Picture three small companies, and three surprises.
A property management company adds an AI chatbot to its website. It answers tenant questions at night. One day a tenant asks, “Where do my chats go?” The owner reads the vendor’s terms for the first time. The vendor keeps every chat, with no end date.
A marketing agency gets a security form from a new client. One question says: “Does any AI vendor train on our data?” The agency uses an AI writing tool. The business plan says no training. But half the team uses the no-cost plan, and its terms say something else.
An online store connects an AI agent to its order system. The agent answers refund requests. One evening, a bug makes it approve every refund. Nobody can find the switch to stop it. It runs for 40 minutes.
Nobody in these stories did anything wrong on purpose. They all missed one step: nobody asked the vendor the right questions first.
That step has a name: an AI vendor review. Today I am sharing a template for it, with an Excel file that does the counting for you.
What an AI vendor review is
Think about a house cleaner. Before you give a stranger the key to your house, you ask a few questions. Do you send someone else in your place? Do you make copies of the key? What happens if something breaks? How do I cancel?
An AI tool that sees your data holds a key to your house. A vendor review asks the same kind of questions, before you hand over the key. The template groups them into four:
From the template, section 1:
1. What does the vendor do with our data?
2. How does the vendor protect it, and what happens when something goes wrong?
3. Can we control the AI, check its output, and switch it off?
4. Can we leave, and take our data with us?
Why AI vendors need their own questions
You may already check your normal software vendors. AI vendors bring three new problems.
- Your data can teach their AI. Some AI tools use your inputs to train their models. Your text could shape answers for other customers.
- There is often a second company behind the first. Many AI tools send your data to an AI model company, for example through an API. You signed with one company, but two companies see your data.
- The AI can change without notice. A vendor can switch to a new model overnight. Yesterday’s careful answers can become today’s mistakes.
Go back to the house cleaner. Problem 2 is the cleaner who sends a subcontractor that you never met. You want to know who that person is.
Not every tool needs every question
At a hospital emergency room, a nurse sees every patient first. A sprained finger waits. Chest pain goes straight in. People call this first check triage. It sends each patient to the right level of care.
The template works the same way. Before you send any questions, you answer six questions about your own use of the tool:
From the template, section 5:
– T2. Will the tool see Restricted data, for example personal data, customer data, financial data, or health data?
– T3. Will customers, patients, or the public see the output or talk to the tool?
– T5. Can the tool take actions or connect to our systems, for example send messages, change records, or read our files?
That is half of the list. Your answers set one of three review levels, called tiers:
- Light: 6 questions. For example, a writing tool that sees only public information.
- Standard: 27 questions. For example, a chat assistant that sees internal drafts.
- Deep: all 42 questions. For example, a support chatbot that reads customer records and answers customers.
The Excel file picks the tier for you. It also marks which questions to send. A low-risk use gets a short list, so the vendor answers faster. A risky use gets the full review that it deserves.
One rule makes the triage safe: “If you are not sure about a triage answer, answer ‘Yes’.”
What the questions cover
The 42 questions sit in ten sections. Here they are in plain words:
- A. The product and its AI: which features use AI, and which AI models they use.
- B. Our data: training, how long they keep your data, who else gets it, and how they delete it.
- C. Security: logins, separation from other customers, and attacks on the AI.
- D. Model changes: notice before the AI changes.
- E. Output quality: how they test the answers, and whether a person can check them.
- F. AI agents: what the AI can do in your systems, and how you stop it.
- G. Incidents: what happens when something goes wrong, and how fast they tell you.
- H. Contracts and compliance: the agreements that your laws require.
- I. The vendor’s own AI rules: a written policy and a person in charge.
- J. Exit: how you leave with your data.
Here are the first two data questions, exactly as the vendor sees them:
From the vendor questionnaire, section B:
B1 (Light). Do you or your model providers use our inputs, outputs, or files to train or improve AI models? If yes, can we switch this off?
B2 (Light). Where is your answer to B1 written: in our contract, your terms, or a policy page? Name the document and the section.
Look at B2. It asks the vendor to show where it wrote the promise. That brings us to the most important rule.
A promise counts only if it is written down
A friendly sales email says, “We never use your data.” That is nice to hear. But a sales email is not a contract.
From the template, section 7.2:
3. For each deal-breaker, the answer must point to a contract, the vendor’s terms, or an independent report. A sales email or a marketing page is not enough.
4. If the vendor’s answer and its contract disagree, the contract counts.
This is how the marketing agency in the second story gets its answer. It checks the terms of each plan, not the sales page. Then it moves the whole team to the business plan.
“We don’t train on your data” is only half the answer
Imagine a wedding photographer. They promise never to use your photos in their ads. Good. But they keep every photo forever, in an old laptop at home. The promise is true, and your photos can still leak.
