The 18% confidence drop is the tell. If AI were quietly making everyone's work obviously better, you wouldn't need a mandate — adoption would look like Claude Code going zero-to-number-one in eight months, pulled in by people who wanted it. You mandate the things that don't sell themselves. The fact that leadership reached for a stick at the exact moment trust was falling means the stick is doing the opposite of building belief. It's manufacturing compliance and calling it transformation.

What I'd Actually Do If I Led That Team

If you lead a team and you're under pressure to "show AI adoption," here's what I'd actually do.

Measure outcomes, not usage. Cycle time, change-fail rate, escaped defects, time-to-restore — the DORA-shaped stuff that was real before AI and is still real now. If AI is working, it shows up there. If it only shows up in your "% of commits touched by an agent" chart, you've measured nothing except how well your team reads incentives.

Never let a usage metric touch a compensation or promotion decision. The second it does, you've converted your most honest signal into your most gamed one. You can look at usage as a curiosity — who's experimenting, who needs help — but the instant it has teeth, it lies to you.

And say the quiet part to your team out loud: I am not grading you on whether you used the AI. I'm grading you on the work. That one sentence buys back more real adoption than any mandate, because it removes the reason to fake it. People adopt tools that make their day better. They perform adoption of tools they're scared of.

The Part That's Going to Age Badly

I sit in regulated finance, so let me push on the part that's going to age badly. We are building permanent HR records that say a named employee was a "non-user" of AI. Think about what that is. In an industry that documents everything for the regulator, we're now logging individual technology-adoption compliance into people's files — and tying livelihoods to it — for a class of tools whose own users trust them less this quarter than last. When the productivity numbers don't materialize the way the mandate promised (and the delegation data says they won't, because devs can still only fully hand off a sliver of their work), somebody is going to go looking for why. "We made everyone use it and graded them on it" is not the answer that survives that review.

Here's the pattern I keep watching repeat. Leadership can't directly create the thing it wants — genuine productivity, genuine adoption, genuine judgment — so it mandates the nearest measurable shadow of it and declares victory when the shadow moves. Sales did it with activity metrics. Agile did it with story points. We're now doing it with AI usage, except this time the proxy is wired straight into people's reviews and their fear, not their belief, is moving the number.

The teams that win the next two years won't be the ones with the highest "AI adoption rate." They'll be the boring ones whose leaders had the spine to say I'm measuring the work, not the tool — and meant it. Adoption you have to enforce isn't adoption. It's a number you taught your best people to fake.

Sources: JPMorgan Ties Engineer Performance Reviews to AI Tool Adoption for 65,000 Staff | Let's Data Science · Most companies are requiring employees to use AI; some IT pros think that could backfire | IT Brew · Forced to Use AI: The Corporate Mandate Reshaping Every Career | SmarterArticles · Meta to formally review employees' AI performance from 2026 | HR Grapevine · How Companies Are Using AI in Performance Reviews in 2026 | Reworked