Trust & expectations
Users calibrate trust fast and forgive asymmetrically: expectation-setting as a product surface, the forgiveness design toolkit, and disclosure as strategy rather than compliance.
A probabilistic product's real asset is calibrated user trust — users who rely on it for what it's good at and check it where it's weak. Both failure modes are expensive: under-trust means your feature gets ignored (the adoption graveyard is full of good AI features users didn't believe); over-trust means your feature's mistakes get acted on (worse — now your error rate is their error rate). The PM owns the calibration, through three levers:
- Expectation-setting is copy, placement, and first-run. What the feature claims ('drafts a reply for your review' vs. 'answers your customers') sets the trust budget before the first output. The first-run experience should show a representative interaction — including, deliberately, what a miss looks like and how recovery works. Products that demo only magic in onboarding are writing checks the steady state can't cash.
- Forgiveness design — the toolkit that makes errors cheap: visible provisionality (drafts look like drafts — the edit box is the message), effortless correction (one click to fix beats one form to report), receipts (citations, confidence, 'based on your last 3 orders' — users forgive what they can verify), and undo everywhere the feature acts. Every unit of forgiveness you design buys percentage points of acceptable error rate — which is often cheaper than model improvements for the same trust outcome. (Illustrative: a one-tap undo on a wrong suggestion dropped the felt cost of a miss enough that the feature shipped at ~92% acceptance instead of needing to reach ~97% first — the UX change bought five points of error tolerance that the model would have taken a quarter to earn.) That trade — UX spend vs. model spend for trust — is a resource-allocation decision only the PM can make.
- Asymmetric errors need asymmetric design. Users forgive a miss that wastes ten seconds and remember forever a miss that embarrassed them in front of their customer. Map your feature's error types by user-felt cost (not model-measured severity — these differ, and the delta is pure PM insight), then spend forgiveness design on the expensive ones. Harbor Helper misrouting a question: cheap. Harbor Helper confidently misquoting the refund policy to an angry customer: expensive. Same 'error rate', different products.
Disclosure as strategy
'This is AI' disclosure is legally required in a growing set of contexts (the governance courses cover the map) — but treating it as a compliance checkbox misses that it's a trust instrument: disclosed AI gets graded on a curve users reserve for tools; undisclosed AI discovered later gets graded as deception, retroactively, across every interaction it ever had. Dana's rule for Harbor Lane: disclose at first contact, remind at high-stakes moments ('I'm an AI — for this refund dispute, want a person?'), and never let marketing language ('smart', 'magic') creep into surfaces where it erodes the calibration the product spent months earning. The PM is the only person in the building incentive-aligned to defend this line against both engineering's pride and marketing's adjectives.
One viral bad output costs more trust than a hundred quiet successes earn — the asymmetry every customer-facing AI PM operates under. Practical consequence: your risk register (Module 2) weights visibility of failure as heavily as probability, and your launch plan (Module 5) stages exposure accordingly. The strategy course calls this reputation risk; at product altitude, it's just the budget you manage.