What do investors check in technical due diligence for an AI startup?

Key facts

• Under the Copyright, Designs and Patents Act 1988, a computer program is a literary work (section 3), and the employer is the first owner of copyright in a work an employee makes in the course of employment unless agreed otherwise (section 11(2)). (Source: Copyright, Designs and Patents Act 1988, Part I, Chapter I)
• Under section 90(3) of the Copyright, Designs and Patents Act 1988, an assignment of copyright is not effective unless it is in writing and signed by or on behalf of the person assigning it. (Source: Copyright, Designs and Patents Act 1988, section 90)
• The UK's text and data mining exception, section 29A of the Copyright, Designs and Patents Act 1988, covers copies made for computational analysis for the sole purpose of non-commercial research by someone with lawful access to the work. (Source: Copyright, Designs and Patents Act 1988, section 29A)
• The UK government's Report on Copyright and Artificial Intelligence, published on 18 March 2026, said a broad copyright exception with opt-out is no longer its preferred way forward, and that it will not reform copyright law until confident the reforms meet its objectives. (Source: GOV.UK, Report on Copyright and Artificial Intelligence)
• UK Private Capital, formerly the BVCA, added AI warranties and revised the intellectual property, business systems and data protection warranties in the February 2025 edition of its model documents for early-stage investments. (Source: UK Private Capital, model documents for early-stage investments: February 2025 changes)
• The Llama 3.1 Community License requires prominent "Built with Llama" attribution when the materials, or products containing them, are distributed, and a separate licence from Meta for companies whose products had more than 700 million monthly active users at the model's release. (Source: Meta, Llama 3.1 Community License Agreement)
• The OWASP Top 10 for LLM Applications 2025 ranks prompt injection as the top risk (LLM01), followed by sensitive information disclosure (LLM02) and supply chain (LLM03). (Source: OWASP Top 10 for LLM Applications 2025)
• Cyber Essentials, developed by the UK's National Cyber Security Centre, is the minimum standard of cyber security the government recommends; Cyber Essentials Plus adds independent technical testing. (Source: NCSC, Cyber Essentials overview)

Definitions

Technical due diligence

An investor's or acquirer's structured review of a company's technology, data, security, team and technical risks, carried out before committing money.

Data provenance

The documented origin of each dataset, and the licence, contract or consent under which the company may use it for training, evaluation or serving.

Key-person risk

The risk that one person's departure would stall the product, because critical knowledge, access or unassigned code sits with them alone.

Warranties

Statements of fact the company makes to investors in the investment documents, for example that it owns its IP, with a right for investors to claim if a statement proves untrue.

Model dependence

How much the product relies on a specific model or provider, and how hard it would be to switch if prices, terms or the model itself changed.

How deep does technical due diligence go at each stage?

Depth grows with the cheque. The stages below are typical rather than fixed, and an investor with a technical partner may go deeper earlier. At every stage the same three questions sit underneath: does it work, does the company own it, and will it scale without surprises.

The technical due diligence checklist for an AI startup

These are the areas investors cover, with the evidence that answers each one quickly.

Data rights and provenance: the AI-specific questions

Data is where AI due diligence differs most from ordinary software diligence. Investors want to know where every training, fine-tuning and evaluation dataset came from, under what terms, and whether those terms cover commercial use.

In the UK, scraped or copied material is a particular concern. As of November 2026, the text and data mining exception in section 29A of the Copyright, Designs and Patents Act 1988 covers only computational analysis for non-commercial research. The government's March 2026 report on copyright and AI said a broad exception with an opt-out is no longer its preferred way forward, and that without new legislation the status quo continues.

Customer data needs its own answers: whether contracts allow it to improve the product, how a customer opts out, where it is stored, and whether a data protection impact assessment was done where UK GDPR requires one. UK AI regulation for startups covers the legal side. None of this is legal advice.

Model dependence, cost and licences

Building on a provider's model is normal and rarely a red flag on its own. What investors test is whether the company could absorb a price rise, a change in terms or a model retirement: is there one internal interface for model calls, an evaluation set that can qualify a replacement, and a cost per user that still leaves a margin at ten times today's usage? The LLM cost per user guide shows how to calculate that number.

Open-weight models come with licence terms that need reading. The Llama 3.1 Community License, for example, requires prominent "Built with Llama" attribution when the materials or products containing them are distributed, and a separate licence from Meta for companies that had more than 700 million monthly active users when the model was released. The wider trade-offs are in open-weight vs proprietary LLMs.

Evaluation evidence and model governance

A demo proves that an AI product can work; investors want evidence that it does work, on representative inputs, over time. Show the evaluation set and how it was built, scores for each model and prompt version, the failure categories and what was done about them, and the production signals that would reveal a decline. How to evaluate an LLM application covers the methods.

Governance is the other half. Chapter 17 of From Model to Product, Ethics, Governance & Risk, frames model governance as always knowing which model is running, how it was trained, what data it used, when it was deployed and what happens if it fails. A company that can answer those five questions for every model in production, from a registry and logs rather than from memory, passes this part of diligence easily.

IP ownership and key-person risk

Investors check that the company, not its people, owns the product. Under the Copyright, Designs and Patents Act 1988 an employer owns copyright in code an employee writes in the course of employment, but that rule does not reach code a founder wrote before joining the payroll, or code from contractors and agencies. Those need an assignment, and under section 90(3) an assignment only takes effect if it is in writing and signed.

Research spin-outs should also check the university's IP policy, and every company should check that the open-source licences in its codebase fit how the product is sold. The investment documents then turn these questions into promises: UK Private Capital's model documents, revised in February 2025, include IP, data protection and AI warranties.

Key-person risk is the human side of the same question. If one engineer holds the only production access or the only knowledge of the model pipeline, fix it before diligence does: shared secrets management, written runbooks and founder vesting, which technical co-founder equity explains.

Security evidence for an AI product

Security questions now include threats specific to language models. The OWASP Top 10 for LLM Applications 2025 ranks prompt injection first, followed by sensitive information disclosure and supply chain risks, so expect to explain how untrusted input is kept away from tools and data it should not reach.

For the basics, a UK startup can point to Cyber Essentials, the National Cyber Security Centre scheme the government recommends as a minimum standard, or Cyber Essentials Plus, which adds independent technical testing. Apps built quickly with AI coding tools deserve a separate check; see is your vibe-coded app secure.

How do you prepare a technical data room?

Prepare before the round starts. Gathering evidence against a deadline is when small gaps start to look like big ones.

1. A two-page technical overview: what the product does, the architecture, the models used and why.
2. A data inventory: each dataset, its source, its licence or contract, and permitted uses.
3. Signed IP assignments from every founder, employee, contractor and agency that wrote code.
4. An open-source and model licence report.
5. Evaluation results over time, with a note on how the evaluation set was built.
6. Cost per user per month, by feature, with its trend.
7. Security evidence: policies, certification or penetration test results, and the incident log.
8. A team page: who owns which system, where the documentation lives, and the hiring plan.

Frequently asked questions

Sources

1. Copyright, Designs and Patents Act 1988, Part I, Chapter I (sections 3 and 11)
2. Copyright, Designs and Patents Act 1988, section 90
3. Copyright, Designs and Patents Act 1988, section 29A
4. GOV.UK, Report on Copyright and Artificial Intelligence (18 March 2026)
5. UK Private Capital, model documents for early-stage investments: February 2025 changes
6. Meta, Llama 3.1 Community License Agreement
7. OWASP Top 10 for LLM Applications 2025
8. NCSC, Cyber Essentials overview