Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

AI Vendor RFP Template: The Sections That Force Written Proof

The sections an AI vendor RFP needs to require written proof of memory, tool permissions, observability and ownership terms, with questions for each.

Be Human typographic cover with the title: AI Vendor RFP Template: The Sections That Force Written Proof
Conceptual illustration created with AI · Be Human

An AI vendor RFP for company AI should require written answers in six areas: the role and its success criterion, memory, tool permissions, observability, ownership terms and post-launch operations. Each section should ask for a specific artifact, such as a permission table, a redacted trace or a draft exit clause, so vendors must show how their system behaves before you sign. The sections below give you the structure, the questions to put in each one and the evidence to request.

Key takeaways

  • A generic software RFP misses what makes an AI agent different: it keeps context, it acts through tools, and its behavior can change when the underlying model changes.
  • Ask every vendor for artifacts, not adjectives. A sample permission table is worth more than a sentence that says the system is secure.
  • Memory, permissions, observability and ownership each need their own section with answers you can check after signature.
  • Score the answers against one defined role and one success criterion, not against a feature list.
  • Treat the RFP as the first draft of the operating contract, because the terms you accept now are the terms you will enforce later.

Why a company AI RFP needs more than a software checklist

Traditional software RFPs ask about uptime, support hours, data hosting and pricing. Those questions still matter, but an AI agent that answers customers, updates records or drafts purchase orders has three extra properties. It retains context through memory, it takes actions through connected tools, and its outputs depend on model behavior that can shift between versions. A vendor can pass a conventional security questionnaire and still be unable to tell you which tool the agent called last Tuesday or where a stored fact came from.

The RFP is the place to ask those questions before a contract exists. The sections are ordered so that each one builds on the previous one. The role defines which memory and tools are needed, and those choices determine the permissions, the logs and the ownership terms.

Section 1: Define the role, scope and success criterion

Describe one useful role, not a general promise of company AI. Ask each vendor to respond to that single role, including its users, its inputs, its outputs and the decisions the agent may not make alone. A vendor that cannot describe the role in plain terms is not ready to design memory or permissions for it.

Include these questions in the RFP:

  • Which single role will the agent perform in the first phase, and which people depend on its output?
  • Which actions are read only, which produce drafts for approval, and which may execute without review?
  • What observable result shows the role is working, and how will that result be measured each month?
  • What must our team provide before the first working session, and who on our side owns each item?

The answers become the scope section of your statement of work, so they should be specific enough to be copied there without rewriting.

Section 2: Memory, retrieval and deletion

Memory in a business agent usually has three parts: stable identity and instructions, company knowledge retrieved at answer time, and short-term working context from the current task. Retrieval means the agent looks up relevant documents when it answers, rather than relying only on what the model learned during training. This is why the source of each fact matters. If you cannot trace an answer to a document, a person and a review date, you cannot correct it reliably.

Ask the vendor to answer these points in writing:

  • Which memory stores exist, what each one contains, and which system is the source of truth for each?
  • How is a memory item linked to its source document, its owner and the date it was last reviewed?
  • Who can add, correct or delete memory, and how does a deletion reach search indexes, embeddings and backups?
  • Can memory be scoped by team, location or customer, so that one group's records do not appear in another group's answers?
  • How is an outdated fact retired, and what does the agent do while a correction is pending?

Request a diagram of the memory flow from source to answer. A diagram that omits deletion or the review step is a sign that those steps have not been designed yet.

Section 3: Tools and permissions with written approval boundaries

Every tool the agent can call is a permission. A calendar read, a CRM update and an outbound email are three different risks, even if all three run through the same integration. Ask for a permission table that maps each tool to its scope, its action type (read, draft or execute), its approval threshold and the method used to revoke it. The OWASP authorization guidance is a useful benchmark here: access should be checked on each request, against the identity making it, rather than assumed from an earlier step.

The RFP should also ask how the agent handles hostile or accidental instructions. OWASP's Top 10 for LLM Applications lists prompt injection, where text inside an email, document or web page tries to direct the model to do something its operator did not intend. An agent that can send messages is exposed to whatever it reads, so the vendor must explain the containment.

Include these questions:

  • Provide the permission table for the proposed role, with tool, minimum scope and action type for each row.
  • Which actions need human approval, who approves them, and what happens when that approver is unavailable?
  • How are credentials issued, rotated and revoked, and can access to one tool be removed without disrupting the others?
  • How does the agent treat instructions found inside retrieved documents or incoming messages?
  • Where the platform supports delegated authorization, such as OAuth 2.0, are actions attributed to a named user or to a generic service account?