AI vendors work the same way. A vendor can promise no training and still keep your data for years. The template catches this:
From the template, section 7.2:
5. “We do not train on your data” does not mean “we do not keep your data”. Read the answers to B1 and B4 together.
B4 asks how long the vendor and its model providers keep your inputs and outputs. That was the property management company’s problem in the first story.
Eight answers that stop the deal
A house inspection finds many small problems. A squeaky door does not stop the sale. A cracked foundation does.
The template has eight “cracked foundation” questions, called deal-breakers. If the answer to one of them is “Not acceptable”, you do not approve the use as planned. Here are three of them:
From the template, section 6:
1. B1. Training on our data. The vendor or its model providers train AI models on our data, and we cannot switch it off.
2. B6. Sub-processors. The vendor does not name the sub-processors that receive our data, including AI model providers.
8. F4. Switch-off. We cannot switch off the AI feature or an agent quickly.
A sub-processor is another company that handles your data for the vendor. It is the subcontractor in the house cleaner example. Often it is the AI model company behind the product.
F4 is the online store in the third story. An agent that you cannot stop is a cracked foundation.
A failed deal-breaker does not always mean “no”. The template offers another path: approve a smaller use. For example, you use the tool, but without customer data.
Four simple marks for each answer
When the answers arrive, you give each one a status:
- Acceptable: the answer meets your needs for this use.
- Follow-up: the answer is missing, unclear, or needs a document. You ask again.
- Not acceptable: the answer does not meet your needs for this use.
- Not applicable: the question does not apply to this tool. You write down why.
Note the words “for this use”. A long retention period can be fine for public text. The same period can be a problem for customer records.
The Excel file counts the marks. Then it shows one result, for example “Wait: answers missing or follow-up open” or “Do not approve as planned: a deal-breaker failed”.
The review feeds your AI tool register
In my last template, I shared an AI tool register: a list of every AI tool that your company uses. The vendor review is the step before a tool gets its “approved” mark.
From the template, section 1:
This review feeds our AI Tool Register. The register records the decision. This review records the reasons.
When an insurer asks, “Why did you approve this AI tool?”, you have the answer in one folder.
Four versions: choose the one closest to you
Every industry asks different questions. So the template comes in four versions:
- Generic, for any company: placeholders that you fill in.
- Healthcare clinic group, for clinics and medical groups: patient data agreements that cover the AI, and questions about recording patient visits.
- B2B AI software company, for companies that build on AI model providers: data retention, regional processing, and notice before a model goes away.
- Consulting firm, for consultants and professional services: client data, admin controls, and AI notetakers.
The three industry versions use fictional companies, so you can see the questions in a realistic business.
Start in one afternoon
- Pick one AI tool. Choose the one that sees your most sensitive data.
- Download the Excel file and the guide. Choose the version that is closest to your company.
- Answer the six triage questions with the person who uses the tool. The file picks the tier.
- Send the questions for that tier to the vendor, with a deadline of 10 business days.
- Copy the answers into the Excel file when they arrive, and give each one a status.
- Send one follow-up email for every “Follow-up” answer.
- Decide, and record the decision in your AI tool register.
Steps 1 to 4 take one afternoon. The vendor needs time to answer, so steps 5 to 7 come a week or two later.
Download the template
The templates are on GitHub under the CC BY 4.0 license. You can use and adapt them, also for commercial work.
- All templates: github.com/nosistech/governance-templates
- Generic questionnaire (Excel): download
- Generic review guide (Word): download
- Generic vendor questionnaire (Word): download
The healthcare, software, and consulting versions are in the same collection.
Ask first, then trust
Go back to the three stories. The property management company now asks about retention before it signs. The marketing agency moved everyone to the plan that has the written promise. The online store asks every agent vendor one question first: “How do we switch it off?”
A vendor review does not stop you from buying AI. It turns “the sales rep said it’s safe” into answers that you can show to anyone who asks.
This is the third of four templates. The last one is an AI incident-response playbook: what to do when an AI tool goes wrong.
In my next post, I go deeper. I explain the design decisions behind this questionnaire, the alternatives that I rejected, and the public guidance behind each rule. That post is for IT, security, and compliance people.
Have a question or feedback about the questionnaire? Contact NosisTech.
Keep reading: AI governance templates
Next article: Inside My AI Vendor Questionnaire: 10 Design Decisions and the Alternatives I Rejected
Template 2: Which AI Tools Does Your Company Use? Build a Register in One Afternoon
Template 1: Your Team Already Uses AI. Here Is a Policy That Makes It Safe
Topic hub: AI Governance
All articles: NosisTech Writing
This post is general information, not legal advice. The opening stories and the example companies are fictional. Any resemblance to real people or companies is a coincidence. Laws about AI change often and differ by state and country. Ask your legal counsel before you rely on any template.