AI Product Checklist

Design AI features people can trust

5 topics · 18 checks 0%

Problem & fit

Is AI actually the right tool? Name the user problem first. If a rule, a filter or a plain form solves it more reliably, use that. AI earns its place when the input is fuzzy and the value is in handling that ambiguity.
Cost of a wrong answer Map what happens when the model is confidently wrong. Low-stakes suggestions can ship loose; anything touching money, health, safety or reputation needs a human in the loop.
Reward vs. risk is honest Weigh the best-case help against the worst-case harm for the same feature. If the downside lands on a vulnerable group, redesign the feature, don't just tune the copy.

Know your training data Who is represented, who is missing, and how old is it? Gaps in the data become gaps in the experience for the people the data forgot.
Test for bias, not just accuracy A model can be accurate on average and unfair per group. Break results down by user segment and look for who gets the worse output.
Plan for the cold start Design the first-run and no-data states. New users and thin accounts should still get a usable, non-embarrassing experience.

Say when it's AI People should never wonder whether they are talking to a machine. Label AI output clearly and keep the label near the content, not buried in settings.
Set expectations up front State what the feature can and can't do in plain language. Over-promising is the fastest way to lose trust the first time it fails.
Show confidence honestly Signal uncertainty when the model is unsure. A hedged answer the user can verify beats a confident answer they can't.
Explain, at the right depth Offer a reason for the output that a non-expert can act on. "Because you saved similar items" is more useful than a raw score.

Keep the human in charge Let people edit, undo, override or dismiss what the model produces. Automation should feel like a draft, not a verdict.
Make feedback easy A quick way to flag a bad result does two jobs: it gives the user agency and it gives you signal to improve the model.
Design the graceful failure Decide what happens when the model returns nothing, or nonsense. A clear fallback path beats a dead end every time.
Offer an off-ramp Give users a way to reach a manual route or a person when the automated one isn't working for them.

Measure the experience, not just the model Model metrics don't capture trust, effort or abandonment. Track how the feature changes real user behavior over time.
Protect privacy by default Collect the minimum you need, be explicit about what feeds the model, and make sure people can opt out without losing the core product.
Watch it after launch Models drift as the world and the data shift. Schedule a recurring review instead of assuming launch-day quality holds.
Name who is accountable Someone owns the harm if the system gets it wrong. Decide who that is before you ship, not during the incident.

Progress auto-saves in this browser — no account, nothing leaves your device.