AI Vendor Due Diligence Checklist for Business

If you have already decided an AI tool could help your business, the next question is whether the company behind it is safe and dependable enough to trust. This AI vendor due diligence checklist helps you move beyond price and features, then gather enough evidence to approve, limit or reject the proposed use.

What AI vendor due diligence means

AI vendor due diligence is the process of checking whether a provider is suitable before your business buys, connects or relies on its product. It is not a guarantee that nothing will go wrong. It is a structured way to identify risks, compare the vendor's claims with available evidence, and decide what safeguards are proportionate.

For Morgan, an operations manager at a mid-size business, the old process may be little more than comparing subscriptions, booking a demonstration and asking whether the software integrates with existing systems. The improved process adds a standard questionnaire, evidence review and sign-off. A low-risk writing assistant receives a light check. A system processing employee records receives a deeper review. A tool influencing hiring, credit, health, safety or legal decisions may need specialist input before any contract is signed.

Use a three-tier review model

Small and medium businesses rarely have the time to conduct an enterprise procurement review for every application. A three-tier model keeps the work realistic:

• Tier 1, low sensitivity: public or non-confidential information, limited integration, and outputs checked by a person. Complete the core checklist and retain the answers.
• Tier 2, business-sensitive: internal documents, customer information, staff data, connected systems or material operational reliance. Request security evidence, examine contract terms and run a controlled trial.
• Tier 3, high impact: sensitive personal data, regulated records, automated decisions, critical operations, or use that could materially affect a person. Escalate to privacy, security, legal or industry specialists appropriate to your region and sector.

The tier should reflect both the information entering the tool and the consequence if the tool is wrong, unavailable, breached or changed without warning.

1. Define the proposed use before assessing the vendor

Write down the exact task, users, data and decision involved. Avoid approving a vendor for a vague purpose such as “general AI”. A useful scope might be: five marketing staff will use the tool to draft product copy from public specifications, with human approval before publication.

Record whether the tool will receive personal information, confidential commercial material, intellectual property, credentials or data from connected business systems. Also record whether its output will inform decisions about customers, employees or members of the public. This scope becomes the basis for every later answer and prevents a low-risk trial from quietly expanding into a higher-risk deployment.

2. Check data collection, use and retention

Ask the vendor to explain, in writing:

• what prompts, files, metadata, account details and usage logs it collects;
• whether customer content is used to train or improve models, and whether that setting can be disabled;
• where information is stored and processed;
• how long each category of data is retained;
• how deletion works during the contract and after termination;
• whether administrators can control sharing, exports and integrations; and
• which subprocessors or model providers can access the information.

Compare the response with the privacy notice, product terms and data processing agreement. Differences between sales answers and binding documents are a warning sign. The US Federal Trade Commission has stated that AI companies should uphold their privacy and confidentiality commitments, and it has also warned businesses against quietly changing terms in ways that undermine earlier promises. See the FTC guidance on AI privacy commitments and its guidance on changes to terms.

3. Review security evidence, not just claims

Look for controls that match the proposed use. Common questions include whether the vendor supports multi-factor authentication, single sign-on, role-based permissions, encryption in transit and at rest, audit logs, independent security testing, vulnerability management, backups and an incident response process.

For Tier 2 or Tier 3 use, request current evidence such as a recognised independent assurance report, certification scope, penetration test summary or completed security questionnaire. A badge on a website is not enough. Confirm which product, locations and services the evidence actually covers, when it was assessed, and whether serious exceptions remain unresolved.

The NIST AI Risk Management Framework provides a voluntary structure for governing, mapping, measuring and managing AI risk. Its companion materials also recognise risks arising from third-party AI systems and dependencies. A small business does not need to implement the whole framework to use its basic logic: define responsibility, understand the context, measure important risks and keep managing them after purchase.

4. Understand the model and supply chain

Many AI products depend on other companies for model hosting, cloud infrastructure, search, transcription or data enrichment. Ask the vendor which third parties are essential to the service and what happens if one changes its model, pricing, availability or terms.

Clarify whether your business contracts with the company operating the model or with an intermediary placing its own interface around another provider. The answer affects where data travels, who receives support requests, who gives contractual commitments and how much control the vendor has over product changes.

