# How do you crash-test a board pack with several AI models before the meeting?

> An AI crash-test runs your board pack through several cross-checked models before the meeting, to surface weak signals, internal contradictions and points that won't hold in the room. AI sharpens the reading; the decision stays yours.

## What a crash-test surfaces

A team that produced a document no longer reads it: it recognises it. The crash-test restores a fresh eye, without fatigue or internal politics.

- **Internal contradictions**: an assumption made on page 6 that no longer holds on page 30.
- **Weak signals**: what is half-said, or not said at all.
- **The points that won't hold**: the figures, shortcuts and blind spots a demanding board member will catch in the room.

## The method, in three steps

The document is submitted to several AI models (Claude, ChatGPT, Gemini and others) with adversarial reading instructions. The feedback is cross-checked: what recurs across models deserves attention; what appears only once is set aside.

The output is not an AI report. It is a short note, read and signed by a human, pointing to the three to five items to address before the meeting. Returned within 72 hours.

## What AI does not replace

AI does not decide. It knows neither your sector better than you, nor the people around the table. It speeds up detection; the trade-offs, the ranking of priorities and the accountability stay yours. That is the WYP rule: AI to sharpen the decision, never to make it.

## For a leader
Three questions before your next committee: who has read your pack as an adversary? which internal contradictions has no one flagged? which figures won't hold in front of a demanding board member?

## FAQ

**Can a sale memo or information memorandum be reviewed this way?**
Yes. The crash-test applies to any decision document: board pack, quarterly memo, investment file, teaser or sale memorandum. The point is the same: have the file read adversarially before buyers or board members do.

**Why several AI models and not just one?**
A single model has blind spots and can hallucinate. Cross-checking several models does not remove error, but it surfaces what is robust: a point raised by several independent models is worth pausing on.

**How fast is the feedback?**
A useful return is possible within 72 hours for a committee file, as a short, actionable note rather than a technical report.

**Does the AI make the decision for us?**
No. The crash-test surfaces signals; it does not rule. The decision, the ranking of priorities and the signature remain the leader's and the committee's.

**Is the confidentiality of our documents preserved?**
It is a precondition. Sensitive documents are handled in a controlled setting, without exposing your data or feeding it into the training of a public model.

## See also

- [Map and territory](https://wyp.agency/en/glossary/map-and-territory/)
- [Hallucination](https://wyp.agency/en/glossary/hallucination/)
- [Shadow AI](https://wyp.agency/en/glossary/shadow-ai/)

## Sources

- French Institute of Directors (IFA), guide “Les systèmes d'intelligence artificielle et les conseils d'administration” (AI systems and boards of directors), 19 November 2024 (https://www.ifa-asso.com/lifa-publie-un-nouveau-guide-les-systemes-dintelligence-artificielle-et-les-conseils-dadministration/)
- Improving Factuality and Reasoning in Language Models through Multiagent Debate, Du et al., arXiv:2305.14325, 2023 (https://arxiv.org/abs/2305.14325)
- OECD AI Principles (accountability principle), updated May 2024 (https://oecd.ai/en/ai-principles)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/ai-board-pack-crash-test/ · Last updated : 2026-10-14