VENDOR AI DUE DILIGENCE
The vendor says it uses AI. Here are the questions to send back.
A vendor adds AI to a product you already pay for, or sells you an agent, and procurement has nothing to send back except a security questionnaire written before any of this existed. Answer nine questions about what you are buying, what it will touch, whether it acts, where it runs and which rule binds you, and this page writes the questionnaire: the questions to send, grouped by topic, each with why it is asked, what a good answer looks like and the red flag; the ones that decide the purchase, to read first; and a scoring method with no percentages in it. Download it as Word to send to the vendor, or as PDF, Markdown or JSON. Not legal advice; counsel reads the contract questions before they go out. No email is needed.
NINE QUESTIONS
Answer nine questions about the purchase. The questionnaire writes itself.
Every answer is a choice except the two names, which you can leave blank. The questionnaire is written on this page the moment you finish and gets an address of its own; the Word file is the one you send. Do not enter credentials or customer records.
Keep a copy in your inbox.
Enter a work email, company and role to receive this questionnaire as text with its address. This records an SIT inquiry for follow-up, and the engineer reads it before any call. Marketing emails require a separate choice.
YOUR REGISTER
What this browser has worked out so far.
Kept in this browser only; nothing on this list is stored on our server. Each line opens its own address, so you can come back to it, send it to the owner, or clear it. The procurement pack binds every line into one PDF, rebuilt from the addresses when you press the button, with a cover and a contents list.
WHAT IT IS
What is vendor AI due diligence?
The written questions that establish what of yours the vendor's AI will read, whether it trains on it, where it runs, what it may do on its own, who can stop it, and what the contract says when any of that goes wrong.
The old security questionnaire asks about firewalls and badges. It does not ask whether your prompts train a model, whether the agent runs on a person's login, or whether there is a stop path you control. This page writes the questions that do, with the evidence to expect under each one.
THE INSIGHT
Why does every question need a "good answer looks like" line?
Because the vendor's reply is read by someone who is not the expert, and "we take security seriously" and "a SOC 2 Type II report naming the AI service, under NDA, within a week" look the same to them without it.
Copilot Cowork carries out multi-step tasks across Microsoft 365 and pauses for the user's approval before sensitive actions such as sending an email or posting in Teams. (Microsoft Learn, checked Oct 31, 2026)
Write the good answer and the red flag beside the question and procurement can judge the reply. Leave them out and the vendor answers the question it wishes you had asked.
HOW IT WORKS
How does the builder write the questionnaire?
From a question bank, not a model: each of the nine answers switches questions on or off and marks the ones that decide the purchase for your situation, so every question traces to an answer and the deciding ones are read first.
- You answer nine questions on the page; nothing is sent anywhere.
- The page writes the questionnaire in up to nine sections, each question with why it is asked, a good answer and the red flag, then the deciding questions and a scoring method.
- You download it as Word with an answer line under every question, send it to the vendor, and score the replies as evidence, words only, or not answered.
- Counsel reads the contract questions before they go; the builder is not legal advice.
AGENTS
What changes when the vendor sells an agent that acts?
Five questions switch on and become deciding: its own identity, the approval step per action type, the stop path you control, the test against instructions hidden in content it reads, and how an action is found and reversed.
An assistant that drafts can be wrong. An agent that sends, files, pays or changes access can be wrong in your name, so the vendor's answers on those five decide whether it can be governed at all.
YOUR DATA
What happens to the answers?
They are rendered on the page and encoded in the questionnaire's address, and nothing is stored unless you ask for the email, in which case the lead is recorded through the same intake as the contact form.
The only free text is the vendor's name and your organization's name, and both can be left blank. Do not enter credentials or customer records.
THE MONEY
What is free and what is paid?
The questionnaire is free. Reading the vendor's answers against your own tenant, and writing the approval step and the inventory row the purchase needs, is the ten-hour foundation, and anything built after that is scoped from it.
No price appears on this site; the engagement is quoted after a scoping call, assessment separate from implementation.
AFTER TWO WEEKS
What will be working after two weeks?
- You hold a questionnaire written for this purchase, in up to nine sections, with the deciding questions marked and read first.
- Every question carries why it is asked, what a good answer looks like and the red flag, so a non-expert can judge the reply.
- The Word file has an answer line and an evidence line under every question, ready to send.
- The scoring method has three outcomes per answer and no percentages, and says what each outcome means for the pilot.
HONEST ABOUT LIMITS
What this does not do yet
It does not know what the vendor will say. Agents built with Microsoft Foundry, Copilot Studio and Agent Builder are integrated with Microsoft Agent 365 automatically; agents built elsewhere join through registry sync or the Agent 365 SDK. (Microsoft Learn, checked Oct 31, 2026) An agent the vendor registers there is one your own controls can see; the questionnaire asks whether it will be, and the foundation checks the answer.
It does not evaluate the vendor: it writes the questions, and the answers still have to be read.
It is not legal advice; counsel reviews the contract questions before they are sent.
THE ONE THING TO REMEMBER
A vendor answer in words is a promise; a vendor answer with evidence is a control.
Said another way: ask for the report, not the badge; the clause, not the policy page.
Said a third way: the questions are free; reading the answers against your tenant is ten hours.
DIRECT ANSWERS
Questions before you decide.
Do I need an email to see the questionnaire?
No. It renders on the page and has its own address. The email is only for keeping a copy and it records a lead through the normal intake.
Is the questionnaire written by an AI?
No. It comes from a question bank, so every question traces to the answer that switched it on. A model may tailor wording in a later version, under the same rules.
Can I send it as is?
Yes, as the Word file, after counsel reads the contract questions. Fill in the vendor contact, the reviewer and the answers-due date on the first page.
How do I score the replies?
Each answer is evidence, words only, or not answered. A deciding question without evidence means no pilot on real data. Three words-only answers in the data or agent sections mean a pilot on sample data only.
How does it relate to the agent inventory and the policy?
A vendor agent that passes becomes a row in your inventory with the owner, approval and stop path the vendor confirmed, and the policy's approved-tools list gains the product.
NEXT STEP
Start with a boundary you can explain.
Choose a focused check, or speak with the team about the workflow you need to improve.