The NPS Autopsy Playbook 2026

Three thresholds fire the play, each a moment when the customer has something real to say: an engagement spike sustained for 30 days (they just got a lot more valuable to talk to), a 20-point engagement decline (something broke — find out what before the renewal does), or an activation milestone reached (the value moment is fresh enough to confirm). Accoil computes these from the events you already stream, and the signal carries which threshold fired, the trend behind it, the user's role and the account segment. One hard rule: suppress any account in an open save play. Nothing says "our systems don't talk to each other" like an NPS ask landing mid-escalation.

The threshold routes the survey — no human picks recipients, no list gets pulled. Spiking power users get the advocacy ask, decliners get the friction probe, the newly activated get value confirmation. Cap frequency at one micro-survey per user per quarter regardless of how many thresholds fire; the play's response rate lives or dies on the asks staying rare.

In-product, one question, at the end of a completed workflow: "Would you recommend us to a peer?" with a follow-up ask on a 9–10 — a review, a reference call, a case study. These are the only users who should ever see the classic NPS question, because they're the only ones whose answer you can act on the same week. Route the yeses straight into the advocacy sequence in Customer.io while the enthusiasm is real.

Decliners don't get a 0–10 scale; they get the friction probe: "What's gotten harder about [the feature their usage says they abandoned]?" One question, free text, in-product where the decline is happening. A number from a declining account tells you what you already know; a verbatim tells you whether it's a bug, a pricing gripe, or a new stakeholder — three different plays.

Within days of the activation milestone: "Did [the milestone] get you what you needed?" Yes confirms the activation was real and not a checkbox; no is the earliest churn signal you will ever collect — an account that hit the milestone and didn't feel it is mis-onboarded, and you just found out in week three instead of at renewal.

This is the autopsy. Every response lands next to the respondent's own usage, and the join changes what the answer means: a detractor with high usage is a frustrated power user — an invested, save-able relationship problem. A detractor with no usage is already gone, and the play is a graceful right-size or an exit interview, not a rescue. A promoter who never logs in is a courtesy score, not an advocate — routing them to a G2 ask embarrasses everyone. Out of the join come two fields: the routed class and the owner of the play it triggers.

Each class posts to Slack with the response, the verbatim and the usage context attached, tagged to its owner: frustrated power users to the CSM's save queue, promoters-who-use into the advocacy motion, no-usage detractors to the right-size conversation, value-confirmation noes back to onboarding. The score never gets reported without its class — a "7 from a power user at a renewal-quarter account" is actionable; a "7" is decoration.

The frustrated power user gets a CSM call within 48 hours — this is the highest-leverage conversation in the whole program, an invested user telling you exactly what to fix while they still care. RevOps aggregates the verbatims by feature and ships the ranked list to the product team monthly: "eleven accounts named the reporting rebuild this quarter, $840k ARR attached" gets a roadmap slot that a sentiment chart never will.

Every response actioned within a week — that's the bar, and it's only reachable because volume is low and every response arrives pre-classified. In-product micro-surveys at behavioral moments convert several times better than the emailed quarterly blast, so you'll have more signal from fewer asks. Backtest quarterly: did the friction-probe verbatims precede the churn? Did the advocacy cohort renew above the base rate? When the answers are yes, you have the thing the quarterly blast never was — a survey program that predicts.