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

Personalized AI infrastructure

Superintelligence RFP Questions for AI Vendors, Put in Writing

Written questions to send AI vendors on memory ownership, permissions and evaluation before any contract is signed, with a checklist for buyers.

Be Human typographic cover with the title: Superintelligence RFP Questions for AI Vendors, Put in Writing
Conceptual illustration created with AI · Be Human

When a vendor pitches a superintelligent AI layer for your company, the request for proposal is where broad promises either become checkable commitments or disappear. Send every vendor the same written questions about where memory is stored, who owns it, how permissions are granted and revoked, and how output quality is tested. Keep the answers in writing and compare them line by line before any contract is signed.

Key takeaways

  • Send every vendor the same written question set so the answers can be compared line by line.
  • Memory ownership, export and deletion must be answered in contract language, not only in a demo.
  • Each permission an agent holds should map to a named owner, a limited scope and a scheduled review.
  • Evaluation needs a written method, test cases you can read and a success criterion agreed before launch.
  • A vendor who will not commit answers in writing is giving you a finding, so record it as one.

Why put superintelligence RFP questions in writing

Most AI sales conversations happen in calls and demonstrations. A demo shows what the system does on a prepared path. It does not show who holds the memory after a year of use, which accounts the agent can reach, or what changes when the underlying model is replaced. Written questions force the vendor to commit to a specific answer, and a specific answer can be held to account later.

Written answers also become useful input for the contract. The questions below map to sections that usually appear in an AI agent statement of work: scope, data, access, acceptance tests and exit. If an answer does not fit any of those sections, that gap is worth resolving before signature, not after.

Ask for every answer in the same format: the answer itself, the clause or document that supports it, and whether the capability is live today or planned. Planned features should not count as present capability when you score the response.

Memory ownership and portability questions

Memory is the part of an agent that accumulates over time. It includes the notes the agent keeps, the customer and operations context it builds, the decisions it has recorded and the documents it has indexed. Because memory becomes the most valuable and the hardest part to move, it deserves the most precise questions.

  • Where is each type of memory stored, which country hosts it, and which legal entity processes it?
  • Who owns the memory created from our work, and does the contract define it as company data?
  • Can we export all memory, configuration, prompts and interaction history in a documented, machine-readable format, and how long does an export take?
  • What happens to memory when the contract ends, and what evidence of deletion will you provide?
  • Is our memory used to train or tune any model that serves other customers?
  • Are backups of memory included, and has a restore been tested in an environment separate from the live one?
  • Which parts of the memory depend on a format that only your platform can read?

The last question is the one most buyers skip. If memory is stored in a proprietary structure, moving it later means rebuilding it from scratch. Check the vendor lock-in early warning signs before you renew, and notice that the same signs often appear in the first RFP answers.

Permission and tool access questions

An agent is only as safe as the actions it is allowed to take. Permissions are where a useful agent becomes an operational risk, because a single credential can read a calendar, send a message or change a record. Ask the vendor to describe the permission model in the same terms you would use to set up a new employee.

  • Which tools will the agent call, and what is the narrowest scope each one needs for the first task?
  • Does the agent act with its own service identity, or with delegated access tied to a named person? OAuth 2.0 is the common standard for delegated authorization, so ask exactly how it is applied.
  • Who approves a new tool or a wider scope, and where is that approval recorded?
  • Which actions require a person to approve them before they run, such as sending external messages, moving money or deleting records?
  • How does the system resist instructions hidden in emails, documents or web pages that try to override the agent's rules? OWASP publishes prompt injection guidance that is useful for this question.
  • How quickly can access be revoked for one tool, one user or the whole agent, and who holds that authority?

Ask the vendor to show the revocation step during the demonstration and to time it. A revocation that takes minutes and one that takes days are different risks. Plan a quarterly permission review from the beginning, so the answers about review frequency become operating habits rather than slides.

Evaluation and proof questions

Evaluation is where many AI purchases lose their footing. Vendors often report aggregate scores from public benchmarks, which say little about your workflow, your documents or your customers. What you need is a written method, test cases you can read, and a clear definition of failure.

  • What tasks will be tested before go-live, and who wrote the test cases?
  • Can we read the test cases, add our own, and keep them if we change vendors?
  • How is a failed answer defined, and what happens to a failure found in production?
  • How often are the tests rerun after a change to the model, prompts, tools or memory?
  • Which logs and traces are kept for each run, and can we review one decision together with the tool calls behind it? OpenTelemetry's observability concepts are a useful reference for what a trace should contain.
  • Which measure decides whether the pilot continues, and who signs that decision?

A practical test of the answers is whether a business owner could explain them to a colleague in two minutes. If the evaluation depends on a score nobody on your side can interpret, it is not evidence yet.

A scoring plan for RFP responses

Use a simple scoring sheet with one row per question. Give each answer one of three marks: committed in the contract, shown in a live demonstration, or unavailable. Only committed answers count toward the decision. Unavailable answers stay on the sheet even when the vendor promises them soon.

  • Send the same written question set to every shortlisted vendor, with one response deadline for all of them.
  • Ask each vendor to name the people who will operate the agent after launch, including the contact who responds to incidents.
  • Score each answer as committed, demonstrated or unavailable, and write the reason beside the mark.
  • Ask for a memory export from a test environment you control, then check whether the file opens without the vendor's own tools.
  • Trigger one permission revocation during the demonstration and record how long it takes.
  • Move every committed answer into the statement of work before signature, with the clause number written beside it.

Illustrative scenario (not a real client): An operations director at a regional distributor sends the same question set to two vendors. The first answers the export question with a link to a general help page. The second provides a sample export from a test environment and names the file format. The director marks the first answer as unavailable until a documented export exists, and the second vendor moves forward to the next round. The point is not that one vendor is better in general. The point is that a written answer can be checked, and a link to a help page cannot.

Keeping the answers alive after signature

Answers that stay in an email thread do not protect anyone once the agent is running. Move each committed answer into three places: the contract, the operating runbook and the review calendar. The contract records what was promised, the runbook records who does what when something breaks, and the calendar records when permissions, tests and memory will be checked again.

Be Human designs, implements and operates personalized AI infrastructure for companies, with agents that keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, delivered remotely in English or Portuguese. It is the same discipline the RFP questions ask for, carried into daily operation. If you are not yet sure which task deserves an agent first, the free Be Human AI diagnostic will help you find that first task.

Frequently asked questions

What should a buyer ask an AI vendor before signing? Ask for written answers on memory ownership and export, the permission model with revocation timing, the evaluation method and who decides whether a pilot continues. Require that every committed answer appears in the contract with a clause reference. Answers given only during a demonstration should be scored as demonstrated, not committed.

Do superintelligence RFP questions differ from a standard AI vendor RFP? The core questions are largely the same, but the weight shifts toward memory, autonomy and human oversight. A system that carries context and acts across several tools creates more lasting exposure than a single question and answer tool. Ask how the vendor limits what the agent can do on its own and how a person stops it.

How many vendors should receive the same RFP? Send the identical question set to each shortlisted vendor, usually a small group, so the comparison stays fair and manageable. A long list makes scoring inconsistent and slows the decision without improving it. Keep the same response deadline for everyone.

What if a vendor refuses to answer in writing? Treat the refusal as a finding and record it on the scoring sheet as unavailable. A vendor that cannot describe memory export or permission revocation in writing will likely struggle to provide them during an incident. Decide whether the rest of the proposal can stand without that answer, and say so in writing.

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 →