AI Vendor Review Policy

ai-vendor-review.md·SOC 2 · CC9.2

Generate an AI Vendor Review Policy covering diligence before onboarding an AI provider, data-retention and training-use terms, and the re-review cadence reviewers expect.

Generate this policy Browse all policies

What's in the policy

Defines how AI vendors and subprocessors are assessed before they touch your data, and re-assessed after.

✓Diligence before onboarding an AI vendor

✓Data retention and model-training terms to confirm

✓Security-posture evidence to collect from the vendor

✓Subprocessor disclosure

✓An inventory of approved AI vendors

✓Annual re-review and material-change triggers

Mapped toSOC 2 (CC9.2)ISO 42001

What the document actually says

These are the standing requirements every generated copy of the document is written from — the clause text itself, not a description of it. The document you receive is drafted on top of them: it expands on them and fills the current-state detail from your intake, and the generator is instructed to keep them, to state them fully, and never to claim a certification or invent a fact you did not supply.

The five things checked before an AI vendor is approved

Before an AI vendor is approved for use with company or customer data, the following are assessed and documented: - **Training use** — does the vendor train models on submitted data? A **no-training commitment** is required for any use involving customer or confidential data. - **Data retention** — how long is submitted data retained, and can retention be limited or disabled? - **Security posture** — evidence such as SOC 2 / ISO 27001, encryption in transit and at rest, and access controls. - **Subprocessors & location** — downstream subprocessors and data-processing locations. - **Contractual terms** — DPA, confidentiality, and breach-notification commitments.

SOC 2 CC9.2 · ISO 42001 — five named criteria a general vendor review does not ask: whether the model trains on submitted data, how long it keeps it, its security evidence, its own subprocessors and locations, and the contract terms behind all four.

Approval before access, and what re-review actually checks

- AI vendors handling customer or confidential data are approved before use and recorded in the approved-vendor inventory. - Approved AI vendors are **re-reviewed at least annually** and upon any material change to their terms or a relevant security event. - Vendors that cannot meet the data-handling requirements are not used with restricted data.

SOC 2 CC9.2 · ISO 42001 — approval sits before access rather than after, the re-review clock is named rather than implied, and a vendor that cannot meet the requirements is barred from restricted data outright, not approved with a caveat.

OpenAI — support summarisation, no-training via API terms; Anthropic — code assistance, zero-retention endpoint

Filled in for a real stack

Placeholder

Approved AI vendors: To be confirmed. Customer data sent to AI tools: To be confirmed.

Resolved

Approved AI vendors: OpenAI, used for support-ticket summarisation under API terms that disable training on submitted data; Anthropic, used for code assistance through a zero-data-retention endpoint. Customer data sent to AI tools: ticket text only, routed to OpenAI — no customer data reaches Anthropic.

What the auditor asks for alongside it

A current inventory of AI vendors, each one tagged with what it may see.

The AI Vendor / Subprocessor Summary, built from the same approvedAiVendors field this policy's own inventory prints — a structured vendor-and-purpose table when entered one per line. The Vendor Management Policy has no equivalent: its own vendor section prints the keyVendors field as plain text, whatever you typed, with no parsing into a table at all.

Whether a specific AI vendor trains on the data you send it, and who confirmed that.

The drafted answer to 'Do you train models on customer data?' — a question this document's own criteria exist to let you answer honestly. Where the customerDataToAi field reads a clear 'no', the draft states the no-training position and cites the AI Vendor Review Policy and the AI Customer Data Handling Statement as its source.

The customer-facing statement about what happens to their data once it leaves your systems for a model.

The AI Customer Data Handling Statement — the artifact 'Is customer data sent to AI subprocessors?' resolves to on a questionnaire. It restates the same customerDataToAi answer and notes that approved vendors are reviewed for retention and training terms, which is this policy's own subject.

Which AI vendors sit outside your systems, drawn rather than named in a list.

The same architecture diagram this policy's inventory feeds. Each vendor you list at intake becomes its own box inside an 'AI Subprocessors' group, separate from the ordinary third-party vendors — so a reviewer can see, rather than take your word for, which providers actually receive customer data and which are simply on a payroll list.

Where these documents go stale

An AI vendor review ages faster than a general vendor review, because the terms behind it are the part a provider can change without telling you. A model provider updates its API terms and a default flips from opt-out to opt-in training, and nothing about your own systems changed to trigger a second look. A team on a different plan or a newer API version can end up under different retention terms than the ones actually reviewed. New subprocessors get added to a provider's own stack the same way any subprocessor does — quietly, and disclosed, if at all, in a document nobody re-reads after onboarding. A vendor swapped for a cheaper one mid-project rarely restarts the five-criteria review; it inherits the approval of whatever it replaced. Re-validate by re-checking each approved vendor's current terms against the five criteria, not by trusting the review that approved it originally.

From intake to enterprise-ready in three moves

Tell us about your stack

Answer a short intake — cloud, data types, tools. No agents to install.

We generate a tailored draft

Not a blank template: a document written for your environment and pre-mapped to controls.

Review, edit, and share

Export it or attach it straight to an enterprise security review or questionnaire.

AI Vendor Review Policy, answered

What is an AI vendor review policy?

An AI vendor review policy states how a company evaluates an AI vendor or subprocessor before sending it company or customer data, and re-checks it afterward. It exists because the questions that matter for a model provider — does it train on your data, how long does it keep it — are not questions a general vendor review asks.

What does the review actually check?

Five things: whether the vendor trains models on submitted data, how long it retains that data, evidence of its security posture, its own downstream subprocessors and processing locations, and the contract terms — DPA, confidentiality, breach notification — behind all of it. A vendor is documented against all five before it is approved.

Why does a no-training commitment matter so much?

Because training folds your data into a model's weights, which retention limits and deletion requests cannot undo afterward. For any AI vendor handling customer or confidential data, a no-training commitment is required before approval — not requested, not preferred, required — which is the sharpest line in this document.

How often are approved AI vendors re-reviewed?

At least annually, and again immediately after any material change to the vendor's terms or a security event on their side. It is the same clock the general Vendor Management Policy sets for its high-tier vendors; what differs here is what actually gets re-checked — training and retention terms, not just access.

What happens if a vendor can't meet the criteria?

It is not used with restricted data. The policy states this as a bar, not a risk to accept with a compensating control — a vendor that will not commit to no-training or cannot state its retention terms does not get customer or confidential data, regardless of how useful it is.

How is this different from the Vendor Risk Management Policy?

Vendor Risk Management sets the program every vendor goes through — a tier assigned by data sensitivity, a security review before onboarding, and a yearly check-in afterward. This document narrows in on what only matters for a model provider: whether it trains on what you send it, how long it keeps it, and which subprocessors sit behind its own API.

How does Pentest Today generate the policy?

Answer a short intake about your stack and we generate a tailored draft — not a blank template — pre-mapped to the controls your framework requires. You review, edit, and export it.

Can I edit the generated policy?

Yes. Every document is a starting draft you can edit, brand, and export. It's written to be review-ready but stays fully under your control.

Free · no account

Start your Pentest

Our agent swarm and human experts test your endpoints and deliver an audit, fast.

Start your Pentest How it works

Generate your full security policy pack.

Get the ai vendor review policy plus everything else an enterprise security review asks for — generated from your real environment.

Get started Browse all policies