AI Agent Permission Matrix Template
An AI agent with a service account and no written permissions is the 2020s version of a shared admin login. This matrix gives each agent a defined scope, a spending limit and a human who is accountable.
| Agent | System | Read level (0-3) | Write level (0-3) | Transaction limit | Exposure score (0-9) | Approver role | Kill switch tested | Access review | |
|---|---|---|---|---|---|---|---|---|---|
2 | |||||||||
4 | |||||||||
6 | |||||||||
3 | |||||||||
7 |
Exposure score = read level + 2 x write level, so autonomous writes weigh most. Any agent scoring 6 or more should have a named approver, a tested kill switch and monthly access review. Give each agent its own service account in the ERP, never a shared user or a real employee login.
Review your agent permissions
Share the matrix and we will check it against your ERP security model and point out agents with more access than the use case needs.
Optional. The tool and every export work without this. No spam.
How to use the matrix
Add one row per agent and per system it touches, because an agent that reads the ERP and sends email has two different risk profiles. Fill the read and write levels from the lowest access that still lets the use case work. Record the approver role, meaning the person whose approval is needed before a write takes effect. Then match the matrix to real ERP roles and service accounts, and review it on the cadence in the last column.
The principles behind it
The matrix applies least privilege and separation of duties. An agent that proposes a purchase order change should not be the same identity that approves it. Read access should follow the ERP role of the person using the agent, so a planner cannot extract payroll through a chat window. Write access should default to draft or approval, with autonomy limited to low-value, reversible actions. Log every action with the agent identity, the prompt and the resulting transaction.
- One service account per agent, no shared credentials
- Per-transaction and per-day value limits
- Human approval for anything that posts to the ledger or changes master data
- Tested kill switch that disables the account and queued actions
Reading the exposure score
The exposure score runs from 0 to 9. Read only assistants score 3 or less, even with company-wide read, because they cannot change anything. A purchase order proposer with approval scores 6. An autonomous agent with company-wide read scores 9 and should be rare. Use the score to decide review frequency: monthly for 6 and above, quarterly for 3 to 5, and annual for the rest.
Common mistakes
Teams often grant an agent the same role as its sponsor for convenience, which gives it far more reach than needed. Another mistake is leaving the transaction limit blank: an empty limit means unlimited. Also remember indirect access. An agent that can read documents may expose files the user could not open themselves unless retrieval respects document permissions. Test with a low-privilege user before go-live.
Frequently Asked Questions
Do I need a matrix for read-only assistants?
Yes, a short one. Read-only still exposes data, and the main control is that results respect the asking user's ERP role. Record the scope, the data classes covered and the review cadence.
What is a sensible transaction limit?
Start at or below the lowest approval threshold in your procurement policy, and raise it only after reviewing logged actions. Many teams begin with zero for autonomous writes and use approval for everything.
How is this different from ERP roles?
ERP roles control human users. The matrix documents the agent's identity, its scope, who approves its actions and how it is disabled. The two should align, and the matrix is the evidence an auditor will ask for.
How often should I review it?
Monthly for agents that can write, quarterly for read access across departments, and annually for narrow read-only tools. Also review whenever an agent gains a new tool or data source.
If you want agents that inherit ERP roles, show the query behind each answer and keep writes behind approval, we can review your matrix against how your ERP security is configured.
Related Tools
AI Governance Policy Template for Manufacturers
Fill in your company details and get a ready-to-edit AI use policy covering approved tools, data handling, oversight and incidents.
AI Agents & AutomationAI Pilot Charter Template
A one-to-two page charter for an AI or agent pilot on your ERP: scope, success metrics, data access, risks, budget and stop rules.
AI Agents & AutomationAI Agent Security Review Checklist
A 30-point security review for AI agents that can call tools and write to business systems, covering identity, permissions, prompt injection, data handling, and audit.
AI Agents & AutomationAI Use Case Prioritization Worksheet for ERP
Rank candidate AI use cases on your ERP by value, feasibility, data readiness and risk, then export the shortlist to Excel.
More AI Agents & Automation tools
See all- Business Process Automation ROI Calculator
- AI Agent Use Case Prioritizer
- Document Processing Automation Savings Calculator
- AP Invoice Automation ROI Calculator
- Order Entry Automation Savings Calculator
- Quote-to-Cash Cycle Time Calculator
- Manufacturing AI Readiness Assessment
- RPA vs AI Agent Selector
- ERP AI Maturity Assessment