Should I quietly fix it with a manual process instead of admitting the AI doesn't work yet?

No — undisclosed manual workarounds presented as AI performance create bigger trust and legal exposure if discovered later.

What if this was promised in a signed contract?

Involve legal and leadership immediately — this moves beyond a communication issue into contractual risk that needs proper handling.

How do I avoid this happening again?

Separate 'works in the demo' from 'works reliably in production' explicitly in every future commitment, with edge cases named upfront.

← All playbooks

Tier 4: Total Clusterfuck

What do you do if you promised an AI feature or capability that doesn't actually work as described?

The short answer

Tell stakeholders or the client the specific gap between what was promised and what currently works, with a realistic timeline and what's actually achievable in the interim — vague reassurance that 'the model will improve' is not a plan. If it was sold as a commitment (in a contract or public statement), loop in leadership immediately rather than trying to quietly patch it into working.

What's actually at stake

01AI capabilities are especially prone to overpromising because demos and edge cases diverge fast — the gap gets noticed in production, not before.
02'It'll get better with more data/training' without a concrete plan reads as a stall, not a roadmap.
03If this was sold contractually, the exposure is legal and financial, not just reputational — it needs the right people involved fast.

Do

· State the specific gap clearly: what works, what doesn't, and under what conditions it fails.
· Offer a realistic interim workaround if one exists, even a manual or partial one.
· Loop in leadership or legal immediately if it was a contractual or public commitment.

Don't

· Do not promise a specific fix timeline for a model behavior you don't actually control with confidence.
· Do not let the client discover the gap themselves in production.
· Do not quietly patch it with a manual workaround without disclosing that it's not the AI actually performing as described.

A draft you can send now

I want to flag an honest gap between what we described and current performance: [specific capability] works reliably in [conditions] but not yet in [conditions]. Here's what we can do in the interim: [workaround], and here's our realistic plan to close the gap: [plan and rough timeline]. Wanted to be upfront now rather than have this surface in production.

Tailor this to my exact disaster →

Questions people actually ask

Should I quietly fix it with a manual process instead of admitting the AI doesn't work yet?: No — undisclosed manual workarounds presented as AI performance create bigger trust and legal exposure if discovered later.
What if this was promised in a signed contract?: Involve legal and leadership immediately — this moves beyond a communication issue into contractual risk that needs proper handling.
How do I avoid this happening again?: Separate 'works in the demo' from 'works reliably in production' explicitly in every future commitment, with edge cases named upfront.

Take it further

Ask Cris about this →

Talk it through with the crisis bot using your actual details.

Grade what you already sent →

Incident Autopsy scores the damage and rewrites it.

Rehearse the conversation →

Spar with a skeptical manager, exec, or client before the real one.

Tier 4: Total Clusterfuck explained →

Money, customers, or a very senior person's calendar is involved.

Related incidents

All CYABot tools →