AI vendor
questionnaire.
This questionnaire gives you 30 questions to ask any AI vendor or agency before you sign. The questions cover data and location, security, models and sub-processors, accuracy and testing, cost and contract, exit and ownership, and support. Each one shows what a good answer looks like and has a score column.
When should you send a vendor questionnaire?
Send it after a first call and before you ask for a final quote. By then you know which vendors are serious, and the answers can still change your choice. Send the same questions to every vendor on the shortlist, in writing, so that you can compare the answers.
The questionnaire works for an AI receptionist product, a WhatsApp agent platform, an agency that builds a custom agent, or a freelancer. It also works for AI features inside software you already use. The answers feed directly into your DPIA, because GDPR makes you, the controller, responsible for the processors you choose.
Under Article 28 of the GDPR, you may use only processors that give sufficient guarantees, and you need a written contract with each one. Sections 1 to 3 of the questionnaire collect the facts for that check.
Get the editable file.
DOCX · editable
AI vendor questionnaire (Word)
We use your email to send updates to this template. We do not share it.
The questionnaire in full
Send these questions to every AI vendor or agency on your shortlist. Ask for written answers. Score each answer: 0 = no answer or a poor answer, 1 = partly acceptable, 2 = a good answer. Replace every field in [square brackets]. Remove questions that do not apply.
About this review
| Field | Entry |
|---|---|
| Vendor name | [ ] |
| Product or service | [For example: AI receptionist for a dental practice] |
| Sent by and date | [Name, date] |
| Answers received | [Date] |
1. Data and location
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| Where do you store and process our data, including logs and backups? | Named countries or regions for each type of data, with logs and backups included. | [ ] | [ ] |
| Do you or your providers use our data to train AI models? | No, in writing in the contract, with the setting that controls it. | [ ] | [ ] |
| How long do you keep conversations, transcripts and files? | A fixed period that we can change, and deletion on request. | [ ] | [ ] |
| Will you sign a data processing agreement (DPA) under GDPR Article 28? | Yes, with their standard DPA available before we sign. | [ ] | [ ] |
| How do you transfer data outside the EEA, if you do? | The mechanism (adequacy decision or standard contractual clauses) and a list of the countries. | [ ] | [ ] |
2. Security
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| Who in your company can see our data, and how is access logged? | Named roles, access by need only, and access logs. | [ ] | [ ] |
| How do you store API keys and passwords for our systems? | In a secrets manager, never in code or prompts. | [ ] | [ ] |
| What permissions does the agent need in our systems? | A short list of the minimum read and write permissions, per system. | [ ] | [ ] |
| How do you protect against prompt injection and data leaks between customers? | Specific controls, for example tool permissions limited per user, and test results. | [ ] | [ ] |
| How and when do you tell us about a security incident? | A fixed time limit in the contract that lets us meet our own 72-hour breach duty. | [ ] | [ ] |
3. Model and sub-processors
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| Which AI models and providers do you use? | Named models and providers, not "our own AI" without detail. | [ ] | [ ] |
| Give us the full list of sub-processors. | A current list, with notice before they add a new one. | [ ] | [ ] |
| Can you change the model without telling us? | No. They tell us before a change and test it against our test set. | [ ] | [ ] |
| What happens if the model provider has an outage? | A fallback plan, for example a second model or a handover to staff. | [ ] | [ ] |
4. Accuracy and testing
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| How will you test the agent before launch? | A written test plan with our real examples, pass criteria and our sign-off. | [ ] | [ ] |
| What happens when the agent does not know the answer? | It says so and hands over to a person. It does not guess. | [ ] | [ ] |
| Can we see the conversations and the actions the agent took? | Yes, a full log that we can search and export. | [ ] | [ ] |
| How do you measure quality after launch? | Regular reviews of samples, with numbers they share with us. | [ ] | [ ] |
5. Cost and contract
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| What is the full cost for the first year, including setup, licences, usage and support? | One figure in euro, with the usage assumptions written down. | [ ] | [ ] |
| How do usage costs change if our volume doubles? | A clear formula or price table, and spend limits we control. | [ ] | [ ] |
| What is the minimum contract term and notice period? | Short terms for a first release, for example monthly after the setup. | [ ] | [ ] |
| Which parts of the work are fixed price? | A fixed price for a defined scope, with change requests priced before work starts. | [ ] | [ ] |
