AI Acceptable Use Policy for Company Agents: Clauses and Delegation Rules
The clauses an AI acceptable use policy needs for company agents, with what staff may delegate, what they may never delegate, and who approves changes.

An AI acceptable use policy for company agents is a short written document that states what each agent may do, what staff may ask it to do, what staff must never hand over to it, and who approves changes to its tools, memory and permissions. A usable version covers scope, data, delegation, approvals, monitoring and consequences, and every clause is tied to a named owner rather than general good intentions. The sections below list the clauses and show where the lines usually belong.
Key takeaways
- Write the policy around each agent's role, tools and permissions, not around AI in general.
- Split delegation into three groups: work staff may hand over, work they may only review after an agent drafts it, and decisions that stay with a person.
- Name one accountable owner for every agent, with authority to approve tool, memory and permission changes.
- Require logs, a documented way to pause one agent, and same-day removal of access when someone leaves.
- Review the policy on a fixed schedule and after every incident, because agents, tools and vendors change.
Start with scope: which agents and which work it covers
The policy should apply to any AI agent that acts for the company: a system that reads or writes records, sends messages, calls internal or external tools, or keeps memory about customers, suppliers or staff. It should also cover the informal tools people already use with company data, because those are often where the real exposure sits. Define each covered agent in one line: its role, the tools it can reach, the memory it keeps and the person who owns it. If you cannot name the owner, the agent is not ready to be covered by the policy at all.
Scope also needs an exit rule. A tool that has not been reviewed is not approved simply because a team uses it. The policy should say that new agents and new tools enter the approved list only through the owner, and that unlisted tools may not receive company data. For a practical way to find tools that already exist outside the approved list, see how to find shadow AI at work without punishing anyone.
The core clauses every policy needs
A policy that a manager can read in ten minutes and an agent can be checked against usually contains these clauses:
- Purpose: the approved uses of each agent, written as tasks (for example, drafting replies to internal requests), not as technologies.
- Accountable owner: the named person who approves changes and answers for the agent's results.
- Data classes: which kinds of information the agent may read, store and send, and which it may never touch.
- Memory: what the agent may remember, where that memory lives, and how a person asks for it to be found and removed.
- Tools and permissions: which systems the agent may call and which actions it may take in each one.
- Human approval: the actions that require a named person to say yes before they take effect.
- Disclosure: when staff and customers must be told that they are interacting with an agent.
- Logging and review: what is recorded, who may read the records and how often someone other than the owner checks them.
- Pause and removal: how any agent is stopped and how access is taken away.
- Instructions from content: staff and the agent treat text found in emails, documents and web pages as material to read, never as commands that override the policy.
The last clause is easy to skip and important. Agents that read outside content can be steered by instructions hidden in that content. The OWASP guidance on prompt injection prevention sets out the defenses that the policy should require the agent's builders to have in place.
What staff may delegate to an agent
Delegation is acceptable when the work is one the staff member could do under their own permissions, and when a person sees the result before it has any effect outside the company. The following list is a decision guide for most business teams:
- Drafting or summarizing material that the employee will read, edit and send or file themselves.
- Looking up internal information that the employee is already allowed to see, with the source shown.
- Preparing routine records, such as a summary of a support ticket or a first draft of a status note, for a person to confirm.
- Running a repeatable checklist, such as confirming that required fields in a request are complete.
- Sorting incoming requests into queues that the owner has defined, with a rule for anything the agent cannot classify.
If a task passes this test, the policy should still name the agent and the reviewer. Delegation without a reviewer is not delegation of work, it is delegation of judgment, and the policy should treat the two differently.
What staff must never delegate
Some delegation is prohibited whatever the convenience. The policy should say so in plain language, because people will not guess these limits on their own:
- Decisions with legal, financial or employment consequences, unless a named person makes and records the decision.
- Sending sensitive or regulated personal data to any tool the company has not approved for that data class.
- Granting access, changing permissions or adding tools to an agent without the owner's approval.
- Making commitments to customers or suppliers, such as discounts, refunds, contract terms or guarantees, without a human approving each one.
