Due Diligence · 10 min read

Red Flags in AI Due Diligence Every Investor Should Know

The 10 most common red flags in AI company due diligence — and how to find them before the term sheet is signed.

By Sasan Ghorbani · Independent AI Advisor · April 22, 2026

The most dangerous red flags in AI due diligence are not the ones that look like red flags. They are the ones that look like strengths — impressive demos, strong retention numbers, compelling founders who speak fluently about their infrastructure. The job of AI due diligence is to determine which of those signals are real and which are well-constructed narratives.

These are the ten patterns I encounter most often, and what to look for in each.

1. The AI layer is entirely third-party with no proprietary component

The most common structural weakness in AI companies is a product that wraps a foundation model — GPT, Claude, Gemini — with a thin application layer and calls it an AI product. There is nothing inherently wrong with building on foundation models. The problem is when the entire value proposition lives in the wrapper, not in proprietary data, workflow integration, or switching costs.

Ask: what does this company own that could not be replicated by a competitor using the same underlying model in six months? If the answer is unclear, the moat is narrative, not structural.

2. Pricing that cannot survive commoditisation

AI infrastructure costs have fallen dramatically and will continue to fall. Companies whose pricing is built on the current cost of compute — rather than on the value they deliver — are exposed to a margin compression cycle that is already underway.

The question is not what the gross margin is today. The question is what the gross margin looks like when the underlying model cost drops by 60% and a competitor passes that saving on to customers. Companies without real pricing power — built on switching costs, data advantages, or deep workflow integration — will not survive that cycle at current valuations.

3. PMF claimed through aggregate metrics, not cohort behaviour

Product-market fit is frequently claimed and rarely measured. The most common version of false PMF in AI companies is a high-level retention number that does not survive cohort analysis.

Look for: cohort retention curves, not averages. Net revenue retention by cohort. Expansion revenue patterns. What percentage of revenue comes from customers acquired more than 12 months ago, and what does their usage curve look like? Genuine PMF shows in cohort behaviour. Narrative PMF shows in aggregate numbers.

4. Founders who cannot explain their infrastructure costs

This is one of the most reliable signals in an AI due diligence conversation. Ask the founding team to walk through their AI infrastructure cost at current scale and at 3x scale. Founders who have genuinely thought about unit economics can answer this clearly. Founders who have not — or who deflect to high-level gross margin figures — are a warning sign.

The inability to explain infrastructure costs is not necessarily dishonesty. It is often a sign that the cost structure has not been stress-tested. That is a problem regardless of intent.

5. Technical debt that requires a rebuild within 18 months

Early-stage AI companies move fast and accrue technical debt. That is normal. The red flag is technical debt that has been papered over with narrative rather than addressed — infrastructure that will require significant investment to scale, security exposure that was deprioritised, or an architecture already showing signs of strain at current load.

The question is not whether technical debt exists. It is whether the founding team understands it clearly and has a credible plan to manage it.

6. A single customer driving disproportionate revenue

Customer concentration risk is not unique to AI companies, but it is particularly dangerous in AI investments where the commercial story is often built around a small number of early design partners. A company where one customer represents 40% or more of revenue is not a product business — it is a services business with product aspirations.

Related: design partner relationships that have not converted to paying customers at market rates. Early partners at discounted rates are not evidence of PMF. They are evidence that the company has customers willing to use the product when it is cheap or free.

7. Impressive demos that mask weak workflow integration

AI demos are genuinely impressive. The question is whether the demo represents the production experience. The gap between a compelling demo and a product embedded in customer workflows — with real switching costs and real usage data — is where many AI investments disappoint.

Ask to see usage data, not a demo. Daily active usage patterns, feature adoption curves, and support ticket volume tell you more about real product-market fit than any sales presentation.

8. No clear answer on what happens when the model changes

Every company building on foundation models faces model change risk: the underlying model is deprecated, the API pricing changes, the quality degrades in a new version, or the vendor introduces a competing product. Companies that have thought about this have a clear answer. Companies that have not are exposed to a risk they may not fully understand.

Ask directly: what is your plan if your primary model vendor doubles API prices next quarter? The quality of the answer tells you a great deal about infrastructure maturity.

9. Revenue growth driven by logo count, not usage expansion

Growth in AI companies is often measured in customer logos rather than usage depth. A company with 200 customers that each use one feature at low volume is a very different business from a company with 50 customers whose usage is expanding consistently quarter over quarter.

Net revenue retention is the metric that distinguishes the two. NRR above 110% is a meaningful signal. NRR below 100% — meaning existing customers are collectively spending less over time — is a serious red flag regardless of new logo growth.

10. The team is strong on product but has not shipped AI at scale

Product talent and AI infrastructure talent are different capabilities. A founding team exceptional at product design and go-to-market, but which has not shipped AI systems at production scale with real usage load, will encounter problems that their track record has not prepared them for.

The question is not whether the team is capable of learning. It is whether the current infrastructure reflects the level of AI operations maturity the investment thesis requires.

The common thread

The red flags above share a common characteristic: they are discoverable before the term sheet is signed. None of them require access to information that a properly structured AI due diligence engagement cannot surface. The question is whether the right questions are being asked — and whether someone with the pattern recognition to interpret the answers is in the room.