5. Test reliability, human oversight and limitations

Ask what the system is designed to do, what it is not designed to do, and how the vendor evaluates errors. Request information about known limitations, monitoring, update practices and the process for reporting harmful or inaccurate results.

Run a controlled trial using representative but non-sensitive data where possible. Measure the tool against a defined baseline. Check accuracy, consistency, staff time saved, failure patterns and the effort required for human review. Do not treat a polished demonstration as evidence that the system works reliably in your business.

For use affecting people, document who reviews outputs, who can override them and how affected decisions can be explained or reconsidered. The European Commission describes the EU AI Act as a risk-based framework for providers and deployers of AI systems. Whether or not your business is directly within its scope, the risk-based approach is a useful reminder that a marketing assistant and a high-impact decision system should not receive the same review.

6. Read the contract for operational risk

Check the documents that will actually govern the service, including the order form, master terms, service levels, privacy terms, acceptable use rules and any product-specific conditions. Focus on:

• ownership of your inputs, outputs and uploaded material;
• licences the vendor receives over business content;
• confidentiality commitments;
• security and breach notification wording;
• subprocessor changes and available objections;
• service availability and support response;
• limits of liability and exclusions;
• indemnities, especially for intellectual property claims;
• the vendor's right to change features, models, prices or terms;
• renewal, cancellation and data export; and
• governing law and dispute location.

A contract may place most financial and operational risk on the customer even when the sales material sounds reassuring. For a material deployment, have the agreement reviewed by an adviser familiar with your jurisdiction and industry. This checklist is a planning tool, not an assessment of legal compliance.

7. Check business stability and support

Consider how dependent your operations will become on the vendor. Ask how long it has operated, who owns it, how support is delivered, whether it has customers with comparable use cases, and how it communicates outages or major product changes. For a young provider, seek clarity on continuity arrangements if funding, ownership or key technology changes.

Plan an exit before entry. Confirm export formats, deletion timing, integration removal, account closure and the work required to move to another service. A low monthly subscription can still create a costly lock-in if your data, workflows and staff knowledge cannot be transferred.

Red flags that should pause the purchase

Other warning signs include unclear ownership of outputs, no practical export path, weak administrator controls, unexplained subprocessors, broad rights over customer content, no incident contact, expired assurance reports, or a product marketed for high-impact decisions without information about testing and human oversight.

No single red flag automatically proves that a vendor is unsuitable. It does mean the uncertainty should be resolved, the use narrowed, stronger safeguards negotiated, or the decision escalated before signing.

A practical approval record

Keep a short record containing the use case, review tier, data categories, evidence received, unresolved issues, safeguards, approver and review date. Add conditions such as “no customer records”, “human approval required” or “pilot limited to ten users”. Set a future review for contract renewal, a major product change, a new integration or a move into more sensitive work.

Due diligence is not finished when the contract is signed. Vendor terms, subprocessors, models and business use can all change. A lightweight annual review, plus event-based checks when something material changes, is often more useful than a large questionnaire completed once and forgotten.

Related reading: our AI governance by region.

Free tools: AI Privacy Risk Scorer to score your current AI tool setup against data-privacy best practice | AI Compliance Checker to check whether your AI tools meet your compliance obligations.

How long should AI vendor due diligence take?

A Tier 1 review may take a few hours if the vendor publishes clear documents. Tier 2 usually requires questions, evidence review and a trial. Tier 3 should take as long as needed to involve the appropriate privacy, security, legal or sector specialists. The depth should follow the risk, not the subscription price.

Is a security certification enough to approve an AI vendor?

No. A certification can provide useful evidence, but you still need to confirm its scope, date and relevance to the service you will use. It does not answer every question about model training, data retention, contract rights, output reliability or your specific use case.

Should a small business send every vendor a long questionnaire?

Not necessarily. Use a short core checklist for low-risk tools and reserve detailed questionnaires and specialist review for sensitive data, important integrations or high-impact decisions. A proportionate process is more likely to be followed consistently.

Who should approve an AI vendor?

The business owner for the use case should be accountable, with input from whoever manages operations, technology, privacy, security, procurement or legal matters. In a small business, one person may cover several roles, but the decision and any conditions should still be recorded.