STEP · 01
Test every citation
The first test of any AI feature is: does its citation open? If the citation is a string, the feature is not grounded. Reject it.
Checklist
- Open 5 citations before adopting
- Reject features with string-only citations
- Log any non-resolving citation
STEP · 02
Read the drift budget
Every AI answer should declare how fresh it is. If the vendor does not, assume unbounded drift and treat the answer as unverified.
Checklist
- Find the drift budget in the UI
- Do not forward stale AI outputs
- Escalate features without drift indicators
STEP · 03
Assign an AI owner per team
Every team should name a single operator responsible for auditing its AI outputs weekly. This is a twenty-minute job, not a role.
Checklist
- Name one AI owner per team
- Schedule a weekly 20-min audit
- Share audit notes with the team
STEP · 04
Use AI to draft, not to decide
AI features are drafting tools. No AI output should ship without an operator having approved it, even when it feels redundant. That is how trust compounds.
Checklist
- Every AI output is approved
- Never auto-send AI-authored messages
- Log every skipped approval
STEP · 05
Retire AI that cannot be trusted
If an AI feature fails two consecutive audits, retire it. Do not argue about it. Trust is easier to lose than to rebuild.
Checklist
- Retire features after 2 failed audits
- Communicate retirement to the team
- Re-test in 6 months