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

Personalized AI infrastructure

Enterprise Superintelligence Platform: Requirements to Write Before Buying

Put memory, permission, audit and exit requirements in writing before accepting any vendor as an enterprise superintelligence platform. A buyer's checklist.

Be Human typographic cover with the title: Enterprise Superintelligence Platform: Requirements to Write Before Buying
Conceptual illustration created with AI · Be Human

An enterprise superintelligence platform is the operating layer that lets company AI agents hold memory, use defined tools, act within set permissions and leave a record of every action. Before accepting any vendor under that label, a buyer should have written requirements for memory, permissions, audit and exit. A platform that cannot meet those requirements on paper is a demonstration, and it becomes hard to govern as soon as it touches customers, records or payments.

Key takeaways

  • Require written answers on what the platform stores as memory, who can correct it and how deletions reach backups and logs.
  • Scope permissions per agent, per tool and per action, with human approval thresholds that you can change yourself.
  • Every agent run should produce an audit record that a non-engineer can read and that names the people responsible for reviewing it.
  • Agree on an exit list before signing, so you can leave with your memory, configurations, permission matrix and run logs.
  • Treat the first engagement as a test of these requirements, not as proof of the platform's general capability.

The label proves less than the requirements do

Vendors use the words enterprise and superintelligence with wide latitude. The label tells you about ambition, not about what happens to a customer note when an agent writes it on a Tuesday afternoon, who can read that note, or how it gets removed. Written requirements turn a sales conversation into something you can check, test and hold the vendor to.

It helps to organize the questions around a shared vocabulary. The NIST AI Risk Management Framework groups its work into govern, map, measure and manage. You can borrow those four headings without adopting the framework: who is accountable for the agent, what the agent does and where it touches people, how you check its behavior, and how you change it when it fails. Each heading should end with a requirement a vendor can accept or refuse in writing.

What the memory requirement must cover

Memory is what an agent knows between sessions: customer facts, approved wording, procedures, open cases and past decisions. The vendor should explain where that memory lives, how the agent retrieves it and whether any of it is built into a model's weights. Retrieval-based designs keep facts in stored records that people can read, correct and delete. Facts embedded in a model are much harder to inspect, and much harder to remove on request.

Write the memory requirement with the following elements:

  • Name each memory category the agent will hold, and the source each category comes from.
  • State where memory is stored, which region hosts it and which subprocessors can access it.
  • Define who can add, correct and delete a memory entry, and how a deletion reaches backups, caches and logs.
  • Keep memory from one customer, business unit or location separated unless you approve a specific sharing rule.
  • Require a way to see which memory entries an answer relied on, so a reviewer can check the source of a claim.

If personal data sits in memory, the requirement also touches data protection duties. In Brazil, the LGPD and the guidance of the ANPD shape how access and deletion requests must be handled, so ask the vendor to show the exact steps for each request type rather than describing its privacy posture in general terms.

Permissions: scope every agent, tool and action

Permissions define the boundary between what an agent can read and what it can do. The OWASP Authorization Cheat Sheet makes a principle that applies directly here: access decisions should be made on every request, enforced by the system that owns the resource, and denied by default. An agent that can send messages, edit records or issue refunds should be treated as a user with those powers, not as a helpful script with unlimited reach.

Prompt injection adds a second reason to keep permissions out of the model's judgment. An agent that reads email, web pages or uploaded documents can be steered by instructions hidden in that text. The OWASP guidance on prompt injection prevention is clear that controls must live outside the prompt. Ask the vendor to show the following:

  • Each agent has its own identity and credentials, revocable without affecting any other agent.
  • Tools are listed individually, with read and write actions separated.
  • Any action above a stated financial, legal or data-sensitivity threshold requires human approval, and the threshold is a setting you control.
  • Permissions are enforced by the tool layer, so a persuasive instruction in a document cannot widen them.
  • The complete permission matrix can be exported in a form a non-engineer can review on a regular schedule.

