Trust Is What Users Stop Double-Checking. Write OKRs for It.

Most AI-product OKRs measure the wrong subject. They track output accuracy, model quality, and feature adoption. Jeff Gothelf, co-author of Lean UX, puts the point plainly: “A key result has to be a measure of human behavior.” The AI produced a summary. Fine. What did the person do with it?

Gothelf’s answer reframes trust as something you can watch happen. “Real trust isn’t a feeling ultimately but rather a change in what someone is willing to hand over. So watch what they hand over.” That sentence turns a soft word into a metric. Trust shows up in the moment a user stops rewriting your output, stops fact-checking it, and starts shipping it unchanged. A survey score never captured that.

Accuracy Is a Proxy. Behavior Is the Signal.

An AI feature can hit 94% factual accuracy and still fail. If every user rewrites the output before sending it, the tool saved nothing. It added a review step. Accuracy measures the model. It says nothing about whether the model earned any real work.

Behavioral key results measure the thing you actually sold: less human effort spent on the task. When someone accepts an AI draft as written, that is a data point. When they run three fact-check queries before trusting a claim, that is a different data point. Both are observable. Both belong in a dashboard. Neither shows up in a model-accuracy report.

The shift is uncomfortable for teams used to measuring their own system. Behavioral OKRs measure your users instead. They expose whether the product changed anyone’s workflow, which is the only test that matters after launch.

Four Key Results That Encode Trust

Gothelf offers illustrative targets. Treat these as a template for your own baselines, not as market benchmarks. He is describing the shape of the metric and leaving the actual numbers for you to measure.

Share without rewriting: 40% to 65%. Of the AI outputs a user sends onward, how many go unedited? A rising number means the output is trusted enough to act on. A flat number means people still treat every draft as raw material.

Fact-check sessions: 60% to 30%. How often does a user open a second source to verify an AI claim before using it? Falling verification is the clearest trust signal there is. It is the behavioral echo of “I no longer need to check this.”

Override rate: under 8%. How often does a user delete, reject, or heavily rewrite an AI decision? A low override rate says the AI’s judgment matches the user’s. A high one says the human is still the real operator and the AI is a suggestion box.

Auto-share adoption: 15% to 40%. How many users turn on the setting that lets the AI act without a manual review step? Enabling autonomy is the strongest hand-over of all. Nobody automates a task they distrust.

Read those four together and a pattern emerges. Each one measures a decision the human made about how much to rely on the machine. That is a trust scoreboard.

The Calibration Problem You Cannot Skip

Falling verification is only good news if the AI is actually right. A fact-check rate that drops while error rates hold steady is earned trust. A fact-check rate that drops while errors climb is misplaced trust, and it is more dangerous than skepticism.

So a behavioral OKR needs a guardrail metric beside it. Pair “fact-check sessions down 50%” with “error rate flat or falling.” Pair “override rate under 8%” with a sampled audit of what got shipped unchecked. Without the guardrail, you are optimizing for a number that rewards blind delegation. The goal is calibrated trust, where verification falls because the output deserves it, not because the user got tired of checking.

This is where trust measurement becomes governance. Declining verification plus rising delegation is quantified adoption. The same two numbers, read against an error baseline, tell you whether that adoption is safe. One instrument, two questions: are people trusting the AI more, and should they be?

Why PMs Should Care Before Their Board Does

The behavioral frame answers a question every AI budget eventually faces: did this work? Model-accuracy dashboards cannot answer it. A board never funds higher BLEU scores. It funds workflow change. “60% of users now ship AI drafts unedited, up from 40%” is a sentence a CFO understands. “The model improved 3 points on the eval set” is not.

There is also a defensive reason. If your competitor’s users are handing over more work each quarter and yours are not, you are losing on the only axis that compounds. Trust adoption is sticky. A user who has stopped double-checking your output has rebuilt their workflow around you. That is switching cost you did not have to buy.

The Baseline You Do Not Have Yet

The uncomfortable part: most teams cannot set these OKRs today because they never instrumented the behavior. They log model calls and output volume. They do not log whether the output was edited before it left the building, or how many verification queries preceded a send. The behavioral signal exists in the product. It is usually just not captured.

Gothelf’s rubric is only as good as the events under it. Before you set a “share-without-rewriting” target, you need to be recording edits between generation and send. Before you track override rate, you need to know what an override looks like in your data. The instrumentation is the real project. The OKR is the easy part once the events exist.

This connects to a broader argument we have made about measuring judgment, not just output, and about treating trust as the actual user experience rather than a feature bolted onto it. Gothelf’s contribution is the operational layer: the specific behavioral events a product manager can adopt this quarter and hand to an engineer as a tracking spec.

Do This Now

Pick one AI feature. Define the single behavior that would prove a user trusts it, expressed as something you can count: drafts sent unedited, verification queries per session, or autonomy toggles enabled. Instrument that event this sprint. Set a baseline before you set a target, because a target without a baseline is a wish. Then pair it with an error guardrail so falling verification means earned trust and not fatigue. One behavioral key result, one guardrail. That is a trust scoreboard your board can read and your engineers can build.

This analysis synthesizes How to Write OKRs for an AI Product (Jeff Gothelf, June 2026).

Victorino Group helps enterprises turn AI trust into measurable, governed key results. Let’s talk.