6. Exit and ownership
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| Who owns the prompts, the configuration, the code and the data? | We own our data and the configuration built for us, in writing. | [ ] | [ ] |
| If we leave, what do we get, and in what format? | An export of data, prompts, test sets and logs in standard formats. | [ ] | [ ] |
| How long after we leave do you delete our data? | A fixed period, with written confirmation of deletion. | [ ] | [ ] |
| Can another company take over the agent? | Yes, with documentation and a handover period. | [ ] | [ ] |
7. Support
| Question | A good answer looks like | Their answer | Score (0 to 2) |
|---|---|---|---|
| Who do we contact when the agent fails, and how fast do you respond? | A named contact, support hours and response times in the contract. | [ ] | [ ] |
| How do we pause the agent in an emergency? | A switch that our staff can use without the vendor. | [ ] | [ ] |
| How do you handle changes to our prices, services or opening hours? | We can change them ourselves, or they do it within an agreed time. | [ ] | [ ] |
| Can you give us two references from clients in Ireland or the EU? | Named clients who agree to a short call. | [ ] | [ ] |
Score summary
| Group | Maximum score | Vendor score | Notes |
|---|---|---|---|
| Data and location | 10 | [ ] | [ ] |
| Security | 10 | [ ] | [ ] |
| Model and sub-processors | 8 | [ ] | [ ] |
| Accuracy and testing | 8 | [ ] | [ ] |
| Cost and contract | 8 | [ ] | [ ] |
| Exit and ownership | 8 | [ ] | [ ] |
| Support | 8 | [ ] | [ ] |
| Total | 60 | [ ] | [ ] |
Rule: a score of 0 on any question about data location, model training, the DPA or data ownership stops the review, whatever the total.
How to use it: a dental practice compares two vendors
A dental practice in Cork has two quotes for an AI receptionist. Vendor A sells a packaged product. Vendor B is an agency that builds on a large model provider. The practice manager sends both the full questionnaire with a deadline of ten working days.
- Remove what does not apply. The practice deletes the question about changes to prices, because it does not publish prices by phone.
- Read for specifics. Vendor A answers "data is stored securely in the cloud". That scores 0: it names no country or provider. Vendor B names the region and the provider, and attaches its DPA. That scores 2.
- Apply the stop rule. Vendor A cannot confirm in writing that call recordings are not used for model training. Under the rule in the score summary, the review of Vendor A stops there, whatever its total.
- Call the references. Vendor B gives two references. The manager asks each one what went wrong in the first month and how fast it was fixed.
- Keep the answers. The completed questionnaire goes into the DPIA file as evidence for Parts C and F.
Red flags and common mistakes
- "Our own proprietary AI" with no model named. Most products use one of a few large model providers. A vendor that will not say which one cannot tell you where your data goes.
- No test plan. If the vendor plans to "tune it after launch", your customers do the testing. Ask for a plan like our go-live test plan.
- Usage costs with no formula. Ask what happens to the bill if volume doubles. See our guide to AI agent costs for the parts of a typical bill.
- No exit plan. If you cannot get your prompts, test sets and conversation data out, you cannot change vendor without starting again.
- Scoring on the sales call. Score only written answers. Spoken promises are not in the contract.
- Asking only about features. Features are easy to show in a demo. Data, security and exit are where projects get into trouble.
If you are still deciding between an agency, a freelancer and an in-house build, our comparison covers the trade-offs. Send this questionnaire to us too: we answer it in writing like any other vendor.
Common questions
How many vendors should we send the questionnaire to?
Two or three is usually enough for a small business. More than that costs you a lot of reading time. Shortlist after a first call, then send the same questionnaire to each shortlisted vendor with the same deadline.
What if a vendor refuses to answer in writing?
Treat it as a score of 0 for each question they do not answer. A vendor that will not put its data and security answers in writing will not want to put them in the contract either. That is a risk for you as the controller.
Is the scoring scale fair to small vendors?
Yes, because it scores the content of the answer, not the size of the vendor. A freelancer can give a specific, honest answer about data location and exit. The stop rule applies equally to everyone: data location, model training, the DPA and data ownership are not negotiable.
Do we need this for tools like ChatGPT or Copilot?
A shorter version, yes. For a standard tool, the answers are usually in the provider’s published business terms and DPA. Check the data and model training questions in sections 1 and 3, then record the result in the tool register of your AI use policy.
Sources
We last checked these sources on 19 November 2026. Rules, fees and prices change. Check the source before you act.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.