Jump to a section+
Quick Answer
Prepare a use-case inventory linked to dated evidence of data handling, ownership, human review, account configuration, access, incidents, changes and fallback. Mark each material fact CONFIRMED, PARTIAL or UNKNOWN, with an owner, next action and due date for every gap. This prepares evidence for a review; it does not prove compliance or guarantee buyer acceptance.
Why it matters
A professional firm can be a supplier in an AI procurement questionnaire even when it does not sell AI software. A client may want to understand how AI assists its reports, research, administration or service delivery. The relevant scope is the work touching that client, not every capability advertised by a tool provider.
Client AI due diligence can arrive before a contract, during renewal or after a change in service delivery. A maintained evidence set lets the business show current facts and unresolved boundaries without reconstructing everything under deadline pressure. Reviewers have different requirements; the domains below are preparation prompts, not a universal mandatory standard.
Detailed guide
1. Define the supplier work and exact AI systems in scope
Start with the service being evaluated and identify where AI contributes to its inputs, processing or outputs. Record the product, embedded feature, business use case, users, client context and whether it is a trial or live workflow. A supplier AI questionnaire is not answered by a list containing only ChatGPT, Copilot or the name of a document platform.
Include AI embedded in meeting tools, search, CRM, document processing, coding assistants and reporting software where it is actually used in scope. Distinguish a feature that is available from one that is enabled and used. If you do not know whether a feature is active, record the uncertainty and ask the account administrator rather than assuming absence.
Use one stable use-case ID per material workflow. For example, a fictional UC-014 could describe drafting a client report from approved research, with a separate row for meeting transcription. Different data, access and review arrangements mean these should not disappear into a single generic AI entry.
2. Map allowed data separately from actual data use
For each workflow, identify public, internal, client, confidential and personal data that can enter prompts, uploads, recordings, retrieval sources or connected systems. Note generated outputs and where they are saved or sent. An AI security questionnaire may ask about indirect access as well as information typed into a prompt.
Record both the approved boundary and the observed practice. A policy prohibiting client data does not establish that no client data enters a tool. Use appropriately scoped owner confirmations, procedures and configuration evidence; do not collect unnecessary sensitive prompt histories merely to demonstrate activity.
If actual use is not established, make that gap explicit. Describe what is confirmed, which channel or data category remains uncertain, and who will investigate. Whether a particular use is lawful or permitted under a client contract needs qualified judgement informed by the actual facts.
3. Name the accountable business owner
Give each material use case an accountable business owner who understands the work and can approve boundaries, resolve gaps and stop unsuitable use. The IT administrator may own account settings without owning the quality of a valuation, legal draft, client analysis or financial report.
Record the owner role, current responsible person in your private working register, deputy and last confirmation date. Identify who approves client-facing release and who handles a buyer question. Do not publish unnecessary personal details in an evidence bundle; share the appropriate business contact through an agreed channel.
An owner name alone is partial evidence if the person does not know what the tool does or cannot obtain the relevant records. Record the authority and practical access needed to carry out the responsibility.
4. Show what human review actually checks
“A human checks it” is a starting question, not a complete answer to an AI governance questionnaire. Describe the review point, what is checked, which outputs require review, and what happens before the work affects a client or business decision. Distinguish proofreading from checking sources, calculations, assumptions, omissions and suitability for the task.
Reviewer capability matters. Can the reviewer explain the output, challenge an unsupported conclusion, reject it or override the AI-assisted recommendation? Do they have enough subject knowledge, time and access to supporting material? A click on an approval button does not establish these capabilities.
Link the procedure to evidence that review occurred: a dated review record, tracked corrections, approval trail or appropriately redacted example. Record exceptions and rejected outputs where useful. If the procedure exists but review records do not, label that boundary PARTIAL rather than inventing a history of checks.
For the fictional report workflow, evidence could show that the engagement lead checks cited sources and calculations, records corrections and approves release. If nobody can explain a material conclusion without the assistant, assign a review/override action before describing that conclusion as controlled.
5. Capture the purchased tier and enabled account settings
Record the exact vendor, product edition, purchased tier, account type and administrator responsible for settings. Keep dated evidence of controls that are actually enabled for the account and workflow. Vendor marketing describes available capabilities; it does not prove the customer purchased or configured them.
Where relevant, capture sharing, recording, connector, access and data-use settings, with enough account context to avoid confusing a personal account with a managed business account. Redact secrets, tokens and unrelated personal information before sharing screenshots.
For an AI vendor questionnaire, separate three things: a vendor statement, a contractual commitment and an observed account setting. They can support different questions. If the claim cannot be tied to the account in use, leave it UNKNOWN and request the relevant documentation or administrator confirmation.
6. Date retention, training-use and location evidence
Identify the relevant retention behaviour for prompts, uploads, recordings and outputs, and whether content may be used for training or model improvement. Record the source, effective date where given, date checked and scope. Do not collapse different products or plans into one blanket vendor answer.