- Using an agent alone to evaluate a colleague's performance, to make hiring or dismissal decisions, or to read private staff messages.
- Asking an agent to work around a pause, a log, an approval step or a restriction it was given.
The last item needs its own sentence in the policy. Staff should understand that an instruction to an agent to bypass a control is a policy violation even when the goal is legitimate. For how to set rules about reading employee messages, see the privacy rules to set before an agent reads staff messages.
Approvals, permissions and changes
Most incidents trace back to a permission that was granted for a convenient reason and never reviewed. A controlled process keeps permissions small and visible:
- Step one: list every tool the agent can call and the exact actions it allows in each one.
- Step two: classify each action as read only, draft only or act. Any action that changes money, records outside the company, or access rights needs a named approver.
- Step three: record the owner's approval for each new permission, with the reason it is needed.
- Step four: test the permission in a sandbox or with dummy records before it touches live data.
- Step five: schedule a review that removes permissions the agent no longer uses.
Illustrative scenario: A small support team wants its agent to draft replies that mention account credits. The owner approves drafting, but not sending. Any reply that names a credit goes to a team lead first. Two months later the team asks for direct sending on routine replies. The owner runs the review, sees that the agent's drafts were edited heavily in one category, and keeps the approval step for that category while allowing sending elsewhere. The policy records the change and the reason. For high-risk actions where a second person should also sign, see when two-person approval is worth the delay.
Monitoring, incidents and consequences
A policy without monitoring is a statement of hope. The minimum operating controls are these:
- Logs record which agent acted, which tool it used, what it was asked to do, and which person approved the action where approval was required.
- Someone other than the owner reviews a sample of logs on a schedule the policy names.
- Staff know where to report an agent error or a suspected data exposure, and that channel is monitored during working hours.
- The team can pause one agent without stopping the rest of the operation, and it has written steps for doing so.
- Access for a departing employee is removed the same day, including any agent tools, memory spaces and shared prompts that person configured. A same-day checklist for this is available in the offboarding guide for agent tool access.
The consequences clause should say that misuse of agents is handled under the same disciplinary and data protection rules as other misuse of company systems. It should not invent a separate penalty regime for AI. The NIST AI Risk Management Framework is a useful reference for organizing these governance responsibilities into roles, measures and review cycles without reinventing them.
How Be Human's managed approach handles this
Be Human designs, implements and operates AI agents that keep their identity, memory, channels, tools, permissions and operating context under human control. In a first engagement, the policy clauses become the approval boundaries and permission design for one defined role, so the policy is tied to the tools the agent really has rather than to general rules. Managed operations can then include monitoring, recurring evaluations, permission review, incident response and controlled changes, which keep the policy aligned with the agent as it evolves. The service is remote and available in English and Portuguese. If you are not sure which task should go to an agent first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Do we need a separate policy for AI agents if we already have an acceptable use policy for software? Yes, in most companies the existing policy does not describe delegation, autonomy or memory. An agent can act on records and send messages without a person typing each step, which changes the approval and logging questions. Extend the existing policy with an agent section that names covered agents, owners, delegation limits and review cycles.
Who should approve the policy before it goes to staff? The approval should come from a person with authority over operations and one with authority over legal or data protection questions, with the security lead reviewing the tool and permission clauses. Small companies can combine these roles, but the policy should still record who approved it and which agents it covers. Keep the approval record with the policy so that the next review starts from a known version.
Can staff use agents with customer personal data? Only where the policy names the data class, the approved tool, the purpose and the legal basis your company relies on, and where the agent's memory and logs follow the same rules. Anything outside that list should be treated as prohibited until the owner and the relevant advisers have reviewed it. Data minimization is the simplest control, because an agent that never sees a field cannot expose it.
How often should the policy be reviewed? Review it on a fixed schedule that suits your risk, and also after any incident, any new tool, any change to an agent's permissions and any change in the law that applies to your data. Each review should check the logs sample, the permission list and the approved delegation tasks against what the agents actually do. A policy that nobody reads after the first week will not match the systems it describes.
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.


