Personalized AI infrastructure
Superintelligence Infrastructure Checklist for Regulated Companies
A checklist of storage, access, approval and audit controls regulated companies should set before any AI agent gets live authority.

A superintelligence infrastructure checklist for regulated companies lists the storage, access, approval and audit components to set up before any AI agent receives live authority over customers, money, records or external communication. Each component answers a question that a regulator, auditor or board member will eventually ask: where the agent's memory lives, who granted its permissions, which actions required a person to confirm them, and what evidence remains afterward. Setting these components in place now means that a more capable model later changes what the agent can do without changing who is accountable for it.
Key takeaways
- Set storage, access, approval and audit components before granting any live authority to an agent.
- Keep memory and source documents in company-controlled systems with named owners and retention rules, so the model can be replaced without losing records.
- Give each agent its own identity with the narrowest scopes its single role needs, and enforce those limits in the tools themselves.
- Require a named human approver for actions that touch customers, payments, contracts, regulated records or external communication, and log every approval.
- Treat a capability upgrade as a written change request that re-runs your evaluations, not as an automatic improvement.
Why the infrastructure comes before the authority
An agent with live authority is an operational actor. It reads records, calls tools, drafts messages and sometimes executes changes. Governance frameworks tend to ask the same underlying questions in different words. The NIST AI Risk Management Framework, for example, is organized around governing AI risk, mapping the context in which a system operates, measuring its behavior and managing the responses. Depending on how a system is used, the EU Artificial Intelligence Act adds documentation and human oversight obligations. You do not need to decode every clause before acting. You need to show, for each agent, where its knowledge sits, what it can touch, who approved that access, and what happened when it ran.
Capability is the variable that moves. A model that is better at reading long context, planning multi-step work or choosing tools can attempt actions that a weaker model would not. The controls below should therefore be written so they do not depend on which model is running. If the rule says that any refund above a defined amount goes to a named person, that rule keeps applying when the model behind the agent changes. The design goal is governance that survives upgrades, and that goal shapes every item in the checklist.
A checklist is also not a substitute for judgment about the role itself. Before any component is built, the company should be able to state in one paragraph what the agent is for, what success looks like and what it must never do. The components exist to enforce that statement.
Storage checklist: where memory and records live
Memory is the component regulated companies most often lose track of. An agent accumulates summaries, retrieved documents, notes and conversation history, and some of that material is personal or confidential. Retrieval-based designs keep source documents outside the model and let the agent read them at the moment of use. That separation helps governance, because the company controls the source and can change or remove it, while the model remains a replaceable reasoning component. Work through the following list before any store is connected:
- Inventory every store the agent reads from or writes to, including document folders, ticket history, vector indexes, notes files and chat logs.
- Assign each store an owner inside the company, not only a vendor contact.
- Classify the contents by sensitivity and mark which stores contain personal or regulated data.
- Keep source documents in systems the company can export, so the agent can be rebuilt on a different model or host.
- Write retention and deletion rules for each store, including how a person's data is removed when that person asks for it.
- Version knowledge sources so you can state which document version informed a given answer.
- Test a restore: export the store, load it into a clean environment and confirm the agent answers from the same sources.
The last item is the one most often skipped. A store that cannot be exported and restored is a dependency, not an asset. For a fuller treatment of backup scope and restore testing, see AI agent backup and restore: what to save and how to test it.
Access checklist: what the agent can reach, and as whom
An agent should have its own identity in every system it touches, not a shared login used by staff. That identity should hold only the permissions one documented role requires. Delegated authorization, such as the OAuth 2.0 flows used by many business platforms, lets a user grant a bounded permission that can later be revoked. That is easier to govern than a standing password stored in a configuration file. OWASP's authorization guidance makes a related point: permissions should be checked on every request by the system that owns the resource, not inferred from what the requester says it is allowed to do.
- Give each agent a distinct identity in every system it uses.
- Grant read access first, and write access only where a documented task requires it.
- Separate permissions by action. Viewing a customer record, drafting a message and sending a message are three different permissions, not one.
- Enforce authorization in the systems and tools. The model should never be the only control stopping an action.
- Record an owner, a business reason and a review date for every grant.
- Test revocation. One named person should be able to cut an agent's access quickly, and someone should have verified that the cut actually works.
Access decisions are easiest to defend when they are small. An agent that can read one project tracker and draft messages in one mailbox is far simpler to review than one with broad administrative rights that were added for convenience.
