Personalized AI infrastructure
Decision Rights Matrix for AI Agents: Propose, Approve, Execute
Assign every AI agent action to propose, approve or execute, with one named human owner per row. Use this template to set the rules.

A decision rights matrix for AI agents is a short document that lists every action an agent can take and assigns each one a single mode: propose, approve or execute. Every row also names one human owner who sets the rule for that action and answers for its results. Build the matrix before an agent gets write access to any live system, and the agent's authority stays visible and under human control.
Key takeaways
- Give every agent action exactly one mode: propose, approve or execute. An action with no mode is not granted.
- Name one accountable human owner per row, with a named backup, rather than a team or a shared inbox.
- Start with read-only and reversible actions, and keep anything irreversible at propose until logged evidence supports more.
- Treat text the agent reads from outside sources as untrusted, so any action triggered by that text starts at approve or stricter.
- Review the matrix on a fixed schedule and after every incident, tool change or new capability.
Why agents need a decision rights matrix
An agent that can read a customer record, draft an email and send it has three different levels of authority, even though it looks like one assistant. Most small companies discover this only after an agent does something nobody intended. The access was granted in a tool setup screen, and the decision about what the agent may do was never written down. The matrix makes that decision explicit before the first run.
Governance frameworks point in the same direction. The NIST AI Risk Management Framework treats roles, responsibilities and accountability as part of managing AI risk, not as paperwork added afterward. A matrix is the working version of that idea for a single agent: it says what the agent may do alone, what it may only prepare, and who is accountable when the answer changes. For the wider question of who sits on the decision side, see Superintelligence Governance: A Committee Charter for Small Companies.
The three modes: propose, approve and execute
These three words need precise meanings, because teams often use the word approval to mean very different things.
- Propose: the agent produces a recommendation, draft or plan. Nothing happens in the business until a person decides and carries out the step, or gives a separate instruction.
- Approve: the agent prepares a complete action with every detail filled in, such as the recipient, the reason and the record affected. A named person approves that specific action, and only then does the tool run it.
- Execute: the agent runs the action directly, within the limits written in the matrix. Execute is not the same as unsupervised. The agent still logs each action, and the owner still reviews a sample of that log.
The difference between approve and execute is the per-action human step, not the capability of the agent. A highly capable agent can still belong in propose for an action that cannot be undone. Mode should follow the consequence of a mistake, not the confidence of the team that built the agent.
How to build the matrix in seven steps
Work through these steps with the people who run the process, not only with the person configuring the tool.
- Step 1: List every system the agent can reach, including read-only access to mailboxes, shared drives, calendars and databases.
- Step 2: Break each system into concrete actions, such as read a record, create a draft, send a message, change a status, issue a credit or delete a file.
- Step 3: For each action, note what happens if the agent is wrong, whether the result can be reversed, and who else is affected.
- Step 4: Assign a starting mode using the default rules in the fifth section. Be conservative on the first pass.
- Step 5: Name one human owner and one backup for each row, and write down the evidence the owner will check.
- Step 6: Write the stop rule for each row, meaning the condition under which the agent must halt and wait for a person.
- Step 7: Test the matrix in a sandbox or with copied data before live access, and confirm the agent cannot perform any action the matrix does not list.
The OWASP Authorization Cheat Sheet makes a general point that applies here: permissions should be checked for each request, not inferred from an earlier step or from the fact that the agent is already running. The matrix is the reference those checks are made against, so it has to be written before the checks can mean anything.
What the rows look like in practice
The following is an illustrative scenario, not a description of a real client. A small services company uses one agent to handle routine customer follow-ups. Its matrix, drafted with the support lead and the operations manager, looks like this:
- Read the ticket history for a customer | Mode: execute | Owner: support lead | Limit: read only, no export
- Draft a reply to a routine question | Mode: execute | Owner: support lead | Limit: saved as a draft, never sent
- Send a reply to a customer | Mode: approve | Owner: support lead | Limit: agent prepares the full message, a named person sends it
- Issue a service credit within the policy ceiling | Mode: approve | Owner: finance controller | Limit: ceiling written in the policy, each credit approved individually
- Change a customer's contact details | Mode: propose | Owner: operations manager | Limit: a person edits the record
- Close a complaint or offer a refund above policy | Mode: propose | Owner: general manager | Limit: the decision stays with a person
- Delete or overwrite any customer record | Mode: none | Owner: data owner | Limit: not granted to the agent
Notice what the rows do not describe. They do not explain how the agent reasons. They describe what the agent may touch, who answers for it, and what must be true before an action runs. When a row is hard to fill in, the action is usually too broad and should be split into smaller actions with different modes.
Who owns each row and when a mode changes
An owner sets the rule for that action, reviews a sample of its log and answers for any incident involving it. The owner should understand the business effect of the action, which is why it is usually the manager of the process rather than the person who configured the agent. If an outside partner built the integration, the owner still sits inside your company.
Use these default rules on the first pass:
- Irreversible actions, actions that move money or commit the company legally, and actions that expose personal data beyond the task start at propose.
- Reversible internal actions within a written policy can start at approve.
- Read-only actions that expose no more personal data than the task requires can start at execute.
- Any action triggered by text the agent read from an outside source, such as an inbound email or a web page, starts at approve at most. The OWASP LLM Prompt Injection Prevention Cheat Sheet describes how instructions hidden in content can steer an agent, and a human checkpoint is the simplest control against that.
Move an action from propose to approve, or from approve to execute, only after a review period shows that the logged results match what the owner expected, and only through a written change. Move it back down immediately after an incident. The steps for requesting and recording those changes are in How to Request and Approve Changes to a Managed AI Agent, and the standing check on who still holds which access belongs in a quarterly AI agent permission review.
How Be Human keeps the matrix current
In a first engagement, Be Human defines one useful role for the agent, the memory and tools that role needs, the approval boundaries and an observable success criterion. The matrix is drafted with the owners your company names, so the rules reflect the people who will answer for them. In managed operations, Be Human monitors agent activity against the matrix, runs recurring evaluations, maintains integrations, reviews permissions, responds to incidents and implements controlled changes once the owner has approved them. The matrix is therefore not a one-time document. It is the reference that the operating work is measured against.
Review the matrix on a fixed schedule, and also after any new tool, data source, capability or incident. Each review should confirm three things: that every row still has an active owner and backup, that logged actions match their stated modes, and that no action has widened without a recorded change. Remove any access that no current row justifies.
If you are not yet sure which action should be the first row in your matrix, the free Be Human AI diagnostic will help you find the first task worth handing to an agent.
Frequently asked questions
How many rows should a first matrix have? It should cover only the actions that the first agent role needs, which is usually a short list rather than every system in the company. Start with the tasks in the single role you chose, and add rows only when that role requires them. Anything not listed is not granted.
Does every action need a human click? No. Only actions in approve need a per-action human step, while read-only or reversible internal actions can run at execute within their limits. The purpose is to place people where a mistake would be costly or hard to undo, not to slow the agent down everywhere.
Who should own the rows if an outside partner built the agent? The partner can configure and maintain the integration, but each row still needs an owner inside your company who understands the business effect of the action. Ownership should stay with a named person even if the vendor changes, and every owner should have a named backup.
How often should the matrix be reviewed? Review it on a fixed schedule, such as every quarter, and also after any new tool, data source, incident or change in what the agent can do. Each review should confirm that owners are still active and that logged actions match their modes. Anything without a current justification should be removed.
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.


