AI technical due diligence
The investor's guide to AI technical due diligence
What to inspect, what evidence to request, which red flags matter, and the questions to put to founders before you invest in an AI company.
By Kesho Partners
12 minute read
A strong demo proves an AI workflow can work once. Diligence must show it works repeatedly, at a cost the business can sustain, on data the company has the right to use, and in a way competitors cannot easily copy. Check ten areas: architecture, model dependency, proprietary data, defensibility, evaluation, security, scalability, inference economics, the technical team, and IP and legal. Ask for evidence, not slides, and test the most important claims yourself.
Download
The investor's guide to AI technical due diligence
Know what you are really buying before you sign. Take the full handbook into your next AI deal and see past the demo.
Why AI diligence is different
Standard technical diligence assumes software behaves the same way every time, costs little to run per user, and is protected by the code itself. AI products break all three assumptions.
The same input can produce different outputs, and quality can change when a provider updates a model even if the company changes nothing. Every request costs money in model calls, retrieval, and often human review, so gross margin depends on how customers use the product. And as leading models converge, access to a good model is rarely a moat. Stanford's 2025 AI Index found the score gap between the top and tenth-ranked models had narrowed to 5.4%, while the cost of GPT-3.5-level inference fell more than 280-fold in two years.
Regulators are also watching AI claims. In March 2024 the SEC settled with two investment advisers, Delphia and Global Predictions, for a combined $400,000 over false and misleading statements about their use of AI. In September 2024 the FTC's Operation AI Comply included DoNotPay, which agreed to pay $193,000 after marketing a "robot lawyer" it had not tested against human lawyers. Overstated AI is now a legal risk, not only a valuation risk.
Run diligence in five phases
Go deep only where the investment thesis depends on the answer. A seed review may take days; a growth or buyout review may take several weeks.
- 1. Screen
- Decide whether AI is core to the thesis and whether the claims are plausible. Review the deck, website, and product, and hold one founder call.
- 2. Request
- Send a data room request list early. Ask for architecture diagrams, evaluation results, cost data, data licences, security testing, and IP assignments. Missing items are findings in their own right.
- 3. Inspect
- Walk through the architecture and code. Interview the CTO, an engineer, the person who owns evaluation, and a customer-facing lead.
- 4. Test
- Re-run evaluations, review sampled outputs with a domain expert, rebuild the cost model, and check security against recognised AI threat lists.
- 5. Decide
- Score each area, then agree conditions, price implications, and the first 100 days after investment.
The ten areas to assess
Each area answers one question. The full guide gives the evidence to request, what good looks like, and the red flags for each.
| Area | The question to answer |
|---|---|
| Architecture | Is this a controlled system, or a demo held together by prompts and people? |
| Model dependency | Could the company survive a provider changing price, behaviour, terms, or availability? |
| Proprietary data | Does the company have the rights to its data, and does that data measurably improve the product? |
| Defensibility | What stays hard to copy if a competitor had the same model tomorrow? |
| Evaluation | Can the company prove, with a test set and a baseline, that the product works and is not getting worse? |
| Security | Has the company addressed AI-specific attacks such as prompt injection and excessive agent permissions? |
| Scalability | Will quality, speed, and margin hold at the volume in the plan? |
| Inference economics | What does one successful task cost, and how does margin behave under stress? |
| Technical team | Can this team keep the product working and improving without depending on one person? |
| IP and legal | Does the company own what it sells, and do its claims and use cases comply with the law? |
Inference economics: test the cost of one successful task
Ask for the cost of one successful task, not the cost per model call. Include retries, failed attempts, retrieval, hosting, and the human review needed to make the output usable. Then stress-test it: what happens to gross margin if the model provider raises prices by half, if usage per customer doubles, or if a larger share of outputs needs human review?
A company that cannot answer this does not yet understand its own margins. A company that has modelled it, and can show a credible path to lower costs, is far easier to back.
Security: look for AI-specific threats
Ordinary security reviews often miss the ways AI systems are attacked. Use the OWASP Top 10 for LLM Applications and MITRE ATLAS to check for prompt injection, leakage of sensitive data, poisoned training data, and agents with more permissions than they need. The UK's Code of Practice for the Cyber Security of AI, published in January 2025, sets out 13 principles that make a useful baseline.
IP and legal: ownership, data rights, and regulation
Check that code, models, and data are assigned to the company, and that it has the right to use its training data for the purposes it relies on. The US Copyright Office has published a three-part report on copyright and AI, including guidance on when AI-assisted work can be protected.
If the company sells into Europe, map its products against the EU AI Act. High-risk uses, such as hiring, credit scoring, and some insurance pricing, carry obligations on risk management, data governance, documentation, and human oversight. The most serious breaches can bring fines of up to €35 million or 7% of worldwide annual turnover.
This is a general overview, not legal advice. Involve counsel for jurisdiction-specific review.
Five questions to ask every founder
The full guide includes 43 founder questions grouped by area. If you only have time for five, ask these.
- Evaluation
- Show me the test set you use to decide whether a release is better or worse.
- Economics
- What does one successful task cost you today, including human review?
- Dependency
- What happens if your main model provider changes price or behaviour next quarter?
- Data rights
- Which data do you have the contractual right to use for training and improvement?
- Defensibility
- What would a competitor with the same model still be unable to copy?
Recognised frameworks to anchor the review
Use established frameworks so findings are consistent and easy to defend. The NIST AI Risk Management Framework and its Generative AI Profile give a structure for AI risk. ISO/IEC 42001 is the international standard for an AI management system. OWASP and MITRE ATLAS cover AI security threats.
Read certificates carefully. An ISO/IEC 42001 certificate shows a management system exists within a defined scope. It does not prove a model is accurate, that its data rights are clean, or that its margins hold.
For a quick first pass, use our AI technical due diligence scorecard. For a full review, download the guide or talk to our team about an independent technical due diligence.
Related service
Technical due diligence
Independent AI and technology diligence for investors: evidence-backed findings, risk-ranked issues, and a clear 100-day plan.
Explore the service