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

Tools & integrations

Entitlement Mapping for AI Agents: Who Can Use Which Tool

Map which users, agents and tools may combine, then match each entitlement to the approval rules your company wrote. A step-by-step method for operators.

Be Human typographic cover with the title: Entitlement Mapping for AI Agents: Who Can Use Which Tool
Conceptual illustration created with AI · Be Human

Entitlement mapping is the process of listing which users, agents and tools may interact, under which conditions, and with which approval step. For an AI agent, the map should mirror the approval rules your company already wrote, not the broad permissions a vendor grants by default. When the two match, every tool action can be traced to a person who decided it was allowed.

Key takeaways

  • An entitlement is a combination: a requesting role, an agent identity, a tool and a specific action class.
  • Start from the approval rules your company already wrote, then translate each rule into something the tool can enforce.
  • Grant the narrowest action first (read, then draft, then write), and name an approver for every write.
  • Review the map on a schedule, and again after any new tool, role change or incident.
  • Test allowed and denied requests for each row, so the logs show both outcomes.

What an entitlement means when an agent is involved

In a company, an entitlement answers a simple question: who may ask which agent to use which tool to do what. Agents make this question harder because they act on behalf of a person but usually hold their own credentials. A permission granted only to the agent lets anyone who can reach the agent reach the tool. A permission granted only to the user may not cover what the agent does in background jobs, scheduled runs or follow-up messages.

That is why the map has to cover combinations rather than single objects. A row such as "the agent may read customer records" is incomplete. The useful version says which role can trigger the read, which agent identity performs it, which system and fields are involved, and whether the result can be used in a later write. The OWASP Authorization Cheat Sheet makes the same point in general software terms: access should be checked on each request against an explicit rule, not assumed from the interface a person happened to use.

Start from the approval rules you already wrote

Most companies already have approval rules, even if they sit in a finance procedure, a sales policy or a short email from the operations lead. A rule might say that refunds above a set ceiling need a manager's confirmation, or that contracts are signed only by a director. These rules are the source of truth. The entitlement map is only a translation of them.

Gather every rule that touches a system an agent could reach. For each one, write five facts: who requests the action, who approves it, which system is involved, what exactly the action changes, and where the evidence of approval is recorded. Rules that end with "the team handles it" are not enforceable, so rewrite them until a named role can be checked. If a rule cannot be written this way, it is a signal that the process needs a decision before any agent touches it.

Build the matrix: users, agents and tools

Once the rules are explicit, the matrix has six columns. Each row is one allowed combination, and a combination that is not in the matrix is denied.

  • Requester: a named role, such as account manager or finance clerk, never "all staff".
  • Agent: the identity the agent runs under, with its own name and credentials.
  • Tool and action: one tool and one action class, such as read contacts or create a draft invoice.
  • Condition: the limit that applies, such as one folder, one channel, or a ceiling taken from the approval rule.
  • Approver: the role that must confirm before a write takes effect. Read rows may leave this empty.
  • Evidence: the place where the request, the decision and the result are logged.

Most rows will be read rows, and that is the expected shape. A company with a healthy map has many narrow reads, fewer drafts and very few direct writes. A fuller template with these columns and example values is in the AI Agent Tool Permissions Matrix.

A seven-step method for mapping entitlements

The method below works for one agent and a small set of tools, and it scales by repeating the same steps for each new role.

  • Step 1: List every approval rule that touches a system an agent could reach, and mark the rules that are not written down yet.
  • Step 2: Name the roles that request work and the people who approve it. Remove any role that cannot be named.
  • Step 3: For each tool, list its actions and sort them into three classes: read, draft and write. Keep deletions and payments in their own class, since they carry the highest risk.
  • Step 4: Create one row for each combination of requester, agent, tool and action class.
  • Step 5: Attach a condition and an approver to every write row. A write row without an approver is an unanswered decision, not an open permission.
  • Step 6: Translate each row into the mechanism the tool supports, such as a token scope, a folder permission or a channel restriction. Where a tool cannot enforce a row, the row becomes a process rule with a named person who checks it.
  • Step 7: Run an allowed request and a denied request for each row, and keep the results next to the map.

The translation in step 6 is where most mismatches appear. Access control on the tool side is usually coarser than the business rule, so the gap has to be closed either by a narrower scope or by an explicit approval step in the workflow.

[Image placeholder]

What a mismatch looks like in practice

Illustrative scenario: a support lead writes that agents may draft replies to customers, but only the support lead may send them. The integration is set up with one credential that can both create drafts and send messages. The map has a draft row and a send row with the support lead as approver, yet the credential allows sending without any approval. The log shows messages sent by the agent identity, but no approval, because the tool never asked for one.

The fix is not a better prompt. The send scope needs to be separated from the draft scope, or the send action has to pass through an approval step that the tool itself enforces. The same pattern applies to calendars, shared folders and chat channels. For a closer look at separating drafts from sends, see Read, draft and send: three different AI permissions.

How Be Human handles entitlements in managed operations

Be Human designs personalized AI infrastructure in which each agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools that role needs, its approval boundaries and an observable success criterion. The entitlement map is produced as part of that definition, so the first integration already reflects the rules the company wrote.

In managed operations, the same map becomes a working document. Permission reviews compare it with the current tools and roles, integration maintenance updates it when a vendor changes a scope or an API, and incident response uses the denied-request logs to find where a boundary failed. Service is delivered remotely, in English and Portuguese.

Keeping the map current

Entitlements drift. People change roles, tools add scopes, vendors rename actions and new channels get connected. A map that was correct at launch can become wrong within weeks, so set a fixed review point and add triggers that do not wait for the calendar: a role change, a new tool, a new channel, or any incident where an action was allowed that should not have been. Keep the denied-request logs, because they show where people expected access that the agent was never granted, which is often the first sign that a rule needs to be rewritten. For chat channels in particular, the logic is covered in Slack AI agent permissions: channel scopes and named approvers. If you are not yet sure which single task should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

How is entitlement mapping different from a permissions matrix? A permissions matrix usually lists what a tool allows, while an entitlement map lists which combinations of person, agent, tool and action your company has approved. The map starts from business rules and then checks whether the tool can enforce them. In practice, the map is the more useful document for an operations leader, because it shows gaps between policy and configuration.

Should an agent have its own identity, or use the login of the employee who asked for the work? Give the agent its own identity wherever the tool supports it, so its actions are separate in the logs and can be revoked without affecting people. Delegated access through a person's login is harder to audit and tends to carry more permissions than the task needs. Where a tool only supports delegation, record which person authorized the connection and limit the scope tightly.

How often should the map be reviewed? Review it on a fixed schedule that matches how often your roles and tools change, and always after a role change, a new integration or an incident. Keep the review short and specific: check each write row for an approver, check each scope against the current tool, and read the denied-request log for patterns. A map that is reviewed in small, regular steps stays closer to reality than one rebuilt once a year.

Do we need entitlement mapping before an agent has write access to anything? Yes, and this is the best time to do it. A read-only agent still exposes customer, financial or internal data, so its reads need rows with requesters and conditions. Writing the map first also decides what the agent is allowed to draft later, which prevents a read-only pilot from quietly becoming a write integration.

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 →