Designing AI products people can verify

01 — The trust problem

Trust in AI is a verification problem.

An AI that is right most of the time and a reviewer who can’t tell which answers are the wrong ones will re-check everything. At that point the automation has gained nothing; it has added a step. People don’t withhold trust because a model makes mistakes — they withhold it because the interface gives them no fast way to find out whether this answer is one of them.

That reframes the design task. Not: how do we get people to trust the AI? But: how quickly can a person confirm or reject what it just said — and what happens to their judgement afterwards? A disclaimer answers neither. An interaction pattern can.

Trust isn’t a feeling the interface produces. It’s a check the interface makes cheap.

02 — The patterns

Five interaction patterns that make an AI reviewable.

1. 01Evidence linked to its source.

   Every output points at what it came from — the sentence, the row, the rule — and shows it one interaction away, without leaving the list.

   Reviewers stop re-reading the whole document to check one field. That saved time is the business case.

2. 02Uncertainty shown as priority, not as a score.

   A confidence figure asks the reviewer to interpret statistics. Tiers — clears quietly, worth a glance, needs confirmation — tell them where to look.

   Human judgement is spent where the model is weak, and the confident bulk clears without a click.

3. 03Corrections that feed back.

   Fixing an answer is the same gesture as approving it, and the correction stays attached to the evidence.

   The model improves from real reviews instead of a separate labelling effort. Adoption and model quality stop being two projects.

4. 04The person acts; the model proposes.

   Nothing writes to a system of record without a human action at the point where liability sits — and every action can be undone.

   This is the sentence that gets legal and compliance to say yes.

5. 05Explain in the user’s terms, not the model’s.

   The answer to “why this?” is a sentence, a rule, a precedent — something a reviewer can use in three seconds — not attention weights.

   Explainability nobody opens is a feature. Explainability that speeds up review is a product.

03 — Prototyping it

What an AI workflow looks like in four weeks.

The four-week outline is the same as for any product decision — one decision, a prototype, a study, a written recommendation. What changes is what gets mocked and what has to be real.

• Real: the documents, the data, the tasks, the interface — and the reviewers.
• Mocked: the model. Its outputs are scripted, seeded with the errors your model actually makes, at roughly the rate it makes them — including the confident-but-wrong ones.
• Simulated: confidence tiers, by assigning the scripted outputs to tiers, so the attention gradient can be tested before a calibration curve exists.

The test then asks one thing of reviewers: work through the queue as you would at work. What we measure is whether the seeded errors get caught, how much time goes to re-checking correct answers, and whether anyone can say why an answer was flagged. That is the assumption most AI features rest on — “people will catch what matters without checking everything” — and it can be settled before the model is integrated.

The four weeks, week by week →

04 — Where it applies

Copilots, document intelligence, decision support, agents.

• Copilots — suggestions inside an existing workflow. The design question is when to interrupt, and how to be ignored gracefully.
• Document intelligence — extraction and review at volume. The question is how to make verifying faster than reading.
• Decision support — scores, triage, recommendations. The question is who is accountable for the decision the score nudges, and what they see before they act.
• Agentic workflows — multi-step actions with tools. The question is where the human checkpoint sits, and what a person can see when the agent stops.

05 — Worked example

A review workflow at PwC, built around one question.

Enterprise document extraction, where the reviewer’s confidence was the product’s real feature. The case shows all five patterns in one interface — including a demo you can work through. Done as a member of the product team at PwC, not as an engagement.

Read the case →