BonoficiGrade your business
Menu
Guide8 min read

AI Agent Permissions: Free Template and Checklist

A free template for AI agent permissions: what each agent can read and change, which actions need a person, who can stop it, and where its actions are logged.

AI agent permissions set which systems an agent can read, what it can change, which actions need a person’s approval, who can stop it, and where every action is recorded. Set them in the systems themselves, through the agent’s own login, roles and approval steps, not in the instructions you give the model. Below is a one-page access sheet you can copy for each agent.

Why instructions are not permissions

An AI agent is software that reads, decides and acts in your systems: it matches invoices, updates orders, replies to customers. Most teams set its limits by writing them into its instructions: never pay more than €10,000, never delete a record. Those lines help, but they are requests. A model can misread them, and it reads everything it is sent, so an email or a PDF with the right wording in it can talk it out of them. This is called prompt injection, and no model is fully immune to it.

A permission works differently. If the agent’s login to the ERP cannot release a payment, no email can make it release one. That is the idea behind Microsoft chief executive Satya Nadella’s October 2026 essay Models as Insider Risks in the Super Intelligence Era, which argues that the controls over what a model can access and do “must sit outside the model.”

Six rules for AI agent access

None of these is new. They are the controls companies already use for staff who handle money and data, applied to software that acts like staff.

  • Its own identity. Every agent gets its own account in each system. Never a shared login and never the account of the person who set it up, or you cannot tell what the agent did from what they did.
  • Read-only by default. Start with read access and add each write right only when a task needs it. Least privilege, in security terms.
  • A person approves what costs money or can’t be undone. The approver sees the exact action before they confirm: amount, payee, account number, recipient list.
  • The one who proposes doesn’t approve. An agent never approves its own proposal, and is never the only check on its own work. This is separation of duties, the same rule that stops one employee from creating a supplier and paying it.
  • A record the agent can’t edit. Every action is logged in the system it touched or in a separate log, in words a person can read, and the agent has no rights to change or delete it.
  • An off switch with a name on it. A named person can pause the agent mid task, knows how, and can do it in minutes.

Rate every action by its worst case

You don’t need a risk framework. List what the agent will do, and for each action ask one question: what is the worst single thing that could happen if it went wrong once? Then sort the actions into four levels.

LevelExamplesDefault rule
ReadLook up an order, read an invoice, check stockAllowed and logged
DraftPrepare a reply, a payment proposal or a credit noteAllowed; a person sends or posts it
Change, reversibleUpdate a delivery date, tag a ticket, match an invoice lineAllowed and logged, with a sample checked every week
Money or irreversibleRelease a payment, change bank details, delete records, send to many customers at onceA named person approves every time, or the agent has no right to it at all

When the level is unclear, put a number on the worst case in euros. An action that could cost more than you would let a new employee spend without asking belongs in the last row.

AI agent access sheet template

Fill in one sheet per agent, including AI features inside tools you already use if they can change data. Replace everything in square brackets. Keep it next to the agent’s settings, not in a folder nobody opens.

AI agent access sheet

1. Agent and purpose
Name: [e.g. Invoice matcher]. What it does, in one sentence: [ ]. Live since: [date].

2. Owner
[Name, role]. Answers for what the agent does and approves any change to this sheet.

3. Model and vendor
Model: [ ]. Vendor: [ ]. Data processing agreement signed: [yes / no]. Where data is processed: [country or region]. Vendor trains on our data: [no / switched off on date].

4. Identity
Its own account in each system: [system: account name]. Never shared with a person or another agent.

5. Systems and access

  • [System, e.g. ERP]: read [what]. Change [what, or nothing].
  • [System, e.g. shared inbox]: read [what]. Change [what, or nothing].
  • [System]: read [ ]. Change [ ].

6. Actions that need approval
[Action]: approved by [name or role], who sees [amount, payee, account, recipients] before confirming. Threshold: [e.g. every payment / above €X].

7. Actions it can never take
[e.g. change bank details, delete records, create suppliers, approve its own proposals]. Enforced by: [the account has no right to it / blocked in system].

8. Worst single action
[Action], which could cost up to [€ amount] if it went wrong once. How that is limited: [ ].

9. Record
Actions are logged in [where]. Readable by [roles]. Kept for [period]. The agent cannot edit or delete the log: [confirmed by whom, date].

10. Stop
[Name] can pause the agent by [how]. Backup: [name]. Tested on: [date].

11. When something goes wrong
Tell [name / email] the same day: what the agent did, when, and what it touched.

Version: [x]. Approved by: [name]. Date: [date]. Next review: [date].

A worked example: the invoice matcher

Take an agent that does a three-way match: it compares each supplier invoice with the purchase order and the delivery note, and flags what doesn’t agree. A sheet for it might read:

  • Reads: purchase orders, delivery notes and supplier records in the ERP; the accounts payable inbox.
  • Changes: matches invoice lines and drafts payment proposals. Nothing else.
  • Needs approval: every payment proposal, approved by the finance lead, who sees supplier, amount, IBAN and the matched documents.
  • Never: releases payments, edits supplier records or bank details, creates suppliers. Its ERP account has none of those rights.
  • Worst single action: a wrong match that leads to an overpayment, capped by the approval step and by the payment run limit.
  • Record: every match and proposal in the ERP’s own audit log, which the agent can’t edit. Stop: the finance lead disables the account.

How to roll it out

  1. Find every agent with write access. Include AI features switched on inside your CRM, helpdesk, email or accounting software, not only the agents you built.
  2. Fix identity first. If an agent runs on a person’s login or a shared one, give it its own account before anything else. Everything else depends on it.
  3. Fill in a sheet per agent with its owner, in an hour, from what it does today rather than what it was meant to do.
  4. Remove what isn’t on the sheet. Any right the agent has but doesn’t need goes.
  5. Test the stop. Pause each agent once, on purpose, and time it.
  6. Review on a date: every quarter, and whenever the agent gets a new tool or system.

Where this fits

Your AI acceptable use policy covers people using AI tools. The access sheet covers AI acting in your systems on its own. Most companies need both. For the thinking behind this guide, read our essay Rent the intelligence, keep the authority. If you want to know where AI could help your company first, the free Bonofici grader reads your public pages and shows where to start.

Start with what your team already knows.

Book a 30‑minute call.We come back with the first three tools worth building, in writing.

Not ready for a call?

See where your company stands and which jobs are worth the most, from your public pages.