Personalized AI infrastructure
Superintelligence Procurement: Questions to Ask Before You Sign
The questions a buyer should put in writing before signing with a company superintelligence provider: memory export, model swaps, permissions and observability.

Before you sign with any company superintelligence provider, get written answers on five points: how your memory leaves the system, what happens when the underlying model changes, which permissions an agent holds and who can widen them, what you can observe about its actions, and how the relationship ends. Vague answers on these points are where buyers most often find themselves tied to a system they cannot audit or move. This checklist gives you the questions to send, the evidence to ask for, and the order in which to ask them.
Key takeaways
- Ask for a memory export in a readable format and test it on sample data before you sign.
- Require written notice, a fixed evaluation set and a rollback path for every model swap.
- Tie each agent permission to a named owner, a stated limit and a scheduled review.
- Confirm which logs exist, how long they are kept, and who on your side can read them.
- Write exit terms that cover data return, deletion evidence and continuity of running workflows.
Why the written answer carries more weight than the demo
A demonstration shows what a system can do on a good day with carefully chosen inputs. Procurement decides who controls it on a bad day, after the team has changed, the model has been updated and the vendor's roadmap has moved. The NIST AI Risk Management Framework treats governance as work that runs through the whole life of an AI system, not as a one-time review before launch. Your contract and statement of work are where that life cycle gets assigned to people. If a point is not in writing, assume it is open to interpretation, and that the interpretation will favor whoever holds the system.
A practical rule: every answer you rely on should appear as a sentence in the proposal, the statement of work or the service schedule, with a named person responsible for it. A verbal assurance from a sales call rarely survives a change of account manager.
Memory export: what must be able to leave with you
Memory is more than chat history. In a working agent it includes instructions, reference documents, summaries, retrieved passages, process notes and the rules your team has taught it. Ask the provider to list each category and say where it is stored. Then ask which parts can leave in a form you can open without the provider's software. Source documents are the easiest to keep. Retrieval indexes and embeddings can usually be rebuilt from those documents, so you need the documents more than the index.
Questions to put in writing:
- Which memory categories exist, and which of them are included in the export?
- In what formats are they delivered, and can someone without the platform read them?
- How long does a full export take, and is there a fee or a notice period?
- Can we run an export of a test memory before we sign, and review the result together?
- Does deletion on the provider side cover backups, indexes and logs, and how is that confirmed in writing?
Model swaps: who decides when the engine changes
The model behind your agent will change. Providers update models, retire versions and introduce options that are cheaper or stronger. The question is whether you learn about a change before your staff or customers feel it, and whether the agent's behavior is checked afterward. A swap that the contract does not mention is a swap you cannot contest.
Illustrative scenario: a regional services firm uses an agent to draft quotes from its approved pricing rules. After an unannounced model update, the drafts begin to leave out a standard exclusion clause, and nobody notices for a week. A recurring evaluation, meaning a fixed set of test requests with expected outcomes, would have flagged the omission on the first run after the change.
Questions to put in writing:
- Which model versions run our agents today, and who is allowed to change them?
- How much notice do we receive before a swap, and through which channel?
- Will you run our fixed evaluation set before and after the change, and share the results with us?
- Can we pin a version for a defined period and roll back if the results drop?
- Which other parts of the agent, such as prompts, tools and retrieval, are re-tested when the model changes?
Permissions: what the agent can touch and who can widen it
Start with a complete list of every system the agent can call. For each one, the contract should state whether the agent can read or write, which data it can reach, which actions require a human approval, and what the approval threshold is. Ask whether the agent acts through its own service account or through a person's credentials. The second option makes every action look like that person's work and complicates any review later. The OWASP Authorization Cheat Sheet recommends checking access on every request and deciding it by explicit rules rather than by default, so ask for that rule set in writing.
A short decision guide for the permission section:
- If the agent can write to a system of record, require an approval step with a named approver and a limit.
- If the agent can send messages to customers, keep a review path until evaluations show stable results.
- If a permission was granted for one project, require an end date that someone must actively renew.
- If the provider can widen a permission without your written approval, do not sign.
- If access cannot be revoked from your side within a stated time, treat that gap as a blocker before launch.
A full method for this review is described in how to run a permission review for your AI agents, step by step.
Observability: what you must be able to see
Observability means you can reconstruct what the agent did, when, with which model and which tool, and which human approved what. The OpenTelemetry observability primer describes three basic signals, traces, metrics and logs. The same logic applies here: a trace that follows one request through the model and each tool call, metrics on errors and approvals, and logs of configuration changes. Ask for each signal by name.
Questions to put in writing:
- Do we receive a record of each tool call, input source, model version and approval decision?
- How long are these records kept, where are they stored, and who on our side can read them?
- Which alerts fire, who receives them, and within what window after an incident?
- Can a non-technical manager read a plain-language summary of what the agent did last week?
- If we leave, are the logs exportable in an open format?
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, its approval boundaries and an observable success criterion. Managed operations can include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so the answers above remain true after signature, not only during the sales process.
Exit terms: how the relationship ends
Most buyers spend the least time on exit, which is the moment they have the least leverage. Write the terms before you need them. Specify the notice period, the content of a complete data return, the deletion certificate, the handover of prompts, tool definitions and workflow documentation, and what happens to the agent's external accounts, such as messaging channels and email identities. Also specify how workflows keep running during the transition, because an agent that stops mid-process can cause more damage than one that never existed.
A guide to the early warning signs of lock-in is in how to avoid AI vendor lock-in with business agents.
A checklist to run before signature
- Step 1: Write the one role the agent will perform and the success measure it must meet.
- Step 2: Send the questions from the sections above as a written request with a response deadline.
- Step 3: Request a test export and a sample observability record from a demo or pilot environment.
- Step 4: Score each answer as documented, partial or missing. Any missing answer on memory, permissions or exit blocks signature.
- Step 5: Confirm that the statement of work names a responsible person for each control.
- Step 6: Schedule the first permission review and the first recurring evaluation before go-live.
For a deeper look at platform features to compare side by side, see what to look for in a company superintelligence platform.
If you are not yet sure which task should go to an agent first, take the free Be Human AI diagnostic to find the first task worth handing over.
Frequently asked questions
Can we ask for a memory export after we have already signed? You can ask, but it is much harder to negotiate after signature, when the provider holds more leverage. Put the export format, the timing and a test run into the contract before you sign, and request a sample before you decide.
Do we need to approve every model change? Not every minor update needs your sign-off, but any change that can affect output should trigger notice and a rerun of your fixed evaluation set. Agree in writing which changes count as minor and which require your approval. Keep the rollback path in the same document so it is available when a change goes wrong.
Is a permission review necessary if the agent only reads data? Yes, because read access to customer records, contracts or pricing can still expose sensitive information, and access granted for one project often outlives that project. A read-only review is shorter, but it should still happen on a fixed schedule and leave a written record. Ask the provider to show you the evidence from the last review.
What if our team is too small to read raw logs? You need a summary rather than raw traces. Ask for a regular plain-language report of actions, approvals and errors, together with an alert path for incidents. The technical reading can be handled by the provider or an operations partner, but the accountable owner should still receive the report and make the decisions.
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.


