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

Tools & integrations

AI Agent Entitlement Mapping: Which Tools Each Role Gets

How to map which tools, records and actions each AI agent role may use, step by step, so access grows only when logged evidence supports it.

Be Human typographic cover with the title: AI Agent Entitlement Mapping: Which Tools Each Role Gets
Conceptual illustration created with AI · Be Human

An AI agent entitlement map is a written register that lists, for each agent role, which tools it may call, which records it may read or change, and which actions need a named person to approve them first. Build it from the smallest set of access that completes one useful task, then widen it only when logged results show the agent behaves correctly inside its current limits. Access should grow by evidence, not by convenience.

Key takeaways

  • Define each agent role by one job, one business owner and one success test before listing any tool.
  • Map tools, records and actions as separate entries, because reading, drafting and writing carry different risks.
  • Start every new entitlement at no access, then grant read scope first and only where the role needs it.
  • Tie each expansion of access to logged evidence and to a named reviewer who signs off in writing.
  • Review the map on a schedule and whenever a vendor scope, credential or role responsibility changes.

Why access should belong to a role, not to an agent

Many teams connect an agent with one shared login that can reach everything the owner can reach. That works during a demo and becomes hard to defend in production. A role describes a job, such as answering order questions, preparing a supplier reconciliation or triaging inbound leads. Entitlements attached to a role can be read by a business owner in plain language, and they can be removed without rebuilding the whole agent.

Role-based entitlements also make incidents easier to explain. When something goes wrong, the useful questions are which role made the call, which tool it used, and whether that tool sat inside its approved set. If those answers take a long search, the map was never specific enough. The NIST AI Risk Management Framework treats this kind of accountability as part of governing AI systems across their lifecycle, and the access map is one of the most practical places to start.

Step by step: build the entitlement map

Work through one role at a time. Keep the map in a document the business owner can read, not in a configuration file that only engineers open.

  • Write the role in one sentence and name the business owner who accepts responsibility for its output.
  • List the tasks the role performs in a normal cycle, then list the tasks it must never perform, even where a tool would technically allow them.
  • For each task, name every tool, record and channel it touches, including reads that happen as side effects of another action.
  • Classify each action as read, draft or write, and note whether a write can be reversed by your own team without help from the vendor.
  • Assign an approval rule to each write: no approval for reversible internal notes, sampled review for drafts, or a named approver before execution for anything that moves money, changes a customer commitment or alters a legal record.
  • Record the credential type and where the secret is stored, so that revoking access is a known procedure rather than a search.
  • Set one success criterion for the role and the evidence required before any entitlement is widened.

Keep the first version short. Three roles with four tools each are easier to review than one role with a dozen tools, and a short map is more likely to be read again after the first month.

Read, draft and write are three different entitlements

Most access mistakes come from treating uses the CRM as a single permission. Reading a customer record, drafting a message inside that record and sending a message to the customer carry very different risks. A read can expose information to the wrong context. A draft can be wrong, but it stays inert until someone releases it. A write changes the state of a system that other people depend on.

Grant these separately. An agent that can read a folder does not need edit rights to it, and an agent that can draft an email does not need the send scope. The same tool can appear in three rows of the map, each with a different approval rule. The distinctions for email are worked through in Read, draft and send: three different AI permissions, and the scope review for Google Workspace and Microsoft 365 is covered in AI Agent Google Workspace and Microsoft 365: Grant Only What It Needs.

A decision guide for each entitlement

When a team is unsure whether a role should receive a tool, a few questions usually settle it. Use them in the review meeting and write the answer beside the row in the map.

  • If the action can change money, legal documents or a customer commitment, the default is draft only, with a named approver before execution.
  • If the action can be repeated by a retry or a duplicate trigger, the write needs an idempotency check before it runs.
  • If the record is shared outside the team, limit the role to the folder, channel or account segment it actually serves.
  • If the vendor offers a narrower scope, choose it, even when the broader scope is faster to configure.
  • If you cannot describe how the action could fail, keep the entitlement at read until you can.

A reusable format for this review is described in AI Agent Tool Permissions Matrix: A Least-Privilege Template. Using the same columns for every role makes comparisons easier when someone asks why one agent can do more than another.

An illustrative scenario: widening access by evidence

The following scenario is illustrative and does not describe a real client. A small online store wants an agent to handle first-line support for order status and return questions. The owner begins with one role, a support triage agent, and grants read access to order status and to the published help articles. The agent drafts replies into a review queue, and a person approves or edits each draft before it is sent.

After several weeks, the owner reviews three things: every draft that a reviewer changed and the reason for the change, every tool call that touched a record outside the planned set, and whether the agent asked for a human when a case fell outside its instructions. If the drafts are consistently usable and no out-of-scope calls appear, the owner grants permission to add internal notes to the ticket. Refunds stay outside the role. The agent may prepare a refund summary from the order data, but execution remains with a named person in finance, and that boundary does not change in this scenario.

The order of grants is the point. Each widening has a written reason, a reviewer and a log that can be checked later. Approval steps can run through a message to a named person, as described in Slack Approval Buttons for AI Agent Actions: Decide Before Any Write.

How Be Human manages entitlement maps

Be Human designs and operates this kind of map as part of managed AI infrastructure for companies. 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 the map current through permission review, recurring evaluations, integration maintenance and incident response, so a changed vendor scope or a shifted responsibility is reviewed rather than discovered later. The work is remote and available in English and Portuguese.

Keep the map current, then choose the first role

Review the map when a role changes its work, when a vendor changes a scope, when a credential expires, and after any incident. A reconnection after an expired token can quietly restore broader access than the role needs, so the recovery steps belong in the map as well, as covered in What should an AI agent do when a connection expires?.

Choose the first role by picking a task that is frequent, bounded and easy to check. Avoid starting with anything that touches payments or contracts. The first map should be small enough that one owner can review it in a single sitting. Before you widen anything, use the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

What is an AI agent entitlement map? It is a written record of the tools, records and actions each agent role may use, with an approval rule for every write. A business owner can review it without reading code, and it becomes the reference point for granting or removing access.

Should an agent have one login or separate permissions for each role? Separate permissions per role are safer, because each role can be reviewed, limited and revoked without affecting the others. A single shared login hides which job made a given call, which makes incidents slower to explain. Shared credentials can still be used, but only behind a role-level map that names what each path may do.

How does an agent move from read-only access to write access? It moves when logged evidence shows that its drafts or actions are accurate, inside scope and reviewed by a named person. The owner should write down the reason for each widening and keep the review record. Writes that move money or change legal or customer commitments should stay behind a named approver even after the evidence is strong.

Who should approve a new entitlement? The business owner of the role should approve it, and the person responsible for the connected system should confirm that the scope is the narrowest one the vendor offers. Both approvals should be written in the map next to the row they cover. If either approver is unavailable, the entitlement waits, because a pending grant is safer than an undocumented one.

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 →