Personalized AI infrastructure
Superintelligence RFP Questions on Memory, Tools and Exit Terms
Ready-to-send RFP questions on memory ownership, tool scopes, approval limits and exit terms, so procurement can compare AI vendors before any decision.

A superintelligence RFP should ask each AI vendor three things in writing: who owns the memory the system builds, which tools it can call and under what permissions, and how your data, configuration and history leave with you if the contract ends. Procurement can send these questions before any demo or commercial discussion, then score every answer against the same criteria. The questions below are grouped by memory, tools, approvals, exit and scoring so the buying team can compare vendors on evidence rather than presentation.
Key takeaways
- Ask for written answers on memory ownership before comparing features, because memory is the asset that grows with use.
- Require a tool inventory that lists the scope of each tool, the account it runs under and who may change that scope.
- Treat approval thresholds and audit logs as contract terms, not as settings a vendor can change without notice.
- Write exit terms that name the export format, the deletion method and the conditions of a migration before you sign.
- Score each answer on evidence: a document, a configuration screen or a tested procedure counts for more than a verbal assurance.
Why memory ownership belongs in the RFP
Most AI purchases are judged on what the system does today. For a business, the more important question is what the system accumulates. An agent that answers customer questions, drafts quotes or reviews operating reports builds memory: summaries of past cases, stated preferences, corrections and recorded decisions. If that memory sits inside a vendor's system in a form you cannot read, move or delete, much of the value of the work stays with the vendor even after the subscription ends.
Procurement is well placed to ask these questions because they are contract questions. Ownership, retention and portability are the same kind of questions a finance or legal team asks about any system that holds company records. The RFP is also the point where the vendor still needs the deal more than you need its current answer, so vague language is easiest to fix now.
Memory and data questions to send in writing
Memory is more than a chat history. It includes stored summaries, retrieved records, embeddings, logs and any tuned settings derived from your use. Ask about each layer separately, because vendors often answer for one and assume the others are covered.
- Who owns the memory, summaries, embeddings and logs generated from our use of the system, and does that ownership appear in the contract?
- Where is each memory store hosted, which subprocessors can read it, and can you provide that list in writing with each change notice?
- Is our memory used to train or tune any model that serves other customers, and can we opt out in writing?
- Can we export memory in a documented, machine-readable format with the schema included, without opening a support ticket?
- How is memory separated between our business units or locations, and can you demonstrate that separation on a test account?
- What is the default retention period for conversations and memory, and can we set shorter periods for specific data categories?
- How do you delete one record across memory, logs and backups, and how long does complete deletion take?
- When the agent answers from stored records, can we see which records were retrieved for a given response?
Tool scope and permission questions
A tool is any action the agent can take outside its own conversation, such as reading a CRM record, creating an invoice draft, posting to a team channel or moving a calendar entry. Scope is what a tool can touch. Permission is who authorized the connection and under which identity it runs. The Model Context Protocol is one common way to expose tools to a model, but a protocol does not decide your policy. The vendor's implementation does, which is why the RFP should ask about the implementation directly.
- Provide the full list of tools the agent can call at launch, the system each one connects to and the minimum permission each one requires.
- Does each tool run under a named service account with its own credentials, or under a shared key that also serves other users?
- Can a tool be limited to read-only access, to specific records or to specific fields, and is that limit enforced by the system or only described in instructions to the model?
- Who can add a tool or widen an existing scope, and do we receive a notice and a change record before the change takes effect?
- How are credentials stored, rotated and revoked, and can one tool be disconnected without disabling the whole agent?
- How do you defend against instructions hidden in documents, emails or web pages that try to make the agent call a tool it should not use? Ask for the test results, not only the design.
- Is every tool call logged with its input, output, time and the identity used?
The OWASP authorization guidance makes a useful point for this section: access should be checked by the system at the moment of each action, not assumed from the conversation that requested it. A vendor that answers only with a policy statement has not yet shown that check.
A tool scope that exists only in a prompt is a request, not a permission.
Approval, audit and incident questions
Approval boundaries define which actions need a named person to confirm before the agent acts. The RFP should establish where those boundaries are set, who sets them and how they change. Each question below can be answered with a configuration export or a demonstration on a test account.