Audit trails a leader can read

An audit record should answer four questions: what the agent was asked, what it looked up, what it did, and who approved it. The OpenTelemetry observability primer describes traces, metrics and logs as signals that are most useful when they are correlated. For an agent, that means each run should produce a trace that links the request, the memory consulted, every tool call, the output and the version of the instructions active at that moment.

Records also create exposure. A log that stores full customer messages can become a second copy of sensitive data. Ask which fields are masked by default, who can view unmasked records, how long records are kept and how they are deleted. Then ask how incidents surface: which alerts exist, who receives them, and what response process the service agreement commits to. A record that nobody reviews is decoration, so name the reviewer and the review cadence in the requirement itself.

Exit terms: leave with your memory and records

Exit is the requirement buyers write last and need first. Define, before signing, what you receive on termination: memory entries with their source references, tool definitions, the permission matrix, evaluation cases, run logs for the agreed retention period and the configuration history. Ask for formats that open without the vendor's software. Then ask how long the vendor keeps copies after termination and how deletion is confirmed in writing.

Lock-in usually hides in details: a proprietary memory format, tools that exist only inside one console, or evaluation sets that cannot be exported. The practical tests in how to avoid AI vendor lock-in with business agents are a good companion to this list. If the vendor refuses to export something, treat the refusal as a finding and ask what it provides instead, and whether that alternative is complete enough to rebuild the role somewhere else.

A written buyer plan, and how a managed approach handles it

Use this sequence before you shortlist any platform:

  • Write down the one role the platform must serve, with a success criterion stated in observable terms.
  • List the memory categories and data sources that role needs, and mark each one that holds personal or confidential data.
  • Draft the permission matrix: each tool, each action and each approval threshold.
  • Specify the audit record fields and name the people who will review them.
  • Write the exit list and ask each vendor to mark every item as included, extra or unavailable.
  • Test the shortlisted platforms on one real task in a controlled environment, and score the results against the written criteria, not against the demonstration script.
  • Carry the approved requirements into the statement of work, section by section, so they become obligations rather than intentions.

Illustrative scenario: a regional services firm wants an agent to answer scheduling requests from clients. The written requirements say the agent may read the calendar, propose time slots and send confirmations for standard appointments, and that any message touching fees or cancellations must go to a person. During a trial, the agent sends a confirmation that quotes an old rate, because the client's earlier message contained it. The rule about financial terms existed only in the instructions, not in the tool layer. The firm does not accept the platform until the vendor demonstrates a permission that blocks that action.

Be Human designs, implements and operates personalized AI infrastructure, so these requirements can become the first deliverables of an engagement instead of a checklist left to the buyer. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations then keep those requirements true over time through monitoring, recurring evaluations, permission reviews, incident response and controlled changes. The service is remote and available in English and Portuguese.

If you are not yet sure which role should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Is enterprise superintelligence platform a recognized product category? Vendors use the phrase with wide and inconsistent meaning, so treat it as a marketing label rather than a standard. Judge any platform by the memory, permission, audit and exit terms you can test in writing. If a vendor cannot describe those terms plainly, the label does not change the assessment.

Do we need every requirement in place before a pilot starts? Memory scope, permission boundaries and audit records should exist before any agent touches live customer data or live systems. A pilot on synthetic or limited data can start earlier with a narrower set, as long as the limits are written down. Exit terms should be agreed before signing, not after the pilot ends.

Who in the company should own these requirements? A named business owner should own the role and its success criterion. IT or security should approve the permission matrix and the audit design, and finance or legal should review the data and exit terms. Ownership should stay inside the company, even when a vendor operates the platform day to day.

What if a vendor refuses to export logs or memory? Treat the refusal as a material finding rather than a negotiating detail. Ask what the vendor offers in place of an export and whether that alternative would let you rebuild the role with another provider. If it would not, pause the shortlist until the terms change.

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 →