Save-play recommender

When a renewal is at risk, the useful output is not a generic retention script. It is a play that already worked on a similar risk, with a citation to the past save, and a CSM who still owns the next step.

The quality bar is retrieval, not generation. Return a play from a confirmed save whose risk pattern matches this account. The citation names the prior opportunity, the risk that was present, and what the team actually did. The CSM accepts, adapts, or rejects. Nothing auto-sends. Nothing invents a concession that is not in the cited record. If the opportunity is already closed-lost, the recommender stays silent.

This sits downstream of risk detection. The churn-risk prediction model can flag the renewal. Champion changes can surface through champion-departure monitoring. Those signals are inputs. They are not the play. The play comes from history, not from a template library that ignores why this account is failing.

Platforms in this class (Gainsight, ChurnZero, Salesforce) already hold health scores, renewal opportunities, activity, and playbooks. The gap is whether the next action is grounded in a past save that looks like this one, and whether the CSM can see the cite before they act.

Start from a cited current risk. Attach it to fields and events the team already uses in forecast and QBR: usage drop after a champion left, stalled legal, a competitor evaluation in this cycle, a severity spike on a core workflow. If you cannot point to the evidence, you cannot retrieve a matching save.

Restrict candidates to confirmed saves. A save is an opportunity that was at risk and later closed-won, with a recorded play, not a quiet renewal that never entered a risk queue. Closed-lost records belong in a win-loss reason classifier for learning. They are not source material for a play you will run this week.

Match on risk pattern and on the segment that constrains the motion: commercial or enterprise, self-serve or named CSM, product line, region, contract vehicle. A play that worked for a self-serve mid-market account does not transfer to a named-enterprise renewal because both show usage down. Usage down in a land-and-expand logo often means a stalled rollout owner. Usage down in a self-serve book often means a failed first-value moment. Mixing those is a failure mode.

Retrieval returns three things together: the play, the cite (which save, which risk, which segment), and the constraints (what was not done). If the cited save did not include a discount, do not add one. If it used an executive working session and a scoped success plan, that is the play. Padding it with a commercial concession is invention.

Score the match in terms a CSM can audit: shared risk type, shared segment, recency against current packaging, same product surface. If two of those fail, return nothing. A thin match is worse than empty. Empty forces the CSM to work the account. A thin match lets them copy a motion that does not apply.

The CSM is the owner. Retrieval proposes. It does not fire.

Read the cite first. Confirm segment, risk, and outcome. Different segment: reject. Sunset product sold a different way: reject. Opportunity already closed-lost: stop. Suggesting outreach after Closed Lost is noise on an account that now belongs to win-loss.

If the cite holds, pick, adapt, or reject.

Pick when the play maps without changing commercial terms. The cited save recovered champion-departure risk by introducing the incoming owner to the existing success plan, running a short working session on the workflows they inherited, and putting a named executive on the next business review. Run that here only if this account has the same risk and the same motion is available.

Adapt when the intent is right and the packaging has to change. The cited save used an onsite. This account is remote-only. Keep the working session and the executive presence. Drop the travel. Adaptation is not a license to add a discount, extra term, or product give that is not in the cite.

Reject when the match is cosmetic: same risk label, different buyer, product, or authority. Reject when the only retrieved action is a concession. If history does not show a discount, do not invent one to make the recommendation feel complete.

One illustration. A named-account CSM has a renewal still in legal. Risk is cited as champion departure plus a stalled rollout of the workflow the new owner cares about. Retrieval returns a prior save in the same segment: new champion, same product family, same risk pair. The play was a handover with outgoing and incoming owners, a rewritten success plan against the new owner's workflow, and an executive check-in on the next auto-generated QBR packet. The CSM picks that play, books the handover, and leaves pricing alone. A retrieval that offered a commercial discount to close would be rejected. That concession is not in the cite, and it trains the book that risk equals discount.

Write the decision back: picked, adapted (what changed), or rejected (why). That stops the next retrieval from repeating a bad segment match. The save motion is in progress because a human chose it.

Cross-segment retrieval is the quiet failure. The risk label matches. The play looks professional. The buyer lacks the authority, packaging, or CSM motion the cited save used. You spend a week on enterprise executive alignment for a commercial account that needed product-led re-onboarding, or the reverse. Gate on segment before similarity on risk text.

Invented concessions are the loud failure. The model cannot find a clean play, so it completes the sentence with a discount, a free module, or an unapproved term. That is generation dressed as history. If the cited save has no commercial change, the retrieved play has none. Pricing exceptions stay on deal desk, manager, and documented exception. A recommender that mints discounts gets used once, then ignored.

A save after close is the operational failure. Closed-lost is not an at-risk renewal. Outreach that looks like a save on a closed-lost record confuses the customer, pollutes activity, and contaminates training. Stage-gate: open renewal or at-risk open opportunity only. Closed-won needs no save. Closed-lost goes to win-loss. Last-mile pressure on a still-open deal is a different motion, closer to last-mile deal coaching, and it still requires an open opportunity.

Two smaller failures: retrieving from a save that was never at risk (a healthy auto-renew labeled as a win), and retrieving from notes that describe intent rather than executed action. Confirm the save, the risk, and that the play was done.

Done is a retrieved play, a cite the CSM can open, a decision on that play, and an opportunity that is still open. Forecast should be able to say this account is in a save motion taken from save X, risk pattern Y, segment Z, owned by this CSM. A task that says "save the account" is not inspectable.

No play is a valid result when there is no matching confirmed save. The CSM then works from the cited risk, the QBR, and live conversations, without a fake precedent. Empty retrieval is honest.

Keep the loop tight for a weekly book. Retrieval refreshes only if the cited risk changed and the opportunity is still open. The last pick, adapt, or reject stays visible. Do not re-fire a rejected play from a nightly job. Do not launch a sequence because a similar ticket appeared in another segment.

If health scores and playbooks already live in Gainsight, ChurnZero, or Salesforce, treat this as retrieval and ownership on top of those objects. The objects stay. The recommendation has to earn the cite.