AI agents in your business: useful work, clear limits.

What is an AI agent, in business terms?

In this article, an agent is software that uses AI to work through a task and can call on approved tools or systems. One might gather account information before a sales call. Another might prepare a supplier invoice for review or flag a change in a quality report. Its actual capabilities depend on the systems, access and controls put in place.

The benefit is the work it helps complete. Start with a specific role, such as preparing a daily exceptions report. Avoid a broad instruction to “manage operations”, where neither the expected result nor the limits are clear.

Why a good instruction is not enough

OWASP identifies excessive functionality, permissions and autonomy as causes of excessive agency: the risk that an AI system can take damaging actions. Its guidance includes limiting available tools and access, and requiring people to approve high-impact actions. OWASP.

Our practical recommendation is to enforce the important boundaries outside the AI’s instructions too. A system that only needs to read a customer record should not receive permission to delete it. A draft should stay a draft until the agreed approval has happened. The exact design depends on the platform and the consequences of a mistake.

Give the agent an AI Authority Map

Headland’s AI Authority Map turns a general discussion about control into a working plan. For each agent, we agree six kinds of authority with the business, configure the relevant controls and test how they behave.

Decide who is responsible

There should be a named business owner who approves the agent’s purpose and any change to its authority. An authorised administrator manages its settings. A reviewer checks output, deals with exceptions and knows when to pause it. In a small business, one person may hold more than one role.

Write down who can request a change, who approves it and who tests it before it goes live. A new supplier connection or a wider set of records can change the risk even when the agent looks the same to the user.

NIST’s voluntary AI Risk Management Framework provides a broader basis for considering risk through design, evaluation and use. The Authority Map is Headland’s practical method for agreeing a particular agent’s responsibilities and limits; it is not a NIST certification. NIST.

Make approval a real decision

A person cannot meaningfully approve a message they cannot inspect. Show the proposed action, the supporting information and anything uncertain before asking for approval. Route exceptions to someone who has the time and knowledge to resolve them.

For an accounts payable assistant, that could mean showing the invoice, the matched order and any mismatch together. Preparing that pack may be valuable even if payment approval remains entirely with the finance team. This is an illustrative workflow, not a claim that a particular tool provides it automatically.

Test what it must refuse as well as what it can do

Try a missing record, an unusual request and an instruction that conflicts with the agreed rules. Check whether the agent stops or escalates when it should. Test access with different staff roles and verify that unapproved actions are blocked, not merely discouraged.

Keep an appropriate record of actions and approvals, agree who reviews it and define how the agent can be paused. These are practical controls to maintain over time, not a guarantee that a system will never fail.

Start with useful work you can supervise

You do not have to choose between a chatbot that only talks and an agent with broad freedom. Reading, preparing and flagging can remove a substantial part of a task while keeping important decisions with people. Increase authority only when there is a clear reason and evidence that the controls work.

When Headland builds a controlled agent, we create its Authority Map with you, configure and test the controls, and train the named people who will manage it. The result should be useful work with a clear chain of responsibility.