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

Personalized AI infrastructure

Superintelligence Entitlement Mapping: Who Can Do What, Written Down

Map each AI agent's entitlements to roles and systems in one written register with owners, approval rules and revocation steps you can review.

Be Human typographic cover with the title: Superintelligence Entitlement Mapping: Who Can Do What, Written Down
Conceptual illustration created with AI · Be Human

Agent entitlement mapping is the practice of recording, for each AI agent, the roles it works under, the systems and actions it can reach, the actions that need a named person to approve them, and the exact steps that remove each grant. The result is a short written register that an owner or operations leader can read in one sitting, question line by line, and change without reading code. Every permission in that register should have a reason, an accountable owner and a revocation path.

Key takeaways

  • Give each agent one role, one accountable owner and one written purpose before any access is granted.
  • Map entitlements by system and by action, so the register describes what the agent can actually do, not what its tool is called.
  • Split rights into read, draft, write and spend, and require a named human approval for anything that touches money, customers or external commitments.
  • Store the register in a versioned place and test each revocation before you need it in a hurry.
  • Review the register on a fixed schedule and after every incident, tool change or staff move.

Why the permission list is the first document to write

Agents accumulate access quietly. A pilot connects to one shared drive, then a mailbox is added so it can file attachments, then a CRM connector is switched on so it can read contact history, then a token is created for a reporting step. Each addition looks small on its own. Months later, nobody in the company can say with confidence which systems the agent reads, which ones it can change and who approved each grant. Entitlement mapping closes that gap before it becomes a problem.

Entitlements are broader than logins. They include connector grants, service accounts, API tokens, shared folder rights, memory stores that hold copies of company records, and the channels through which the agent can send messages. The OWASP authorization guidance is useful here: access should be denied by default and checked for each request, not assumed because a trusted team member set the agent up. The register is where those defaults become visible and reviewable.

The fields every entitlement record needs

Write each grant as a record with the same fields, so records can be compared and reviewed side by side. A useful record contains:

  • Role: the job the agent performs, in one sentence, such as answering order status questions for the support queue.
  • Accountable owner: a named person who approves changes and receives alerts about the agent's use of the grant.
  • System and object: the exact system, folder, table, mailbox or channel, not a vague label such as the CRM.
  • Actions: read, draft, create, update, delete, send or spend, each listed separately.
  • Approval rule: whether the action runs alone, needs a named approver, or is not allowed at all.
  • Revocation path: where the grant lives, who can remove it, and how long removal takes to take effect.
  • Evidence: where logs of the agent's activity are kept and who reviews them.

A method for building the map in eight steps

  • Write the agent's job in one sentence and name its owner. If the sentence needs a second job, split the agent into two.
  • List every connection the agent uses today or plans to use, including accounts, tokens, folders, inboxes, databases and memory stores.
  • Break each connection into the actions the job requires, and delete any action the job does not need.
  • Mark which actions run alone, which need approval and which are forbidden, based on the owner's judgment rather than the tool's default settings.
  • Record the revocation path for every grant, including who holds the credential that makes the grant work.
  • Test one revocation in a low-risk environment and observe how quickly the agent loses access.
  • Send the register to each owner and ask for written confirmation that the record matches their understanding of the job.
  • Put the register under change control, so any new grant requires an updated record before it is switched on.

Illustrative scenario: a quote assistant for a small installer

This scenario is illustrative and does not describe a specific client. A small installation company wants an agent that drafts quotes from its price sheet, reads job notes and sends routine quotes to customers. The owner's register could read like this. The agent reads the current price sheet in one shared folder, read only. It reads contact and job history in the CRM, also read only. It creates draft quotes in a dedicated folder, and it can send a quote to a customer only when the total and any discount fall within limits the owner set in advance. A discount above that limit goes to the operations lead for approval. The agent has no access to invoicing, the bank account or the payroll folder, and its memory holds the price sheet version and job notes but no bank details.

The revocation path is written next to each grant. The CRM token lists the person who can revoke it and the expected time for removal to take effect. Before go live, the owner removes the send permission on a test copy and confirms the agent can no longer send. Later, a request to add invoicing access becomes a visible change to the register instead of a quiet extension of what the agent already does.

Separating read, draft, write and spend rights

Four tiers make the register easier to reason about. Read rights carry the least direct risk, but they can still expose sensitive records, so they should be limited to the folders or tables the job needs. Draft rights let the agent prepare something a person will check before it takes effect. Write rights change records. Spend rights move money or commit the company to something. Most avoidable failures happen when a read grant is treated as harmless, or when a write grant is added for convenience and never reviewed again.

Prompt injection makes the tiers more important. An agent that reads inbound email or web pages may be steered by text it should not obey. The OWASP prompt injection guidance points toward controls outside the model, so the register should limit what the agent can do no matter what it is told. Approval thresholds are the practical control here, and they should be raised slowly and only with evidence from the agent's real behavior, as described in approval thresholds for superintelligent agents.

Revocation must happen at the system that issued the grant. Revoke the token, remove the role binding or disable the service account. Telling the agent to stop is not revocation, because the agent may still hold a working credential. Stop rules belong in a short runbook that names the trigger, the person who acts and the systems to cut, as outlined in the one-page agent runbook.

The NIST AI Risk Management Framework organizes AI risk work into govern, map, measure and manage functions. Entitlement mapping is a concrete piece of that work: it identifies what the system can reach, who is accountable for it and how the organization can take the access back.

How managed operations keep the map current

A register written once goes stale. Set a quarterly review, and add reviews after any incident, any new connector or token, any change of tool vendor, and any staff move that changes who owns an agent. Logs should identify the agent as its own actor, separate from the human who configured it, so reviewers can see what the agent did with each grant. A review compares the register with what the systems actually grant, removes unused rights and records the evidence. The steps for that review are set out in the quarterly permission review guide.

Be Human's managed approach writes the entitlement register as part of the first engagement, ties each grant to a defined role, keeps memory and tools scoped to that role, and routes approvals to named people. Through the operating period, recurring permission reviews check the register against the live systems, and any change to a grant goes through the same request and approval path as other changes to the agent.

If you are not yet sure which task should be the first one handed to an agent, the free Be Human AI diagnostic will help you find it.

Frequently asked questions

Who should own an agent's entitlement record? The owner should be the person accountable for the business process the agent supports, not the technician who connected it. That person approves changes, receives alerts and signs off on reviews. Technical staff implement the grants, but the business owner decides whether each one is needed.

How detailed should the register be? It should be detailed enough that someone who did not build the agent can tell what it can read, change and send. Name the exact system and object, list actions separately and record the revocation path. If a single grant takes a long time to describe, it is probably too broad and should be split.

Can one register cover several agents? It can, but each agent should keep its own record with its own role and owner. A shared folder used by two agents should appear under both records, so that removing access for one agent does not silently change what the other can do.

What should trigger an update outside the quarterly review? Any new connector, token, folder or channel, a change in the underlying tool, an incident, a staff change affecting the owner, or a request to move an action from draft to write or from write to spend. Each trigger should produce a recorded change to the register and a written confirmation from the owner.

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 →