Approval checklist: which actions wait for a person
Approval gates are where live authority becomes a deliberate decision rather than a side effect. Agents that read untrusted content, such as inbound emails or uploaded documents, face a known risk called prompt injection, in which text inside that content steers the model toward actions the company never intended. OWASP's prompt injection guidance treats this as a design problem, not only a wording problem. An approval gate that sits outside the model limits what a manipulated agent can do, whatever it was persuaded to attempt. Classify every action before the agent is switched on, using this decision guide:
- Reading within granted scope: allowed without approval, and logged.
- Drafting internal material: allowed, reviewed by the owner of the task before anyone relies on it.
- Reversible change inside one system, such as updating a ticket status: allowed, with notification to the task owner.
- Any change that affects a customer, a payment, a contract or a regulated record: requires a named approver to confirm each instance before execution.
- External communication that speaks for the company: requires approval until the owner has reviewed a track record for that message type.
- Actions the company has decided the agent must never take: blocked at the tool layer, not only forbidden in the instructions.
Write these thresholds in plain language and keep them in a document the approver can read without technical help. Raising a threshold is a change to the system, and it should be recorded like one. The discussion in Human approval thresholds for stronger AI agents explains how to adjust these lines as capability changes without weakening oversight.
Audit checklist and an illustrative scenario: what must be recorded
Audit is the component that makes the other three checkable after the fact. OpenTelemetry's observability primer describes telemetry in terms of traces, metrics and logs, and that vocabulary maps well onto agent work. Each run, tool call and approval can be a traceable event with a time, an actor and an outcome. Recording everything is not the goal, because complete prompts and outputs often contain personal data that the retention rules must then cover. The goal is to record enough to reconstruct what happened and why.
- The agent identity and the model version in use at the time of each run.
- The business request or task that started the run.
- Each tool call: the system, the action, the scope used and the outcome.
- For every gated action, the approver's identity, the time of approval and exactly what was approved.
- References to the source documents and versions the agent used.
- Errors, refusals and escalations to a human.
- The retention period for each log category, aligned with the storage rules above.
Illustrative scenario: a regional services firm wants an agent to prepare first drafts of client status updates from its project tracker. The team sets the tracker as the only source, grants read access to tracker projects and draft-only access in the email tool, and requires a project lead to approve every outgoing message. Later, a newer model becomes available. Because the audit log records the model version, the tool calls and the approvals, the team can run the same evaluation set on both models and compare the results before changing anything. The approval rule does not need to be rewritten, and the decision to upgrade is recorded with its evidence.
How Be Human handles this, and where to start
A practical first sequence keeps the work small and ordered:
- Name one accountable owner for the agent, and one owner for each store and each permission.
- Inventory the stores, tools and actions the agent would touch, and mark which actions are irreversible.
- Write the approval rules before granting any write access.
- Turn on logging, then confirm that a sample run can be reconstructed from the logs alone.
- Schedule a recurring permission review, and keep the evidence from each review.
Be Human designs and operates this layer as part of a managed engagement. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. Ongoing operations can include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so that changes to models, tools or permissions pass through the same written process. The work is remote and available in English and Portuguese. For the mechanics of a review, see How to run a permission review for your AI agents, step by step.
If you are not yet sure which process should be the first one you hand to an agent, the free Be Human AI diagnostic can help you find that task.
Frequently asked questions
Do we need every component in place before the first agent goes live? Not every detail, but the storage, access, approval and audit components should exist for the first role before it receives any write permission. A read-only pilot can start with lighter controls, provided it has a named owner, logged activity and no path to change records or contact customers. The gate for live authority is the point where the agent can change something that matters.
Does a more capable model make these controls unnecessary? No. A more capable model can reach more systems and attempt more complex actions, which raises the consequence of a mistake rather than removing the need for controls. Stronger models can also be harder to inspect, so the audit trail becomes more valuable over time. The controls are what let you adopt a better model with evidence instead of hope.
Can we use the same checklist if the model runs on our own infrastructure? Yes. The categories of storage, access, approval and audit apply regardless of where the model runs, although the evidence sources differ. A self-hosted deployment still needs identities, scoped permissions, approval gates and logs. Hosting choices affect who operates those components, which is a separate decision that should be documented.
Who in the company should own these components? The business owner should own the purpose and the approval thresholds for the role, while an operations or technology lead should own the stores, identities and logs. A named person should be accountable for each permission and for each change request. A small governance group can review the arrangement on a fixed schedule, as described in Superintelligence governance: a committee charter for small companies.
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.


