ChecklistMergers & Acquisitions

Due diligence on AI companies: what to verify before you buy

When a company's value rests on AI, standard technical diligence leaves the expensive questions open: is the model owned or rented, can the data behind it lawfully be used, do the published results hold on data the company has never seen, and what does each customer action cost to serve? This checklist covers those questions in the order buyers need the answers, and shows how findings become price changes and protections.

Reviewed 10 October 2026 8 min read

On this page

Classify the target's AI before anything else
Model provider and license checks
Training-data rights and personal data checks
Reproducing performance claims on a buyer-held test set
Inference unit economics: what each customer action costs
Regulatory exposure specific to AI targets
Security, people and intellectual property checks
Red flags that should change the price or the protections
A hypothetical contract-analysis target under test
Questions and answers
Sources

Classify the target's AI before anything else

Most AI companies follow one of three patterns, and each puts value in a different place. Establish which applies from the code, the infrastructure bills and the contracts, not the pitch deck.

Question	Proprietary model	Fine-tuned open-weight model	Orchestration over third-party APIs
What the buyer acquires	Weights, training pipeline, data and the team that built them	Adapted weights, tuning data and evaluation work on someone else's base model	Prompts, workflows, integrations and customer relationships
Main external dependency	Compute supply and data sources	The base model's license terms	The provider's pricing, terms and availability
How hard it is to copy	Hard if the data and evaluation are unique	Moderate: the base model is public, the tuning data may not be	Often easy unless workflow data or distribution are unusual
Where cost exposure sits	Training and hosting spend	Hosting, and retraining when the base model changes	Per-call fees the target does not control
Evidence to ask for	Training runs, experiment logs, compute invoices	Base model version, license, tuning datasets, evaluations	Provider contracts, usage logs, prompt repositories

Many products mix patterns. Classify each revenue-bearing feature, not the company as a whole.

Model provider and license checks

These items establish whether the business can keep using its models, at a predictable cost, under a new owner.

0 of 7 checked

Model inventory by product featureWhich model, version and provider sits behind each revenue-bearing feature, and where it runs.
Assignment and change-of-control terms in provider contractsWhether consent is needed, and whether the provider can terminate or reprice when ownership changes.
Pricing, rate limits and notice periodsCommitted capacity, minimum spend and the notice owed before prices change or a model is retired.
Terms on outputs and data useWho owns outputs, whether the provider may train on the target's or its customers' data, and how long prompts are kept.
Tested fallback to another modelWhether the product has been evaluated on an alternative provider, and how much rework a switch would take.
Open-weight model license obligationsUse restrictions, acceptable-use policies, attribution or naming rules, and conditions that change with company size.
Open-source software scanCopyleft components in distributed software or network services, and complete license notices.

Training-data rights and personal data checks

Data rights decide whether the model can keep being used and sold. Problems here are costly because the fix can mean retraining.

0 of 6 checked

Source register for every training and tuning datasetOrigin, license or contract, and which models were trained on it.
Licenses that permit commercial trainingA research-only or non-commercial dataset can taint every model trained on it.
Scraping practices and complaintsHow site terms and opt-out signals were handled, and any takedowns or claims received.
Customer data used for trainingWhether customer contracts allow it, and whether any customer has opted out.
Lawful basis for personal dataThe basis relied on under GDPR Article 6 and any impact assessments². The European Data Protection Board's Opinion 28/2024 discusses legitimate interest for AI models and how unlawful processing during development can affect later use³.
Deletion and objection requestsWhether requests can be honored for data already inside a trained model.

Reproducing performance claims on a buyer-held test set

Pitch-deck scores are usually measured on data the company chose. A controlled test on data it has never seen tells the buyer far more.

List the claims that carry the valuation
Record each accuracy, quality or automation claim the business case relies on, with the exact metric and conditions.
Output
Claims register
Owner
Buyer deal team
Assemble a held-out test set
Build a sample from the buyer's own data or a fresh source that matches real use, and keep it away from the target until the run.
Output
Sealed test set
Owner
Buyer data lead
Agree the protocol in writing
Fix the metric, model version, configuration and pass threshold before anything runs, so the result cannot be argued away.
Output
Signed protocol
Owner
Both technical leads
Run it under observation
Execute in a clean room or observed session with logs kept, so nobody tunes the system to the test set mid-run.
Output
Run logs and raw outputs
Owner
Target engineers, observed by the buyer
Check for contamination
Look for overlap and near-duplicates between test and training data, and for public benchmarks that leaked into training. Contamination inflates scores.
Output
Contamination report
Owner
Buyer data lead
Compare against a plain baseline
Run the same test on a general-purpose model with a sensible prompt. If it comes close, the technical advantage is thinner than claimed.
Output
Baseline comparison
Owner
Buyer technical lead

Inference unit economics: what each customer action costs

Gross margin in an AI business depends on the cost of one unit of value: a document reviewed, a ticket resolved, a call answered. Ask the target to compute it from production logs: model calls per action, compute or tokens per call, retrieval and storage, and the share of actions retried or escalated to a person.

