← All AI Guides
#001 · AGENT REVIEW · an ecommerce business · SEP 07, 2026BLOCKED

A green checkmark said 4 people were reachable. The real number was 250.

The agent hadn't broken the rule. It had applied the rule to half the data — and every step after that looked correct.

The job I handed over

  • Build the list of people we can email for a campaign and apply our firm consent rule: nobody gets mailed unless their marketing consent is current.

What happened

  • The step reported done: 4 sendable contacts out of roughly 3,100.
  • It also warned that 8,314 contacts would be blocked across the whole database.
  • It cited the consent rule it had applied and spot-checked samples to prove it.
  • It looked rigorous. Nothing in the output suggested a problem with the work itself.

How I checked it

  • A separate agent owns one job: when work is marked complete, question the starting assumption, not just the steps.
  • The assumption it questioned: is checking consent in only one system even correct for these contacts?
  • It isn't. Consent lives in two systems — the current ecommerce platform and the legacy email platform. The step had validated against the ecommerce platform only, silently discarding everyone whose consent came from the legacy email platform.
  • A cross-check pulled records from both systems (~24k profiles each) and merged them. About 250 people were reachable, not 4; after 134 cross-source consent conflicts were excluded, 228 were sendable.
  • The alarming 8,314 blocked figure came from looking at only one system. The underlying data was fine.
  • The actual defect: an unnecessary repeat check against only the ecommerce platform buried in the tool that builds the audience, crushing a 250-person audience to 4 on every run.

What it took from me

  • Fixed the tool that builds the audience to use merged consent from the ecommerce platform OR the legacy email platform.
  • Fixed the ~134 genuine conflicts between the systems — people who really had opted out in one system.
  • Changed how the systems share updates so the two sources stop overwriting each other.
  • Re-ran a three-way pull (ecommerce platform + legacy email platform + our database, ~24k contacts) to confirm conflicts went 134 → 0.
  • My own time went almost entirely into the decision, not the repair: deciding that merged consent was the correct rule.

What I took from it

  • A confident checkmark sitting on a plausible-but-wrong assumption is the most dangerous kind of done. Re-running the same logic would have returned the same green forever.

Try this

  1. Before you delegate a recurring job, write down how you will check the result first — the specific evidence you'd accept as proof it worked.
  2. Then write down one assumption the check itself should question. Usually it's "does this source have all the information?"
  3. Decide in advance what happens when the check fails: who is told, what stops, and what does not get sent.
  4. Give the check to something other than the worker. A worker that grades itself will grade its own assumptions too.

Applies to any job where the same fact lives in more than one system — customers, consent, inventory, availability, pricing.