Entitlement Mapping for AI Agents: Tools, Records and Actions
Learn how to map each AI agent role to the exact tools, records and actions it may use, and how to review that map as its capability grows.

An AI agent entitlement map is a written record, kept for each agent role, of the exact tools, data records and actions that role may use, the conditions that apply to each one, and the person who approves changes. You build it by starting from a single job, listing what that job needs, removing everything else, and attaching an owner and a review date to every line. The map earns its keep only when it is reviewed each time the agent gains a tool, a memory source or more autonomy.
Key takeaways
- Map entitlements per agent role and job, not per model, vendor or chat window.
- Write every entitlement as tool, record scope, action, condition, approver and evidence.
- Deny by default: anything not written on the map stays out of bounds until someone adds it.
- Treat memory, retrieved documents and connected inboxes as entitlements, not background detail.
- Re-review the map on every capability change and on a fixed calendar, whichever comes first.
Why a written map beats a general permission setting
Most teams first meet agent permissions as a single switch: the agent can access the CRM, or it cannot. That switch hides nearly every decision that matters. Reading a customer record, drafting a reply, updating a status field and sending an email to that customer are four different exposures, and they usually need four different answers about who may approve them and what must be logged.
A written map forces those distinctions into the open. It also gives a business owner something to read without an engineering background. The OWASP authorization guidance describes the same discipline in technical terms: check every request against an explicit policy rather than assuming a caller is allowed because it sits inside the system. The NIST AI Risk Management Framework asks organizations to govern and document AI risks across the lifecycle, and an entitlement map is one of the most practical documents that supports that work.
Start from one role and its job
Begin with a single role that has a clear output. For a first engagement, that is usually one useful role such as an invoice follow-up agent, a supplier intake agent or a booking coordinator. Avoid "general operations assistant," because a broad role accumulates permissions that nobody can justify later.
For that role, work through the following steps in order:
- Describe the outcome in one sentence and name the business owner who accepts it.
- List every system the role touches today, including the ones staff use only occasionally.
- For each system, write the smallest set of records the job needs, such as one customer segment or one ledger account rather than the full database.
- Mark each action as read, draft, change, send or spend, and keep the list of verbs short.
- Remove anything you cannot tie to the outcome sentence. If nobody can name the job behind a permission, the permission goes.
- Assign an approver for each action that changes a record or reaches an outside party, and set a review date.
This sequence runs from purpose to permission. Teams that start with a tool list tend to grant whatever the tool exposes by default, and the map then describes the software rather than the job.
Write each entitlement line in the same format
Every line on the map should answer the same six questions. A fixed format makes the map readable to auditors, to managers and to whoever maintains the agent's operating instructions.
- Tool: the exact integration or connector, with its account identifier or version.
- Record scope: the filter that limits which records are visible, such as open invoices for accounts assigned to one team.
- Action: the single verb allowed, such as create draft, update due date or send reminder.
- Condition: the limit that applies at runtime, such as an amount ceiling, a time window or a required field being present.
- Approver: the named role that must confirm the action, or none with a written reason.
- Evidence: what gets written to the log and where that log is kept.
Illustrative scenario (not a client example): an invoice follow-up agent may read open invoices for one team, draft reminder emails and create follow-up tasks. It may not send an email to a customer without a staff member approving the draft, and it may not change payment terms at all. The map records each of those three lines separately, so a manager can see at a glance that sending and changing terms belong to people, while drafting belongs to the agent.
The scenario contains no permission called email access. Splitting actions this way is what makes the map reviewable, because each line can be checked, changed or removed on its own.
Separate reading, writing and acting
Entitlements fall into three broad tiers, and each deserves a different default. Reading records carries the lowest operational risk but still creates privacy exposure, so it needs scope limits and logging. Writing or drafting changes internal state, which is reversible in most systems but still needs a named owner. Acting outward, meaning sending a message, moving money, publishing content or contacting a third party, is where mistakes become visible to customers and suppliers.
Give outward actions their own section of the map, and decide in advance which of them need a second person. Guidance on two-person approval for high-risk actions explains how to choose that set without slowing routine work. Also record whether the agent can draw on memory from earlier conversations, because remembered context can widen what the agent knows about a person even when its tools are narrow. A useful companion is the authorization guidance on permission tiers, which shows how trust levels can be stepped up one line at a time.
Model Context Protocol documentation describes how an agent connects to tools and data sources through defined interfaces. That helps the map, because each connector becomes a named line you can switch off, rather than a capability buried inside a prompt.
Review the map as capability grows
A map is a snapshot of one moment. Agents change: a new model version may handle longer tasks, a connector gets added, a memory store is expanded, or a staff member leaves and their approvals need a new owner. Each of these should trigger a review of the affected lines before the change goes live, not after an incident.
A practical review cycle has three parts. First, a change trigger list that the operations lead owns and checks before any release. Second, a scheduled review, for example monthly for outward actions and quarterly for read scopes, where the owner signs off on each line and removes any permission the job no longer uses. Third, a log review that compares what the agent actually did against what the map allows, using the evidence field on each line. OpenTelemetry's semantic conventions for generative AI give teams a shared vocabulary for recording agent operations, which makes that comparison easier to automate later.
Raising autonomy should be a deliberate map change, not a side effect of a model upgrade. Move a line from draft only to send with approval only after the evidence for the current level has been reviewed and signed. The guidance on raising autonomy safely as capability grows describes this progression in more depth.
Be Human's managed operations treat this as part of the service rather than leaving it to the client's team alone. The engagement defines one role, its memory and tools, its approval boundaries and an observable success criterion, and that definition becomes the map of record. Ongoing work covers monitoring, recurring evaluations, permission review, integration maintenance and controlled changes, with each change recorded against the affected line before it takes effect.
Where to start
Choose one role this month and write its map before adding any new tool to it. Take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How often should an agent entitlement map be reviewed? Review it whenever a tool, memory source, model version or approver changes, because those events alter what the agent can do even if the job looks the same. Between those triggers, run a scheduled review in which the owner confirms each line and removes permissions that are no longer used. Outward actions usually deserve a shorter cycle than read-only scopes.
Who should own the map in a small company? One named person in operations or the business function that the agent serves should own it, and that person should be able to approve or reject changes without needing to read code. Engineering or the vendor can maintain the connectors, but ownership of what the agent is allowed to do should stay with the business. Write the owner's name on the first page of the map so the responsibility is never ambiguous.
Does agent memory count as an entitlement? Yes, and it should be written down like any tool. Memory lets an agent recall information from earlier interactions, which can include personal data, so it needs its own scope, retention rule and deletion process. Listing memory on the map also prevents a narrow tool set from hiding a wider knowledge base.
Can one map cover several agents? Keep one map per agent role, even if several roles run on the same platform. A shared map makes it hard to see which agent holds which permission, and it invites the common mistake of granting one role what another role needed once. Link the role maps together under a single index so that a reviewer can still see the whole picture in one place.
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.