Then test sensitivity to a provider price rise, to heavy customers growing faster than light ones, and to a forced move to a more capable, more expensive model. A business charging a flat fee per seat while paying per call is exposed to all three. The results belong in the valuation model.

Regulatory exposure specific to AI targets

Sector rules in credit, insurance, health or employment apply on top. Check which obligations apply on the expected closing date.

EU AI Act (Regulation (EU) 2024/1689)

European Union

Applies whenThe target places AI systems or general-purpose models on the EU market, or their outputs are used in the EU¹.

Establish the target's role for each system: provider, deployer, importer or distributor. A company that substantially modifies a high-risk system or puts its name on one can become its provider.
Classify each system as prohibited, high-risk, subject to transparency duties or minimal risk, and check general-purpose model duties if the target trains its own.
Ask for the technical documentation and risk-management records that the role and class require.

GDPR (Regulation (EU) 2016/679)

European Union and EEA

Applies whenTraining data, prompts or outputs contain personal data of people in the EU, or the target is established there².

A documented lawful basis for training and for operating the model.
Transparency to data subjects and working processes for access, objection and erasure requests.

Security, people and intellectual property checks

0 of 6 checked

Access controls on model weights and training dataWho can download them, and whether access is logged.
Prompt-injection testing on features that take actionsEspecially agents that send messages, move money or write to customer systems.
Incident historyData leaking through outputs or product abuse, and how each case was handled.
Invention assignment from employees and contractorsCovering code, models and datasets, including founders' work before incorporation.
Key contributors and retentionWho could rebuild the core system, and what keeps them after closing.
Claims from former employers or universitiesWhether a founder developed the core idea under another institution's IP policy.

Red flags that should change the price or the protections

A proprietary model turns out to be orchestration over a public API

Early signalNo training pipeline in the repositories and no compute spend beyond per-call fees.

MitigationRevalue the business as software and distribution; the AI premium in the price may not hold.

Results were only measured on public benchmarks

Early signalThe target cannot run a test on unseen data before signing.

MitigationRequire a buyer-held test before signing, or tie part of the price to an earn-out with clear quality measures.

Training data of unclear origin

Early signalThe source register has gaps, or datasets arrived with former employees.

MitigationSeek a specific indemnity backed by an escrow or holdback, and a covenant to retrain without the affected data.

Provider terms block a change of control

Early signalThe main model contract allows termination or repricing on a new owner.

MitigationMake provider consent a closing condition, or price in the cost of switching models.

Margin depends on introductory model pricing

Early signalUsage runs on promotional credits that expire soon after closing.

MitigationModel margin at list prices and adjust the valuation, or link deferred consideration to margin.

A hypothetical contract-analysis target under test

Questions and answers

How is diligence on an AI company different from normal technical due diligence?

Technical diligence covers architecture, code quality, scalability, technical debt and engineering talent⁴. An AI target adds questions those checks miss: whether the model is owned or rented, whether its training data can lawfully be used, whether results hold on unseen data, what each action costs at future model prices, and which AI rules apply.

Can a company built on a third-party model API still be worth buying?

Yes, if you value it for what it owns: workflow design, integrations, usage data, distribution and customer relationships. Then price the dependency by reading the provider contract, testing how much work a switch to another model would take, and modeling margins at list prices rather than introductory ones.

Can AI tools speed up due diligence on an AI company?

They help with reading. ColdAI's use case on AI agents for private equity due diligence describes agents that index a data room and draft a cited red-flag memo. Judgments on this checklist, such as whether a model is defensible or a dataset properly licensed, still need people testing the evidence.

Sources

Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) — EUR-Lex · checked 10 October 2026
Regulation (EU) 2016/679 (General Data Protection Regulation) — EUR-Lex · checked 10 October 2026
Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models — European Data Protection Board · checked 10 October 2026
Mergers & Acquisitions: technology due diligence and AI and data asset valuation — ColdAI

More in Mergers & Acquisitions

Back to Mergers & Acquisitions

Next step

Send us the AI claims your valuation depends on

Share the target's main performance and technology claims and your timetable. We will reply with the checks that matter most for this target and whether a buyer-held test can fit before signing.

Discuss an AI target