AI vendor due diligence: the questions to ask before you buy
By the InTheory team · Updated November 2026
Before a small business signs with an AI vendor, it should understand what happens to its data, how secure the product is, how accurate it is on real work, what the contract says, and how to leave. This guide groups the essential questions into six areas, explains why each matters, and notes which answers should stop a deal.
- Ask in writing whether your data is used for training, how long it is kept and who can access it.
- Test accuracy yourself on your own examples rather than relying on demos or vendor claims.
- Check the contract for data ownership, liability, required agreements such as a BAA, and exit terms.
- If the tool scores or ranks people, ask about Colorado SB 26-189 documentation duties from 2027.
Why due diligence matters more with AI vendors
Buying AI software involves the usual questions about price and support, plus a few new ones. AI tools often process large amounts of your client and company information, may send it to third-party model providers, and produce outputs that can be wrong in ways that are hard to spot. Terms also change quickly as vendors update models and plans.
For regulated firms, your obligations follow your data. The FTC Safeguards Rule, for example, requires covered businesses such as tax preparers to take steps to ensure that their service providers safeguard customer information. Health practices need a business associate agreement with any vendor handling protected health information. Due diligence is how you meet those obligations in practice.
Small firms sometimes skip this because the subscription is cheap. But the risk scales with the data, not the price. A $20-a-month tool holding client files carries the same exposure as an expensive one.
Treat it as routine, not as a sign of distrust.
1. Data use and training
Start here, because these answers decide whether the tool can be used with client data at all.
Ask for the answers in writing, ideally with links to the relevant terms or trust documentation. A sales rep's verbal assurance is not something you can rely on later.
- Is our data, including prompts, uploads and outputs, used to train your models or anyone else's? Is that the default, and where is it stated?
- Which third-party AI model providers process our data, and under what terms?
- How long do you retain our data, and can we set or shorten that period?
- Can we delete specific records, and our entire account, on request?
- Where is our data stored and processed geographically?
2. Security and access
Small vendors can be excellent, but ask the same questions you would ask a large one.
If the vendor uses other AI providers behind the scenes, as many do, the answers about training and retention need to cover those providers as well. Ask specifically, since a vendor's own policy may not describe what its model provider does with the data.
- Do you support single sign-on and multi-factor authentication, and can we enforce them?
- Who at your company can access our content, and how is that access controlled and logged?
- Do you have an independent security assessment or certification you can share, such as a SOC 2 report?
- How do you handle and notify us of security incidents?
- Can admins see activity logs and remove users immediately?
3. Accuracy, limits and human review
Demos show the best case. Your decision should rest on how the tool performs on your work.
During the trial, have the people who will use the tool log every correction they make. Ten to twenty real examples is usually enough to see whether accuracy is acceptable for your use.
- Can we run a trial on our own real examples before committing?
- What are the known limitations, and in what situations should the tool not be used?
- How does the product support human review, such as approval steps, drafts or confidence flags?
- Does it show sources or cite the documents it used, so reviewers can check answers?
- How do you notify customers of model changes that could affect output quality?
4. Contracts, compliance and decisions about people
Have someone with contract experience read the terms. Points to check:
If you are not sure whether a tool counts as influencing a consequential decision, describe how you plan to use it and ask the vendor how it is marketed and configured. The developer duties in SB 26-189 depend partly on how the tool is intended to be used, so the vendor's answer is relevant.
- Who owns the outputs and any configuration we create?
- Will you sign the agreements our obligations require, such as a HIPAA business associate agreement or a data processing agreement?
- What are the liability limits, and what indemnities are offered?
- What happens to our data if you are acquired, change model providers or shut down?
- Does the tool score, rank or recommend anything about individuals for hiring, housing, credit, insurance or health care? If so, will you provide the developer documentation Colorado's SB 26-189 requires from January 1, 2027, including intended uses, known limitations, training data categories and human review instructions, and notify us of material changes?
- Note that SB 26-189 makes indemnity clauses void where they try to shield a party from liability for its own discriminatory use of such tools in violation of Colorado anti-discrimination law. Ask counsel how that affects your contract.
5. Integration, cost and exit
The last set of questions covers how the tool fits your systems and what happens if you leave.
Year-two cost deserves attention. Introductory pricing, per-seat growth and usage charges can change the economics noticeably after the first renewal.
Ask for a sample export during the trial, so you can see the format before you depend on it.
- Which of our systems does it integrate with natively, and which need an automation platform or custom work?
- What will it cost at our size in year one and year two, including setup, add-ons and usage charges?
- How much notice do you give before price or plan changes?
- Can we export all our data in a usable format, and how long do we have after cancelling?
- What is the minimum contract term, and can we start month to month?
Answers that should stop a deal
Some answers are red flags for client or company data: the vendor cannot say clearly whether your data trains models; there is no business plan with admin controls; it will not sign a business associate agreement when you need one; there is no way to export your data; or it will not allow a trial on your own examples. Vague answers to security questions are also a warning sign.
Keep the answers with your approved tool list and revisit them at renewal, since terms change. InTheory runs vendor evaluations against questions like these as part of technology vendor selection work. We do not resell software, so the recommendation is based only on fit, and the legal review stays with your counsel.
A short, consistent questionnaire also saves time. Send the same questions to every shortlisted vendor, and comparing answers becomes straightforward.
Common questions
What is the most important question to ask an AI vendor?
Whether your data is used to train models, and where that is stated in writing. The answer decides whether you can use the tool with client information. Close behind are how long data is kept, who can access it, and whether the vendor will sign agreements your obligations require.
Do small vendors deserve the same scrutiny as large ones?
Yes, and the questions are the same. Small vendors can be responsive and well run, but they may have fewer formal certifications. If a vendor lacks something like a SOC 2 report, ask how it handles security in practice and weigh that against the sensitivity of the data involved.
How long should AI vendor due diligence take?
For a small business evaluating one tool, a week or two is typical: send the questions, review the answers and terms, and run a trial on real examples at the same time. Higher-risk uses, such as tools handling health information or decisions about people, deserve more time and legal review.