FIG. 12 · FIELD NOTES / AGENTS · JUL 2026 · 8 MIN

How the winback agent works — and what it measured

A lapsed-customer email program, rebuilt as a scored, QA'd, human-supervised agent run. Every artifact below is the real one, anonymized.

TL;DR

  • 4,180 lapsed accounts scored, drafted, and QA'd by a two-agent pipeline; 212 drafts routed to human review.
  • Measured against a 10% holdout: reactivation 4.8% → 6.3% (+31%) over 6 weeks.
  • Steal this: the guardrail "never invent usage data" did more for reply rate than any copy tweak.

The client’s winback program was one generic email, sent quarterly, converting at rounding-error rates. The dream was “AI-personalized winback at scale.” The system that delivers it is small: a scoring pass, a drafting prompt with hard guardrails, a QA agent, and a holdout. Nothing exotic — the leverage is in the joins and the rules.

1 · Score before you draft

The CRM already knew last_active, plan, and top_feature — it had just never been joined for the ESP. The scoring agent ranks the 4,180 lapsed accounts by reachability and likely reason for churn, so drafting effort goes where it can pay.

ARTIFACT — THE DRAFTING PROMPT, ANNOTATED

SYSTEM
You are drafting a winback email for a lapsed customer. Use the account facts below. Never invent usage data.


FACTS
last_active · plan · top_feature


RULES

- one specific fact in the first line
- one CTA, plain text, no button
- 90 words max
- if facts are missing → route to human
fact-led first lines scored 0.58 on reply-intent vs 0.31 for generic openers
the missing-facts rule is why QA passed 95% of drafts

2 · QA is an agent too — with humans behind it

A second agent scores every draft on factuality, tone, single-CTA, and length. Anything below the confidence bar goes to a person. Out of 4,180 drafts, 212 needed review — about 40 minutes of human time, instead of a week of writing.

ARTIFACT — THE RUN LOGone batch, real timestamps
09:14:02 segment — 4,180 lapsed accounts scored
09:14:07 draft — 4,180 emails generated
09:14:11 qa — 212 low-confidence → human review
09:15:40 send — 3,968 + 212 approved, 10% holdout withheld
+6 wks measure — reactivation +31% vs holdout

3 · No holdout, no claim

Ten percent of the segment never got the email. That’s the whole measurement system — and the reason the number below is a claim I’ll put on a homepage rather than a vibe.

RESULT — 6 WEEKS POST-SEND
control
4.8%
agent
6.3% — +31% reactivation