Skip to content

THIRD-PARTY AI REVIEW · Checklist

AI Vendor Due Diligence

Before business data goes to an AI vendor, find out what is documented and what is only promised. Get ranked questions and a list of documents to request, with no score and no verdict.

  • Runs in your browser
  • No sign-in
  • Result before any contact step
Illustrative · Example values
17 ITEMS · 6 EVIDENCE SEEN · 11 OPEN · EXAMPLE VALUESNO TRAINSAYS YESLIMITED USESEENRETENTIONNOT OFFEREDDELETIONSEENSUBPROCSEENSUB NOTICENOT SURESEC REPORTSAYS YESPEN TESTNOT OFFEREDMODEL CHGSAYS YESPINNINGNOT SUREDPASEENAI TERMSSAYS YESINCIDENTSAYS YESTESTINGNOT OFFEREDLIMITSSEENEXPORTSEENEXITNOT SUREDOCUMENTS TO REQUEST · PRIORITY ORDERCRITICALNo training on your dataVENDOR SAYS YESCRITICALIndependent security reportVENDOR SAYS YESCRITICALIncident noticeVENDOR SAYS YESHIGHPenetration test summaryNOT OFFEREDHIGHRetention period and settingsNOT OFFERED+ 4 MEDIUM · 2 LOWNO SCORE · NO VERDICT · ONLY EVIDENCE SEEN CLOSES AN ITEM
17 ITEMS · 11 OPENSEEN 6SAYS YES 5NOT OFFERED 3NOT SURE 3REQUEST FIRSTCRITICALNo-training clauseCRITICALSecurity reportHIGHPen test summary
Open items
11 of 17 · 3 critical · 2 high · 4 medium · 2 low
Rule
only evidence seen closes an item · no score · no verdict

17 items, 6 with evidence seen. Eleven stay open, ranked 3 critical, 2 high, 4 medium, 2 low.

A sales page and a call with the vendor tell you what the vendor chooses to say. A contract clause, a report or a policy page tells you what the vendor will put in writing. This tool separates the two for one AI vendor, so you know which questions are still open before real data goes in.

Rate the data and the use, then mark 17 items in nine areas as evidence seen, vendor says yes, no or not offered, or not sure. The result is a ranked list of questions and documents to request. It gives no score and no verdict, and it never asks for the vendor's name.

Your inputs · worked out in your browser

0 of 19 answered · empty

Everything is worked out in your browser. Anvaylo does not receive the numbers or answers you enter, and nothing is saved when you leave the page. Use estimates and roles, not names, contract text, customer records or credentials.

Choose Evidence seen only when you have read the document yourself. Vendor says yes without a document, No or not offered and Not sure all stay open. This tool never asks for the vendor's name.

Your context

Two answers that decide how far the data, security and incident items rise in priority.

How sensitive is the data the vendor will receive?

Think of the most sensitive thing someone could paste or upload. Low: public or non-personal information. Moderate: internal business material or a little personal data. High: customer, health, financial or children’s records, or anything protected by law or contract. Not sure counts as High.

How critical is this use?

Low: a wrong answer or an outage is a minor nuisance. Moderate: it costs time or money. High: it affects customers, staff, safety or legal duties. Not sure counts as High.

01 · Data use and training
No training on your data

Ask the vendor: Do you commit in writing that our prompts, files, outputs and feedback will not be used to train or improve any model, whether yours or another company's?

A document that would show it: The contract clause, data processing terms or product terms that state this, with the section number.

Use of your data limited to the service

Ask the vendor: Is your use of our data limited to providing the service to us, with no sale, no sharing and no use for your own analytics or marketing?

A document that would show it: The terms or privacy documentation that list the permitted uses of customer data.

02 · Retention and deletion
Retention period and settings

Ask the vendor: How long do you keep our prompts, outputs and logs, and can we shorten the period or switch off storage?

A document that would show it: The retention schedule or settings guide, covering prompts, outputs, logs and backups.

Deletion on request and at exit

Ask the vendor: Will you delete our data on request and when the contract ends, and confirm the deletion in writing?

A document that would show it: The deletion clause and a sample deletion confirmation.

03 · Subprocessors and model providers
Subprocessor and model provider list

Ask the vendor: Do you publish a current list of subprocessors and model providers, with what each one does and where it processes data?

A document that would show it: The current subprocessor list, with the location and the role of each.

Subprocessor change notice

Ask the vendor: Do you tell us before you add or change a subprocessor or model provider, and can we object?

A document that would show it: The clause or policy that sets the notice period and the right to object.

04 · Security evidence
Independent security report

Ask the vendor: Can you share a recent independent security report or certificate, with its scope and dates?

A document that would show it: The latest independent audit report or certificate (for example a SOC 2 Type II report or an ISO/IEC 27001 certificate), with its scope and dates.

