Personalized AI infrastructure
Superintelligence Procurement: Terms to Request From Any Vendor
A buyer's checklist for superintelligent agent vendors: the contract, data, access and exit terms to request before signing, and the evidence to ask for.

Superintelligence procurement is the process of buying agent infrastructure from a vendor that promises capabilities beyond today's business software. The claim cannot be verified in a sales meeting, so a buyer should judge the vendor by the terms it will put in writing: contract and ownership, data handling, access and permissions, evaluation evidence and exit.
Key takeaways
- Treat any capability claim as unverified until the vendor puts testable terms in writing.
- Request terms in five groups: contract, data, access, evaluation and exit.
- Confirm in the contract who owns prompts, workflows, memory records, logs and outputs.
- Require one named role, one memory scope and one success test before any long-term commitment.
- Write the exit plan before signing, so leaving stays possible after the first renewal.
Start with the role, not the superintelligence claim
Many vendor conversations begin with what the system will eventually do. A better starting point is a single role with a defined job, a defined set of tools and a named internal owner. When a vendor cannot describe the role in those terms, the capability claim is carrying more weight than the offer can support.
The NIST AI Risk Management Framework is a useful reference at this stage because it treats trustworthiness as a set of practices an organization governs across the life of a system, including accountability and how risks are measured. A buyer can borrow that structure for an internal procurement file even when the vendor brings its own documents. For a broader list of questions to bring into the room, see What to look for in a company superintelligence platform.
Contract terms: ownership, change control and liability
The contract decides what you own when the relationship ends, so read it before the demo ends. Ask for each clause explicitly rather than relying on general terms pages, which vendors can revise without notice to you.
- Ownership of prompts, workflows, configurations, memory records, logs and generated work product, stated as belonging to your company.
- Advance notice before the vendor changes the underlying model, a tool integration or a data handling practice that affects your agent.
- A current list of subprocessors and your right to object to a new one before it receives company data.
- Liability for data incidents and for actions an agent takes through connected tools, not only for service downtime.
- Service levels expressed in terms you can observe, such as response times for incident reports, rather than vendor uptime alone.
- A written change process so that every change to scope, tools or permissions is recorded and approved by a named person.
Change control matters more with agents than with static software. The behavior of an agent can shift when a model, prompt or connected tool changes underneath it, so the contract should say who must approve such changes and how you learn about them.
Data terms: memory, training use and deletion
Company memory is the asset that makes an agent useful, and it is also the part most likely to cause harm if handled carelessly. Ask where memory is stored, which regions host it, who inside the vendor can read it and whether any of it is used to train shared models. These answers belong in the contract and in a data description your team can file and review.
- Where each category of data is stored, and in which country.
- Whether prompts, outputs or memory feed any model used by other customers.
- How long memory, logs and backups are kept, and how deletion is confirmed to you in writing.
- Which vendor staff can read company records, and whether that access is logged.
Personal data needs its own track. Companies operating in Brazil should review obligations under the Lei Geral de Proteção de Dados with counsel, and every company should confirm how deletion requests are fulfilled across memory, logs and backups.
Agents that read emails, web pages or customer messages also read text written by people you do not know. The OWASP Top 10 for LLM Applications lists prompt injection as a central risk for this reason. Ask the vendor how it separates instructions from retrieved content and what happens when an agent is asked to act on text it should only summarize.
Access terms: permissions, tools and approval boundaries
An agent is only as safe as the access it holds. Each tool connection should use its own scoped credential, granted with the smallest set of permissions the role needs and revocable without touching the rest of the system. The OWASP Authorization Cheat Sheet describes the same discipline for applications: check permission on each action, not only at sign-in.
Approval boundaries define which actions an agent may take alone, which need a person to approve, and which it must never take. Ask the vendor to show these boundaries as a written table for your first role, with a reason for each line. A boundary that exists only in a sales slide is not a control.
Be Human designs personalized AI infrastructure so that an agent keeps its identity, memory, channels, tools, permissions and operating context under human control. In a managed engagement, the first role is defined with its memory, tools, approval boundaries and success criterion before anything is connected, and permission reviews and change requests are recorded as the work proceeds. The process for handling those requests is described in How to request and approve changes to a managed AI agent.
Evaluation terms: evidence you can inspect
Ask for evidence in the form of runs, not slides. A buyer should be able to see the test cases, the expected results, the failures and the decisions taken after each failure. Logs should record which tool an agent called, with what inputs and with what outcome, so an operations lead can reconstruct a bad action without guessing. The OpenTelemetry semantic conventions for generative AI offer a shared vocabulary for recording this kind of trace, which helps if you later change platforms.
- Test cases drawn from your own records and workflows, not only the vendor's examples.
- A written success criterion for the role, measured the same way before and after each change.
- Access to traces and logs for every action the agent takes, kept for a retention period you set.
- A recurring evaluation report that lists failures next to successes.
- Cost per completed task, so usage can be compared with the work the role actually finishes. The FinOps Foundation's guidance on unit economics is a practical reference for defining that measure.
Treat an evaluation without your own test cases as a marketing exercise. The vendor's examples show what the system can do under favorable conditions, while your cases show what your company needs it to do on an ordinary day.
Exit terms: leaving with your records and your agents
Exit is the term buyers most often skip and most often need. Before signing, agree on what a complete export contains: memory records, prompts and instructions, workflow definitions, tool configurations, logs, evaluation sets and the change history. Agree on the format, on a transition period during which both systems can run, and on written confirmation of deletion from the vendor at the end.
Avoid terms that make an export technically possible but practically useless, such as memory stored only in an opaque format with no link to the source documents. Read more about avoiding AI vendor lock-in before you accept a proprietary format.
Illustrative scenario, not a real engagement: an operations director at a distribution company sees a strong demo of an agent that answers account questions. When she asks for a full export, the vendor says memory is available only as a summary, and the contract is silent on deletion. She pauses the decision, asks for the export format in writing and adds clauses requiring source documents and a deletion confirmation. The demo did not change. The contract did.
A pre-signature checklist and the first role
Use this list as a gate before any long-term commitment. A vendor that answers each item in writing has given you something to compare. A vendor that answers with a roadmap has not yet answered the question.
- A one-page role description with its owner, tools, memory scope and success criterion.
- Written ownership of all company artifacts, including logs and outputs.
- Confirmed data location, training use, retention and deletion practices.
- Scoped credentials and a written approval boundary for each action the agent can take.
- An evaluation plan built on your own test cases, with logs retained for review.
- An agreed exit package and transition period, stated in the contract.
Take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can a vendor prove that its agents are superintelligent before we sign? No buyer can settle that question in a sales process, so it should not be the basis of the decision. Ask instead for testable terms, evaluation runs on your own cases and a written success criterion for the first role. Those items give your team evidence it can inspect after the contract starts.
Which single term matters most? Ownership and exit usually matter most, because they determine whether your company keeps its memory, prompts and records if the relationship ends. Ask for ownership of company artifacts in the contract and for a complete export that you have tested in principle. If the vendor resists either one, treat that resistance as information about the relationship.
Do we need a lawyer for this? Yes, for the contract and for any personal data handling, and the procurement team should bring counsel a finished checklist so the review stays focused. The internal team can settle the operational questions first, such as which role, which tools and which approval boundaries. Counsel then checks the wording that turns those answers into obligations.
How many roles should the first contract cover? Start with one role, one memory scope and one success test. A narrow first scope makes permissions, logs and exit easier to verify, and it gives the company a working example before it decides on additional roles.
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.


