← All posts

ai:vendors

AI vendor due diligence: what to ask before signing.

AI vendor due diligence asks the questions a standard security questionnaire misses when the product uses AI: what happens to prompts and uploaded files, whether the vendor trains on them, where processing happens, which model runs behind the feature, and what the contract says about exit and incidents.

This guide is for procurement, IT security and risk leads about to sign with a vendor whose product uses AI: a copilot, a chatbot platform, or an AI feature inside software you already license. Each section is questions to put to the vendor, and what to do with the answers.

What AI vendor due diligence adds to a normal vendor review.

A normal vendor review covers uptime, support, financial health and general security. An AI product adds questions of its own, because the product sends your content to a model, and the model, its version and its location can change under you. This review adds those questions and records the answers where the contract can point.

The pattern is not new to AI risk frameworks. The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) puts third-party AI in its Govern function: GOVERN 6.1 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, and GOVERN 6.2 calls for contingency processes to handle failures or incidents in third-party data or AI systems deemed to be high-risk. The answers you collect before signing give those policies their content and become the risks you monitor after.

Your data: training, retention, and who at the vendor can read it.

Three questions, in this order: does the vendor train its models on what you send; how long does it keep it; and who at the vendor, and at its contractors, can read it. Ask per plan you are buying, since terms can differ between consumer, business and enterprise tiers.

Two vendor pages show the kind of statement to look for. OpenAI's enterprise privacy page says that by default it does not use your business data for training, that you own your inputs and outputs where allowed by law, that Enterprise plans let you control how long your data is retained, and that access to stored API business data is limited to authorized employees for support, abuse investigation and legal compliance, plus bound contractors reviewing for abuse and misuse (checked 17 November 2026). Microsoft's Copilot privacy page states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs (checked 17 November 2026). Your vendor's page may say less; record exactly what it does say.

If the training answer is yes, ask which plan changes it and whether the control is contractual or a setting. Retention answers come with a number; get it in writing. Access answers name roles, and no answer is itself a finding for the record.

Where processing happens: storage, inference, sub-processors.

A region on a pricing page says where data is stored. It does not say where the prompt is processed: for hosted models the text travels to wherever the model runs, and logs and retrieval indexes add locations of their own. Our guide to AI data residency in the UAE works through the question, storage region against inference region. Ask for both, and for the sub-processor list and how you are told when it changes.

Microsoft's Copilot page, for one, points to its documentation of models provided by OpenAI and by Anthropic as subprocessors within Copilot experiences (checked 17 November 2026). If the vendor cannot name the model provider behind its product, that is an answer to record.

For UAE entities the transfer rules come from Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data (PDPL): Article 22 permits transfers where a proper protection level exists, and Article 23 lists alternatives such as contracts with equivalent safeguards. We describe the text and leave its reading to your counsel. Government entities sit outside the federal law under Article 2; for them, our guide to scoping AI work before a tender covers the earlier stage, and these questions apply at contract time.

The model: which one, who hosts it, how you hear it changed.

Ask which model version answers your traffic, who hosts it, and what changes without notice. Models are replaced on the vendor's schedule, and the model that passed your evaluation may not be the one that answers next quarter. Microsoft's Copilot page says the models that power Copilot are regularly updated and enhanced, and that model updates do not change security, privacy or compliance settings (checked 17 November 2026); that is one vendor's statement of the pattern. Your questions: does the vendor publish model changes, how far ahead, and can you pin a version for a defined window?

Then test the outputs yourself, on your own work: a small fixed set of real tasks, in every language your teams work in, re-run when the model changes. The result is a baseline you own and a check you can repeat at renewal.

Security evidence to ask for.

Ask for the certifications and reports the vendor can show, and read the scope. Two to know by name. ISO/IEC 42001:2023 is an international standard that specifies requirements for establishing, implementing, maintaining and continually improving an Artificial Intelligence Management System within organizations, and it is designed for entities providing or using AI-based products or services. SOC 2 is an examination performed by a licensed, independent CPA under the AICPA attestation standards that evaluates controls relevant to the security, availability, processing integrity, confidentiality or privacy of an organization's systems.

Whatever either covers, ask for the answers on data use, processing location and model changes in writing. When a vendor has neither, ask for what exists: the security page, the latest report or audit summary, who can access customer data, and the sub-processor list. The gap goes into the risk record with a decision someone signs.

Contract terms: exit, incident notice, liability, the off switch.

Put the findings into the contract. Four terms:

  • Deletion at exit. What the vendor deletes when the contract ends, within how many days, and what it certifies. Ask the same about backups and about any fine-tuned models built on your data.
  • Incident notice. A number of hours or days for telling you about a security incident that touches your data, and a channel someone answers at that hour.
  • Liability for outputs. Who carries what when the model gets something wrong in work you deliver with it. Read how far the contract limits it before the output reaches client-facing work.
  • The off switch. The right to switch the AI feature off without losing the rest of the software. AI features arrive inside software you already license; the question is whether they can be separated.

After signing: the tool goes through your approval process.

Once signed, the tool still goes through your AI model approval process, tiered by consequence, with the evidence recorded and a model change first on the re-approval triggers. NIST AI RMF 1.0's MANAGE 3.1 says AI risks and benefits from third-party resources are regularly monitored, and risk controls are applied and documented; the answers you collected at signature are what you monitor against.

If you are still choosing which AI tools to adopt, out:grow's AI readiness assessment audits how your teams actually work, the data AI would touch and the tools already in use, then hands you a written roadmap of the tools to use and the order to roll them out in, inside your existing security review, procurement and data-residency rules.

Do it yourself, or bring us in.

Everything above can be run in-house by procurement with IT security and the data owner: pull the vendor's published pages, put the questions in writing, and hold signature until each has an answer on record. The output is a one-page risk record the contract can point to.

Outside help fits when the answers need checking against your own setup: where inference actually happens in your deployment, and how the model does on your own work in the languages your teams use.

Book a 30-minute call

Not ready for a call? Take the 10-question AI readiness check

Frequently asked questions.

Is a SOC 2 report enough?

A SOC 2 report covers the controls in its stated scope, relevant to security, availability, processing integrity, confidentiality or privacy, examined by an independent CPA. Check whether the AI feature is inside that scope, and ask the AI questions (training on prompts, processing location, model changes) in writing either way. Treat the report as one input.

What if the vendor will not name its model provider?

Record the refusal as the answer. You can still contract on outcomes: where processing happens, deletion at exit, incident notice, liability. Count the unnamed provider as a risk when you set the approval tier and the data you allow.

Does a UAE cloud region settle data residency?

It settles storage. Processing follows the deployment, and logs and sub-processors add locations of their own; our residency guide works through the settings that decide. Ask for the processing answer in writing rather than the region alone.

Who in the organization should own the review?

Procurement runs it, with IT security and the data owner for their parts, and one named owner assembles the answers into the record that goes to signature. The same owner holds the renewal date and the re-approval triggers.

What about AI features added to software we already use?

The same questions apply, at renewal if not before. Ask whether the feature can be switched off without losing the software, what it does with your content, and put the tool through the approval process with its own data ceiling.

Keep reading.

:

How to run an AI model approval process before production.

:

AI data residency in the UAE: where your prompts are processed.

:

What data employees can put into AI tools: a classification guide.