Guide

Vendor diligence for AI tools: the questions that matter

AI tools are bought like other software, and the questions that matter most often go unasked: what happens to our data, what the tool can do on its own, which systems it needs, how we hear about changes, and how we get out. This page sets out those questions and who should judge the answers.

The short answer

Short answer

Before buying an AI tool, ask five sets of questions: what happens to your data, what it can do without a person, which systems it needs, how you hear about changes, and how you leave.

The owner of the process it will touch decides whether the answers are good enough, advised by security and legal.

Today

Definition

AI vendor diligence: the review a company carries out before an AI tool touches its data or processes, covering data use, autonomy, system access, change notice and exit, with a named owner deciding whether the risk is acceptable.

What gets missed when AI tools are bought like any other software?

The usual software checklist covers price, features and a security questionnaire. AI tools add questions it does not ask: whether your inputs train the vendor's models, whether the tool can act on its own, what happens when the vendor swaps the model underneath, and whether the output can leak what you put in. Those questions go unasked, and the answers arrive later as surprises.

Who should run this review?

The owner, operations lead, IT lead or procurement lead choosing an AI tool that will touch company or customer data. It matters most for tools that read customer records, draft anything sent outside the company, or take actions in your systems.

Which questions matter, and why?

Area	Questions to ask the vendor	Why it matters
Your data	Is it used to train models? Where is it stored, and for how long?	What goes in can come back out
Autonomy	What can the tool do without a person approving it?	Actions are harder to undo than drafts
Access	Which of our systems does it need, and with what permissions?	Broad access widens every other risk
Change	How are we told when the model or its behaviour changes?	The tool you approved may not stay the same
Exit	How do we get our data back, and is it deleted when we leave?	Leaving should not mean losing control

How it works, step by step

Diligence on an AI tool is a short, repeatable review, and an agent can carry most of the paperwork.

01

Name the owner and the job

Decide which process the tool will touch and who owns that process. That person decides at the end.

02

Send the questions

Send the vendor the questions in the table above, adjusted to the job, along with your usual security questionnaire.

03

Gather the evidence

Collect the vendor's answers, their published terms and data policies, and any independent reports they offer, in one file. An agent can assemble and compare these, quoting each source.

04

Mark the gaps and conflicts

List every question left unanswered and every place where the answers and the published terms disagree.

05

Human gate: the process owner decides, with advisers

The process owner reviews the file with security and, where data or contracts are involved, legal. They decide whether to buy, buy with conditions, or walk away, and record why. NIST's AI RMF Playbook frames this kind of control under its Govern function, which calls for policies and procedures that address AI risks associated with third-party entities, including risks of infringement of a third party's intellectual property or other rights.

06

Set conditions and a review date

Any conditions, such as data never used for training or a person approving every action, are written into the purchase. A date is set to review the tool again.

07

Recheck when the vendor changes

When the vendor announces a model, terms or pricing change, the review is reopened for the parts that changed.

Making the review count

What makes the review worth doing?

Questions about data first

OWASP's GenAI Security Project notes that users need to understand the risks of unintentionally providing sensitive data, which may later be disclosed in the model's output.

Autonomy in writing. What the tool may do without a person is agreed before it is switched on.

One decider

The process owner signs off, advised by security and legal.

Evidence, not assurances

Answers are checked against the vendor's published terms.

A review date. Approval is not permanent.

Where this fits

Concept

AI vendor diligence sits underneath every AI tool a company adopts, which is why it belongs with the Platform page. The paperwork side, sending questions and assembling answers, is a workflow in its own right, close to vendor onboarding without the email ping-pong, and answering the same kinds of questions about yourself is security questionnaire automation. The commercial side follows vendor negotiation prep and coordination, and the sign-off follows approval gates. The readiness checklist and the use case library help decide what the tool should do in the first place.

How does AI vendor diligence differ from ordinary software diligence?

Question	Ordinary software	AI tool
Your data	Stored and processed	Stored, processed and possibly used to train
Behaviour over time	Changes with releases you see	Can change when the model underneath changes
What it can do	What its features do	Possibly act on its own judgment
What stays the same	Security review and exit terms	Security review and exit terms

Security questionnaire automation: the same diligence, answered about yourself.
Vendor negotiation prep and coordination: turning diligence findings into terms.
Supplier scorecards unified across systems: rating vendors after you buy.

AAOS Platform

How It Works

Questions, answered

It helps, but it rarely answers the AI questions: training on your data, autonomy and model changes. Ask those separately.