- Which actions can the agent take without approval, and which require a named approver? Show the default configuration as delivered.
- Where are approval thresholds stored, and does any vendor-side update require our written agreement first?
- What record exists for each approval, including who approved, what they saw and when?
- What is the incident procedure: who is notified, how quickly, what is paused, and what must happen before the agent resumes?
- Can we pause the agent ourselves, immediately, without contacting the vendor?
- How are model or tool changes tested before they reach our environment, and what evidence do we receive?
Ask for these answers in the order listed. If a vendor cannot show the default configuration, it is difficult to judge any later promise about approvals.
Exit terms: the questions most RFPs skip
Exit terms are the least exciting part of an RFP and the part that costs the most when it is missing. Ask what you will receive, in what format, how long it takes and what happens to the data held by the vendor afterward. A vendor may be cooperative during the sales process and still take weeks to produce a usable export.
Illustrative scenario: a distribution company with two warehouses signs with a vendor whose agent keeps dispatch notes and supplier preferences in the vendor's memory store. A year and a half later, the company decides to move to another provider. The export arrives as a folder of conversation transcripts, but the structured supplier preferences are missing, and the vendor explains that rebuilding them is a separate service. Nobody had asked for a schema or a test export during the RFP. Two questions in writing would have made the gap visible before signing.
- What exactly do we receive on exit: memory, configuration, tool definitions, instructions, logs and evaluation results?
- In what formats, with what documentation, and within what number of days?
- Are export costs defined in the contract now, rather than set at the time of exit?
- What happens to our data at the vendor after exit, how is deletion confirmed in writing, and how are backups handled?
- Can both systems run in parallel for a defined period, and who carries the cost of that overlap?
- Does any part of the setup depend on a proprietary format that only the vendor can read?
For a deeper treatment of this risk, see How to Avoid AI Vendor Lock-In With Business Agents.
How to score the answers before deciding
A written answer is only useful if the team can compare it. Use the same scale for every vendor, and record who reviewed each section. The steps below keep the evaluation practical for a team of operations, legal and the business owner.
- Score each question from 0 to 2: no answer, a verbal answer, or a document or test that supports the answer.
- Ask for a live demonstration of data export, tool restriction and pausing the agent, run on a test account rather than a sales environment.
- Ask the vendor to attach the answers to the contract as exhibits, or to confirm them as written representations.
- Flag any answer that depends on a feature described as planned or under development.
- Compare total commitments across vendors, including exit and migration effort, not only the first-year terms.
- If two vendors score equally, choose the one with the clearer exit terms, because that is the position you will need if the relationship changes.
For a fuller list of what a buyer should inspect in the platform itself, see What to Look for in a Company Superintelligence Platform.
Where a managed approach fits
Be Human designs, implements and operates personalized AI infrastructure for companies, so the questions above map to work we set up at the start. For each role, the engagement defines the memory and tools the agent needs, the approval boundaries, the observable success criterion and the export procedure, which is tested before it is needed. Each agent keeps its identity, memory, channels, tools, permissions and operating context under the company's control. Managed operations add recurring permission reviews, monitoring, incident response and changes recorded through a documented process. This does not remove the need for an RFP, but it gives procurement the same written answers to check.
Take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Should we send these questions before vendor demos? Yes, send them in writing before the demos so that the answers are not shaped by a presentation script. Ask each vendor to respond to every question and to name the document or test that supports each answer. Keep the responses in the procurement file so the same wording can be checked during the contract stage.
How many vendors should receive the RFP? Send the same questions to every vendor under consideration, because comparison is only fair when the criteria are identical. Three to five serious candidates is usually enough to test the answers without making the process hard to manage. Vendors that refuse to answer memory or exit questions in writing have already given you useful information.
Who inside the company should review the answers? Operations should review tool scopes and approval boundaries, since they will live with the daily behavior of the agent. Legal or the finance lead should review ownership, retention and exit terms, and the business owner should decide how the scores are weighed. A short review meeting with a written record is enough for most teams.
Do these questions apply if we only plan a small pilot? They do, and a pilot is often the best time to ask them, because the memory and tool decisions made early tend to persist. A pilot that stores records in a form you cannot export can become expensive to leave. Asking the exit questions during the pilot keeps the option to change direction open.
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.


