Personalized AI infrastructure
Superintelligence RFP Template: The Sections Your Written Request Needs
A practical RFP outline for company superintelligence: the sections to write so vendors answer on memory, access, approvals and success.

A written request for company superintelligence should contain six sections: the single role to be built, the memory and data sources it may read, the tools and access it needs, the approval boundaries, the success criterion and the operating model after launch. Vendors who answer each section in writing show how they will handle memory, access and approvals before any build starts. Use the outline below as a working template, and cut any question a vendor cannot answer in concrete terms.
Key takeaways
- Write the request around one role with one observable job, not around a platform.
- Ask vendors to name each memory source, tool and access path in writing.
- Put approval boundaries in the document as decisions that a named person owns.
- Require a success criterion that can be checked from records you already keep.
- Ask who monitors, reviews permissions and approves changes after launch.
Start with the role, not the platform
Vendors who receive a request that begins with a list of AI capabilities will answer with a list of AI capabilities. A stronger request begins with one role: who the agent serves, which recurring task it takes over, what inputs it receives and what output counts as finished. Name the people who will use it day to day and the person who approves changes to its work. Keep the first role small enough to describe in one paragraph. A document that tries to cover sales, support, finance and operations at once usually leaves each of them undefined.
Describe what the role does not do. Exclusions matter as much as scope in a written request, because they tell a vendor where the agent must stop. A line such as "the agent drafts replies but never sends them, and never changes a price or a delivery date" is more useful to a vendor than a paragraph about ambition.
Describe memory and data sources in writing
Memory is the section most requests leave vague. Ask the vendor to list each source the agent may read: documents, customer records, ticket history, price lists, internal policies and meeting notes. For each source, state who owns it, how current it is, and whether the agent may only read it or also write to it. Then ask how the agent retrieves information when it answers. A search index over approved documents, direct database queries and a mix of both each carry different risks for stale or conflicting records. The Google Cloud architecture material on retrieval-augmented generation describes the general pattern of grounding answers in retrieved material, which gives buyers a shared vocabulary for this question.
Ask also what the agent must not keep. Personal data, confidential client terms and unfinished negotiation notes often belong outside long-term memory, and the request should say so in plain words. Before a vendor can answer well, your source records need to be clean enough to trust. The six-point data readiness audit is a useful check to run before the request goes out.
Name every tool and access path
For each tool the agent may call, the request should state the action allowed (read, draft, create, update, send or pay), the system it touches, the account it uses and how that account is issued and revoked. Ask whether the agent acts under its own identity or under a delegated user identity, and how that delegation is granted and removed. Authorization guidance from OWASP favors deny by default, checking access on every request and keeping permissions in a place where they can be audited, not scattered through application code.
Ask the vendor to list every access path in a table of the tool, the permitted action, the scope and the reviewer. If a vendor proposes broad administrator access to make the first demonstration easier, treat that as a scope problem to fix before the build, not after.
Write approval boundaries as decisions
An approval boundary names the action, the condition that requires a person, the approver and the time the approver has to respond. A sentence such as "the agent should ask before doing anything risky" gives a vendor nothing to build. Write the boundaries in four short lists: actions the agent may take alone, actions that need one named approver, actions that need two approvers, and actions it may never take. Keep the first list short. Approval thresholds can rise later as evidence accumulates, but the starting point should be conservative and written down.
The request should also ask how the design resists instructions hidden in content the agent reads, such as an email or a web page that tries to change its behavior. The OWASP Top 10 for LLM Applications lists prompt injection among the main risks for systems like this, so vendors should explain how untrusted input is separated from permissions. For a deeper view of how thresholds change over time, read approval thresholds for superintelligent agents.
Define success so a vendor can be measured
A success criterion has to be observable from records the business keeps or will keep. "Better customer service" cannot be checked. "Each inbound question in the shared queue receives a first draft with its policy source cited, and a reviewer accepts or edits it" can be checked. Ask vendors to propose one criterion for the first role, the evidence that will show it, and the period over which it will be measured. Ask how the evaluation set will be built, usually from real past work, and how often it will be rerun. The NIST AI Risk Management Framework offers a structure for thinking about measurement and risk across the life of a system, and it is a reasonable reference when you ask how a vendor plans to measure.
Illustrative scenario: a regional equipment distributor writes a request for an agent that drafts replies to purchase-order questions. Its success criterion says that a reviewer accepts the draft without editing the price or the delivery date, measured on a fixed set of questions taken from the last review period and rerun each month. A vendor who can explain how that set is built and who reviews the results is answering the request. A vendor who only shows a demonstration is not.
Assign operations and change control before signing
A request that ends at launch leaves the hardest work unassigned. Ask each vendor to describe, in writing, how the following will run after the first period, with named roles on both sides:
- Monitoring: what is watched, which alerts exist and who receives them
- Evaluations: how often the test set is rerun and who reads the results
- Permission review: which access is reviewed, on what schedule and with what evidence
- Incident response: how a failure is reported, contained and closed
- Change control: how a request to change the agent is written, tested and approved
- Documentation: where prompts, tool settings and memory rules are recorded, and who holds copies
Ask how agent actions are recorded. OpenTelemetry's observability concepts describe traces and logs that let an operator reconstruct what happened and in what order, which is the minimum needed to investigate an incident.
Be Human designs and operates this layer as a managed service. The first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion, so these answers are settled inside the scope. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, with changes requested through a documented process described in how to request and approve changes to a managed AI agent.
If you want to find the first task worth handing to an agent before you finalize this request, take the free Be Human AI diagnostic.
Frequently asked questions
Is a superintelligence request the same as a formal RFP? A written request can be lighter than a formal procurement document, but it should cover the same core questions: scope, data, access, approvals, success and operations. Companies with procurement rules can add their own legal and commercial sections. The technical sections above should stay in either version.
Should we send the request to more than one vendor? Yes, when the answers will be compared on the same terms. Send the same role description, the same source list and the same boundary lists to each vendor so that the differences in their answers are visible. Ask each one to propose its own success criterion and evidence method, since that is where the design choices show.
Can a small company use this template without a technical team? Yes. A business owner or operations leader can answer the role, source, approval and success sections directly, because those are decisions about the business. The vendor or an outside reviewer should supply the access and architecture answers, and the owner should confirm them in plain language before signing.
What if a vendor will not answer questions about memory or access? Treat that refusal as information about the proposal. A vendor who cannot name the sources, the permissions or the revocation process has not yet designed the part of the system that carries the most risk. Ask for a written answer on those sections before any commercial discussion continues.
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.