(Illustrative) A regional distributor drafts supplier replies with an agent. In its RFP, the distributor asks which tool can send email without approval. The strongest vendor answer lists that sending always waits for a named buyer's confirmation, and that the draft is stored with the source thread. The weakest answer states that the agent is 'fully secure'.

Section 4: Observability and evidence you can read

Observability means you can reconstruct what happened after the fact. OpenTelemetry's observability concepts separate three kinds of signal: traces, which show the steps taken during one request; metrics, which show trends over time; and logs, which record discrete events. For an agent, you also need evaluation records, meaning a fixed set of test cases that runs after each change and records whether the agent still behaves as expected.

Ask for samples, not descriptions:

  • Show a redacted trace for one request, including the memory items retrieved, the tool calls made and any approval step.
  • What is retained, for how long, and who can read it? Can it be exported in an open format?
  • Which evaluations run before any change goes live, and what result is required to pass?
  • How are failures classified, and which classes trigger an incident review?
  • For any past answer, can the vendor show the model version and configuration that produced it?

A vendor that cannot produce a redacted trace within the RFP process is unlikely to produce one during an incident.

Section 5: Ownership, exit and data terms

Ownership is the section vendors most often answer in general terms. Separate the assets you own from the ones licensed for the service. Source documents, instructions, memory content, configuration, logs, evaluation cases and exported data should each have a stated owner, a stated format and a stated transfer time. Ask directly whether your data is used to train or tune any model shared with other customers, and require the answer to appear in the contract rather than only in a sales conversation. The NIST AI Risk Management Framework is a helpful reference for assigning accountability across the life of an AI system, which is the same discipline this section needs.

Include these questions:

  • Which assets do we own outright, and which are only licensed to the vendor for delivering the service?
  • In what open formats can we export memory, instructions, configuration, logs and evaluation cases?
  • Is our data used to train or tune any model shared with other customers? State yes or no in the contract.
  • What is the deletion process at termination, including backups, and how is completion confirmed in writing?
  • What would a planned exit look like, and how long would it take to run the same role somewhere else?

Compare the answers with the early warning signs in vendor lock-in for AI agents before you renew, not only before you sign.

Section 6: Operations after launch, scoring and the decision

The RFP must also ask who runs the agent after launch. Ask for the vendor's approach to monitoring, recurring evaluations, integration maintenance, permission review, incident response and change requests. Each one should have an owner, a cadence and a record. For changes, ask how a request is raised, who approves it and how the previous version is restored if the change fails. A change request process is where most unplanned risk enters a live system, so it deserves a precise answer.

When you score responses, use a simple rule. Give a two when the answer includes an artifact you can inspect, a one when it describes a process without evidence, and a zero when it is a promise or is missing. Any vendor that cannot state its training and deletion terms in writing should be removed from the shortlist, whatever its other scores.

(Illustrative) Three vendors answer the same RFP for a supplier-reply agent. Vendor A supplies a feature list with no permission table. Vendor B supplies a permission table, a redacted trace and a commitment to a monthly permission review. Vendor C describes its memory as proprietary and offers no export format. Vendor B is the only one with evidence in all four core sections, which is the comparison that matters at this stage.

Be Human designs and operates this layer as a managed service. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, which correspond directly to the sections above. Managed operations cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled changes, with records kept so the evidence exists when you ask for it. The work is remote and available in English and Portuguese.

If you are not yet sure which process should become the first agent task, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Should every AI vendor receive the same RFP? Yes, and that is the main reason to use a template. When every vendor answers the same sections in the same order, you can compare the permission tables, trace samples and ownership terms directly. Adjust only the role description, which should reflect the single function you are buying.

How many vendors should receive the RFP? A short shortlist of two to four vendors is usually enough to let your team read each artifact closely. A longer list tends to produce shallow answers, because reviewers cannot inspect every sample in depth. Start with the vendors that already passed your basic security and data review.

Should the RFP answers become part of the contract? They should, at least for the points that matter most. Memory handling, permission boundaries, export formats and training commitments are easy to misremember after signature, so reference the written answers in the contract or in an attached schedule. This also gives you a clear baseline for the first permission review.

What if a vendor refuses to show a sample trace or permission table? Treat the refusal as a finding, not a negotiation point to be smoothed over. Ask whether the vendor can provide a fully synthetic example instead, and if it still refuses, the vendor cannot yet demonstrate the control you need. Keep it on the list only if it can show evidence before the decision date.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →