Launch planning

A deterministic launch asks 'does it work?'; an AI launch asks 'do the rates hold as exposure grows?' — a question only answerable in stages, each with its own gate. The pattern (which the agents course builds from the engineering side; here's the PM's version):

• Stage 0 — internal: the team and the support org use it on real (or shadow) traffic. Gate: quality contract bars hold on live-ish data; the support team signs off on the miss-path handoffs. That sign-off is not a courtesy — support absorbs every AI mistake with a customer attached, and a support org that learned about the feature at launch becomes its most credible internal critic by week two. Enable them first: what it does, what its misses look like, how to override it, where their feedback lands in the loop. Your first launch ally is never marketing.
• Stage 1 — fractional (5-10%): real customers, bounded blast radius, segments randomized. Gate: bars hold per segment, miss-path rates within forecast bands, no severe incidents — held long enough for the lagging signals (recontact, complaint rates) to mean something; a week of green leading indicators is not a gate, it's a good morning.
• Stage 2+ — ramp with brakes: 25 → 50 → 100%, each step holding for the lagging window, with the pre-agreed auto-pause: 'any severe incident, or [dimension] under bar for 48h, pauses the ramp automatically.' Pre-agreement is the whole trick — mid-ramp, with a quarter ending, nobody negotiates well with sunk cost (the same sentence the strategy course wrote about kill criteria; launches are where PMs live it).
• Rollback is a designed path, not an emergency: what turning it off looks like (the reactive flow still works — your v1 scoping made this true), who can pull it (named, and not only you — you take vacations), what customers mid-interaction experience, and what the comms say. Rehearse it once at Stage 0. Products with rehearsed rollbacks pull them a day earlier than products without — which is often the difference between an incident and a story.

Comms: promise the floor, not the ceiling

Launch messaging for AI features has one iron rule: the marketing claim must be survivable by the p5 experience, not the median. 'Harbor Helper now spots delivery hiccups and offers fixes before you ask' — survivable, because the miss paths degrade to silence or a human. 'Never wait on a delivery problem again' — a promise the first missed delay converts into a screenshot. The PM reviews launch copy against the error map personally, resists the superlatives with the trust-budget argument (Module 1 gave you the vocabulary), and preps the two assets everyone forgets: the support macro for 'what is this AI thing?' and the honest FAQ entry that says what it does, what it doesn't, and how to turn it off. Products that explain their own limits accrue the strangest asset in software: users who defend them when they err.