Design an AI Workflow That Sticks · Lesson 5 of 7

Lesson 5 — Approval loops and audit trails

30 minutes. The two safety mechanisms every workflow needs.

---

The approval loop

Three states every AI-touched piece of work passes through:

Proposed. AI generated something.
Reviewed. A human looked at it.
Acted on. It went out / got posted / hit the system.

Workflows without all three steps are workflows that will eventually embarrass you. The middle step — *reviewed* — is non-negotiable for anything customer-facing.

The two-tier rule

For each workflow, define two tiers:

Tier 1 — Routine, low-risk. A human eyeballs before send. (Most workflows.)
Tier 2 — Sensitive or high-impact. Two humans review, or one human + a checklist before send. (Customer complaints, regulatory filings, anything financial.)

You don't need fancy software for this. A "send" button that you click, after reading, is enough.

The audit trail

Every AI action should be retrievable later. At minimum, log:

Timestamp
Input (what AI received)
Prompt version
Output (what AI produced)
Reviewer (who approved)
Final action (what got sent / posted / saved)

Why: when something goes wrong six months from now and the customer says "your AI told me you'd refund the deposit," you can look up exactly what was sent, by whom, after which review.

Simplest implementation: a shared Google Sheet or Notion database where each AI workflow run drops a row. More sophisticated: a real audit table in a database (this is what AAT modules use).

→ Lesson 6 — Testing with real users

Previous: Lesson 4 — Connecting AI to email, calendar, docs · Next: Lesson 6 — Testing with real users