When an investor or a board asks us to look under the hood of an AI company or programme, the headline claims rarely survive contact with three questions. A good diligence isn't about the demo. It's about what happens the day after.
The questions that matter
- What's actually in production? “Live” can mean anything from a pilot for ten internal users to a system carrying real load. Get specific.
- How do you know it's good? If there's no evaluation framework, every quality claim is a vibe. Mature teams can show you their evals — and their failures.
- What's the moat — and is it borrowed? A thin wrapper over a foundation model is a feature, not a company. Proprietary data, workflow lock-in, and switching costs are the real assets.
A thin wrapper over a foundation model is a feature, not a company. Diligence is the art of telling the difference.
The red flags a demo won't show
No eval harness. No audit trail. PII flowing where it shouldn't. Costs that scale linearly with usage and no plan to bend the curve. A roadmap that's all model upgrades and no distribution. None of these are visible in a polished walkthrough — which is exactly why they belong in diligence.
The point isn't to be cynical. It's to separate the teams who've done the hard part from the ones still hoping the model will do it for them. We've built the things we're assessing — so our questions are specific, not theoretical.