Data location or residency may matter to a particular buyer or contract. Separate storage location, processing location and any stated regional controls rather than treating them as interchangeable. If the source does not answer the question for this account or feature, say so.
A current settings screenshot and a dated vendor document may together support an answer, but a mismatch needs resolution. Do not assume a privacy toggle establishes every retention or location fact. Route interpretation of contractual or regulatory implications to the relevant professional.
7. Trace connected systems, permissions and access boundaries
Map integrations and retrieval sources: mail, file stores, CRM records, meeting platforms, databases or other business systems. Record what the connection can read or write, which users or service accounts authorize it, and whether it reaches beyond the intended client or task.
List administrator, user and shared-account access, approval for new connectors and how access is removed when someone leaves. An assistant may receive sensitive information through a connector even when staff avoid entering it directly into prompts.
Attach configuration or access-review evidence where available. Separate configured permissions from proven behaviour; if you cannot establish whether existing access controls are respected in a particular integration, assign that question to the technical owner instead of making a security assurance.
8. Record known subprocessors and upstream model providers
The vendor logo is not necessarily the whole delivery chain. Where relevant, identify subprocessors and upstream model providers disclosed for the product or feature in scope. Link the current vendor list, contract or technical documentation and note what it actually covers.
A third-party AI risk questionnaire may ask whether a model provider receives customer information through another application. Do not infer an answer from the application brand or a generic list that does not identify the feature. Record known relationships and distinguish unknown routing or provider choices.
Note who monitors material provider-list updates and who evaluates their implications. Commercial, legal and security interpretation remains separate from the factual inventory. Unknown subprocessors remain UNKNOWN; they are not evidence that none exist.
9. Explain incident reporting and escalation
Describe how staff report an AI-related issue and who receives it. Examples include unintended disclosure, an incorrect client output, an unsupported recommendation or an integration acting outside the intended boundary. Record immediate containment options and the owner able to stop the workflow.
Connect this process to the existing incident pathway rather than creating a policy that nobody uses. Identify who decides on client communication, technical investigation and specialist advice. Applicable notification duties and contractual deadlines need professional review; this checklist does not determine them.
Keep an incident record or exercise record where available, including the issue, affected workflow, decisions, correction and follow-up. No recorded incident is not proof that the escalation process works. If it has not been exercised, describe that evidence limitation honestly.
10. Refresh evidence after material changes
Assign responsibility for material model, feature, provider, tier, configuration and integration changes. Record how a change is noticed, who evaluates its effect on data and review, and which evidence items or working answers need updating.
A procurement AI review can rely on a setting captured months earlier unless someone owns freshness. Keep a change log with the date, affected use-case IDs, decision and linked replacement evidence. Agree a review cadence appropriate to the workflow, alongside event-driven checks.
Do not reuse an old answer merely because the product name is unchanged. A newly connected file store or changed provider can alter the evidence boundary. Preserve prior versions internally where needed to explain what supported an earlier response.
11. Document fallback, dependency and exit
Explain what happens if the tool, provider or knowledgeable owner becomes unavailable. Identify the work that stops, the manual alternative, the deputy, access to necessary source material and any client delivery implications. A fallback needs an owner and a practicable method, not just “we would work manually”.
Record what can be exported, where reusable work is held and how access or integrations can be revoked. Identify dependencies such as proprietary histories, unavailable prompts or a sole employee who knows the workflow. Do not assume a vendor export feature captures everything needed to continue.
Where a fallback has been tried, link the dated record and limitations. Where it has not, distinguish a documented plan from a demonstrated alternative. Exit decisions involving contracts, data deletion or client commitments may require professional advice.
12. Build a reusable evidence index with visible states
Give each material source a stable evidence ID and link it to the relevant use-case ID. Record its title, controlled source link or location, owner, date checked, relevant scope and next review point. A reviewer should be able to trace a factual statement without searching an unstructured folder.
Use CONFIRMED when the stated operating fact has applicable supporting evidence; PARTIAL when support covers only part of it; UNKNOWN when it cannot currently be established. These are your working evidence states, not independent verification or a security rating.
For a fictional EV-021 account-settings record, the product tier might be CONFIRMED while connector permissions remain PARTIAL and regional processing is UNKNOWN. State each boundary rather than applying one reassuring label to the whole workflow.
Keep sensitive records controlled. An index can identify evidence without distributing every raw document to every buyer. Agree what the reviewer needs, redact appropriately and use the authorized sharing method. Reusable records should reduce reconstruction, not broaden unnecessary disclosure.
13. Turn gaps into owner, next action and due date
For each gap, record the missing fact, affected workflow, owner, specific next action and due date. “Improve governance” is not a useful action. “Account administrator to confirm connector permissions for UC-014 and attach a dated settings record by the agreed date” can be followed through.
Prioritize gaps against the real request and the service being evaluated. Separate a missing document from a missing operating control: writing a procedure does not demonstrate that a check takes place. Identify what can be answered now and what needs investigation, restriction or professional judgement.