Penetration test summary

Ask the vendor: Do you have independent penetration tests run, and can you share a summary of the latest one and how you handled serious findings?

A document that would show it: A summary or letter from the latest penetration test, with its date and the status of serious findings.

05 · Model and change notices
Notice of model changes

Ask the vendor: Do you give advance notice of changes to the model, its version or its features that could change the outputs we get?

A document that would show it: The change notice policy or release notes process, with the notice period.

Version testing or pinning

Ask the vendor: Can we test, or pin, a model version before a change reaches our users?

A document that would show it: Documentation of version pinning, test environments or staged release options.

06 · Contract controls
Data processing terms

Ask the vendor: Will you sign data processing terms that cover this use, including limits on purpose, subprocessors and transfers across borders?

A document that would show it: The data processing agreement or addendum, in the version you would ask us to sign.

AI terms, liability and audit rights

Ask the vendor: Does the contract set AI-specific terms (who owns the outputs, cover for IP claims), a liability limit that fits the risk, and audit or assessment rights?

A document that would show it: The master agreement, order form and AI terms, with the liability, indemnity and audit clauses marked.

07 · Incident notification
Incident notice

Ask the vendor: Do you commit to tell us about a security incident or data breach within a stated time, and how will you tell us?

A document that would show it: The incident notification clause and a summary of your incident response process.

08 · Evaluation and testing evidence
Testing results for your use

Ask the vendor: Can you share test results for accuracy, bias and misuse that fit our use case?

A document that would show it: A model card, evaluation report or test summary, with its date and the data it covers.

Known limits and intended use

Ask the vendor: Do you document the known limits of the system, its intended uses and the uses you do not support?

A document that would show it: The documentation page or policy that lists intended uses, known limits and prohibited uses.

09 · Exit and data return
Data export

Ask the vendor: Can we export our data in a usable format, on request and at the end of the contract, within a stated time?

A document that would show it: The clause or documentation that describes export formats and timing.

Exit and transition terms

Ask the vendor: What happens if you end the service or we leave: the notice period, help during the move and continued access while we move?

A document that would show it: The termination and transition clauses, including notice periods and transition help.

Waiting for inputs

Your evidence request list

19 of 19 questions are still to answer. Answer the two context questions and every item, or choose Not sure, and your list appears here. Nothing is treated as covered until you choose Evidence seen.

What you get

  • Counts of items by status: evidence seen, vendor says yes, not offered and not sure
  • Questions for the vendor, ranked by priority, each with the document that would close it
  • A checklist of documents to request, and the evidence you already hold
  • Flags, questions to settle on your side, suggested reviewer roles and an action plan
  • Downloads as text, Markdown or CSV, with a request list you can send once you add the vendor name

If you accept analytics cookies, we record only which tool was opened, that a result was shown and which export or service link was used. Inputs, answers and results are never part of that record.

Method version 2026-10-14. Review aid from self-reported answers. It is not legal advice, a regulatory classification, a certification or an audit, and it cannot see or test any system.

How it works

Start with two context questions: how sensitive the data is that the vendor will receive, and how critical the use is. Each is Low, Moderate, High or Not sure. Not sure counts as High, because an open question about your own data is a reason to ask more questions, not fewer.

Then answer 17 items in nine areas: data use and training, retention and deletion, subprocessors and model providers, security evidence, model and change notices, contract controls, incident notification, evaluation and testing evidence, and exit and data return. Each item has four answers: Evidence seen, Vendor says yes with no evidence yet, No or not offered, and Not sure.

Only Evidence seen closes an item, and only when you have read the document yourself. A yes without a document, a no and a Not sure all stay open. The answer to an item never changes its priority, so choosing Not sure can never move an item down the list.

Every item has a base importance of low, medium or high. Ten items deal with data use, retention, subprocessors, security evidence, the data processing terms and incident notice, and they move with context. When data sensitivity or use criticality is High or Not sure, those ten rise one level, to at most Critical, and they rise once even when both answers are High. A Low or Moderate context never lowers anything.

Open items are ordered by priority. Inside one priority the order is: not offered, then not sure, then a claim without a document. The result shows the counts, the flags for critical and high items that are claimed or not offered, the questions to send, the documents to request, the points you still need to settle yourself and the reviewer roles to involve.

