AI teams drift when feedback is private. Rituals make judgment visible.
Weekly eval review is the simplest honesty practice we saw repeated.
Incident salons that include brand and legal prevent heroics from hiding risk.
Honest teams document failure with the same care as launch announcements.
Ritual is how culture survives model churn.
Honesty in AI work is difficult because demos reward confidence.
A weekly failure review changes the incentive. The team brings examples where the assistant was wrong, weak, strange, or too confident. No one is allowed to defend the model as a personality. The question is practical: what did the system need and not have?
One incident salon began with a customer screenshot. The assistant had answered a returns question with language that was correct in one country and misleading in another. The issue was not only the model. The policy source was unclear, the market routing was incomplete, and the escalation rule was too vague.
Legal, brand, customer service, and engineering sat together. That mattered. If engineering had handled it alone, the fix might have been technical. If legal had handled it alone, the answer might have been restrictive. Together they created a better rule: when policy varies by market, the assistant must state the market or ask for it.
Honest rituals also protect teams from quiet decay. A feature that was good in March may be weak in June after products, policies, or model behavior change. Without ritual, teams keep believing the launch version is still true.
The point is not to shame failure. It is to make failure useful before customers make it public.