Once the request arrives, use the separate How to Answer a Client AI Questionnaire guide for triage, response assembly and submission. This supplier checklist prepares the underlying AI due diligence evidence; it does not replace question-by-question interpretation or promise that a buyer will accept remaining gaps.
14. Keep evidence separate from professional conclusions
An AI risk assessment questionnaire may combine operational questions with contractual warranties, legal interpretations, security assertions or insurance statements. Gather the operating facts, but do not let the existence of an evidence folder substitute for qualified review of those conclusions.
RFP AI questions can ask for commitments beyond current practice. Flag them for an authorized business decision and appropriate legal, regulatory, security or insurance advice. Do not sign an absolute statement simply because an AI-generated answer sounds credible.
This checklist is evidence preparation, not certification, audit assurance, legal advice or a compliance determination. It helps the supplier explain what is supported and what remains open. Acceptance, contract terms and specialist conclusions remain separate decisions.
Practical supplier AI evidence checklist
Actionable steps employers can implement immediately.
- Identify the supplied service and every material AI product, embedded feature and use case in scope.
- Assign stable use-case IDs; record users, inputs, outputs and client context.
- Map client, confidential, personal and internal data; separate allowed from actual use.
- Name an accountable owner and deputy for each material workflow.
- Describe review checks, reviewer challenge/reject/override capability and dated review evidence.
- Confirm the purchased product tier and actual enabled account configuration.
- Date sources for retention, training/model-improvement use and relevant data location facts.
- Map integrations, connected systems, read/write permissions and access removal.
- Record known subprocessors and upstream providers; leave unsupported relationships UNKNOWN.
- Link incident reporting, escalation and containment arrangements to operating records.
- Assign change monitoring and refresh evidence after material feature/provider/settings changes.
- Document fallback, owner absence, export and exit dependencies; distinguish plans from tested alternatives.
- Use stable evidence IDs, source links, dates and CONFIRMED / PARTIAL / UNKNOWN states.
- Give every gap an owner, next action and due date; route specialist conclusions appropriately.
Already have a real client, procurement or security request?
Show WorkplaceReady the questionnaire or evidence request and deadline. Scope, fit and delivery are reviewed before work is accepted.
Show us your request and deadlineFurther reading
- How to Answer a Client AI QuestionnaireUse this guide for the response workflow once an incoming request arrives.
- Fictional AI questionnaire evidence sampleA labelled fictional illustration of evidence states and open gaps, not customer proof.
Common mistakes
Sending policy language without operating evidence
A policy states intended rules. Link it to actual workflows, settings, review records and owner confirmations; do not present an intention as established practice.
Claiming controls from an unpurchased tier
An advertised enterprise capability is not your enabled account setting. Confirm product edition, tier and configuration before citing it.
Forcing every answer into YES or NO
Follow the requested format while preserving partial evidence and explanatory limits. An unsupported YES does not make a missing control exist.
Ignoring embedded AI and subprocessors
Existing software can use AI without a new purchase. Scope the active feature and known delivery chain instead of inventorying only standalone assistants.
Having no change, exit or fallback record
A static tool list cannot explain a provider change or an outage. Record refresh responsibility, dependencies and the practical alternative.
Rebuilding evidence for every buyer
Maintain stable source IDs and dated records. Tailor the response to each request without rewriting the underlying operating facts from memory.
Frequently asked questions
- Is this only for suppliers that sell AI software?
- No. Professional and B2B firms can face supplier AI questions about AI-assisted service delivery, even when AI is not the product they sell. Scope the actual work and data relevant to the buyer.
- How is this different from the client questionnaire guide?
- This article identifies the evidence domains and records to have ready for vendor due diligence. How to Answer a Client AI Questionnaire explains the response workflow once a real request arrives: triage, assemble factual answers, handle gaps and submit.
- What if we receive a security review for AI use before our records are complete?
- Preserve the request and deadline, identify supported facts and label remaining points PARTIAL or UNKNOWN. Assign owners, actions and dates; ask for clarification where necessary. Do not invent assurances to complete a field.
- Can we rely entirely on vendor documentation?
- No. Vendor documentation can support product facts, but it does not establish your enabled settings, actual data use, access or human review. Tie the source to the product tier and workflow in use.
- Does CONFIRMED mean independently verified or compliant?
- No. It means the stated operating fact has applicable supporting evidence in your working index. Legal compliance, security assurance, insurance cover and contractual sufficiency require separate professional judgement.
- Should we send raw prompts and client files as evidence?
- Not automatically. Use the minimum relevant evidence, appropriate redaction and an authorized sharing channel. A dated review record or scoped configuration extract may address a question without exposing confidential material.
- Where can a small supplier start for free?
- Start with the AI Use Register Starter to inventory tools and use cases, owners, data, review evidence, access and fallback. It is browser-local; keep your own saved/exported copy and build the supporting evidence index alongside it.
Author
WorkplaceReady Editorial Team
WorkplaceReady publishes practical workplace readiness guidance across the risks covered on this site — heat, flooding, power outages and cybersecurity — written in plain language for the people responsible for keeping a workplace running.