Formula and rules

Status, priority and order, as the tool applies them
Closed item: Evidence seen. Open item: Vendor says yes with no evidence yet, No or not offered, or Not sure.
Priority = base importance of the item (low, medium or high), listed below.
Context rule: If data sensitivity or use criticality is High or Not sure, items marked as moving with context rise one level (at most to Critical). Raising happens once, even when both are High.
A Low or Moderate context never lowers a priority. The answer to an item never changes its priority.
Order: priority first. Inside one priority: No or not offered, then Not sure, then Vendor says yes. Then the order of the list.
Result: counts, flags, questions, documents to request, open questions. No score. No verdict.
Data use and training | No training on your data | high | moves with context
Data use and training | Use of your data limited to the service | medium | moves with context
Retention and deletion | Retention period and settings | medium | moves with context
Retention and deletion | Deletion on request and at exit | medium | moves with context
Subprocessors and model providers | Subprocessor and model provider list | medium | moves with context
Subprocessors and model providers | Subprocessor change notice | low | moves with context
Security evidence | Independent security report | high | moves with context
Security evidence | Penetration test summary | medium | moves with context
Model and change notices | Notice of model changes | medium | fixed
Model and change notices | Version testing or pinning | low | fixed
Contract controls | Data processing terms | high | moves with context
Contract controls | AI terms, liability and audit rights | medium | fixed
Incident notification | Incident notice | high | moves with context
Evaluation and testing evidence | Testing results for your use | medium | fixed
Evaluation and testing evidence | Known limits and intended use | low | fixed
Exit and data return | Data export | medium | fixed
Exit and data return | Exit and transition terms | low | fixed

Assumptions

  • The base importance of each item (4 high, 9 medium, 4 low) is an illustrative judgement used to set the review order. It is not a legal or industry standard.
  • Raising ten items by one level for a High or Not sure context is also illustrative. A real review may weigh your data and your use differently.
  • The 17 items are a general starting list. Your sector, country or contract may need more questions.
  • An item is closed because you say you read the document. The tool cannot check that and cannot see the document.
  • Not sure about data sensitivity or use criticality is treated as High until you find out.

Worked example

Illustrative worked example · example valuesSample

A team plans to let an AI writing assistant read customer emails, so it sets data sensitivity to High and use criticality to Moderate, then answers all 17 items. Of the 17 items, 6 have evidence seen, 5 are vendor claims with no evidence yet, 3 are not offered and 3 are not sure.

Only the 6 items with evidence seen are closed, so 11 of the 17 stay open. Because the data is High, 10 of the 17 items rise one level. The 11 open items rank as 3 critical, 2 high, 4 medium and 2 low.

The list starts with No training on your data: the vendor says yes, the team holds nothing in writing, and the item is Critical. Next come the independent security report and the incident notice, then the penetration test summary that the vendor does not offer. The tool does not say whether this vendor is acceptable. It gives the team 11 documents to ask for before it relies on any answer.

Evidence to gather

Records that make the result easier to check and act on:

  • Contract clauses or data processing terms on training, purpose, retention, deletion and subprocessors, with section numbers
  • The current subprocessor and model provider list, with the location and role of each
  • The latest independent security report or certificate, with its scope and dates, and a penetration test summary
  • The incident notification clause and a summary of the incident response process
  • The change notice policy, version pinning or staged release options, and release notes
  • A model card, evaluation report or test summary that fits your use, and the documented limits and intended uses
  • Termination, export and transition clauses, with notice periods

What it does not do

  • The tool cannot see any document. It relies on you to mark Evidence seen only after you have read the document yourself.
  • The questions are a general starting list. They do not cover the rules of one sector, country or contract.
  • The order shows what to ask first. It does not rate the vendor, and it is not legal advice, a certification or an audit.
  • A document can be out of date or describe a different product, so check its date and scope.

Questions

Does the result say whether I can use the vendor?

No. There is no score and no verdict. The result lists what is still open: the questions to ask and the documents to request. What to accept is a decision for you and your advisers.

Why does Vendor says yes not close an item?

A promise on a call or a sales page can change and cannot be checked later. A clause, a report or a policy page can be read, dated and kept. Until you hold the document, the item stays open and is listed with the request to send it.

What counts as evidence seen?

A document you have read yourself: a contract clause, data processing terms, an independent report or certificate with its scope and date, a saved policy page or a test summary. Check its date and whether it covers the product, the plan and the use you intend.

Why does Not sure stay open and count as High for context?

Not knowing is not the same as having it covered. An item you are unsure about stays on the list and also appears as a question for your own side. If you are unsure about the data or the use, the tool treats it as High, so priorities lean cautious until you find out.

Is the vendor name, or anything I choose, stored or sent?

No. Everything is worked out in your browser. There is no box for the vendor name or for free text, and nothing you choose is saved or sent. The files are made on your device, and the request list has a blank line where you add the vendor name yourself.

How is this different from an AI readiness survey?

A readiness survey asks whether your own business is prepared to adopt AI. This checklist asks what one vendor can show you in writing before you share data with it. It looks at the vendor's documents, not at your own practices.

Can I send the request list to the vendor?

Yes, after you add the vendor name and delete the internal sections at the end of the text file. The list asks for documents and records no score. Have your legal or procurement team read it first, because it is a starting list and is not tailored to your contract.

More governance tools

All governance tools