When an AI agent's job changes, who updates its permissions?

In a marketing system I've built, an AI agent can organise campaigns, check work against the brief and prepare material for review. It cannot publish a post, send an email, spend money or change the strategy.

That separation is deliberate. The agent has a useful job, but it does not have the authority of the person who owns the campaign.

The first permission check is only the beginning

When we designed the agent, we defined the work it could support and what it was allowed to do. The result is a working role with tests and approval limits.

But useful work rarely stays where it started. Later, the agent might be connected to a calendar, customer database or publishing tool. It might begin coordinating with another agent. A task that once ended with a draft could end with an action outside the business.

That changes the risk. Reading a customer list is different from editing it. Preparing an email is different from sending it. Summarising results is different from deciding where the budget goes.

A policy signed when the agent was launched will not catch those differences on its own.

Human review only works when its job is clear

“Keep a human in the loop” sounds reassuring until you ask what the human is checking.

Who approves the audience and timing? Who checks the current information? At what point does the agent stop? Can someone see what it attempted and correct or reverse a mistake?

In our published marketing-agent case study, the agent can coordinate defined work while publishing, sending, spending and strategy remain human decisions. That boundary is part of the system, rather than a sentence in an AI policy.

It also needs to be proportionate. Filing a draft does not need the same approval as emailing customers. If every minor action returns to a person, we recreate the admin we were trying to reduce.

Governance should learn from the work

An AI governance conversation raised a question I keep coming back to: how do we make governance adaptive?

Policies, risk registers and approved tools provide an important foundation. But AI keeps evolving, and so does how people use it.

Take a team using AI to respond to customer enquiries. Approving the tool is only the beginning. Which replies were useful? Which needed human judgement? Did the team discover a question the AI should not answer, or information it was missing?

Those observations should lead to a decision: change the guidance, adjust the permissions, improve the information, or keep a person responsible for a particular response. Then review what happens next.

That feedback should not sit outside governance. It should inform it. This is where organisational intelligence becomes practical: retaining what the team learns so the way it uses and governs AI can improve together.

The wider guidance is getting more practical

Australia's Digital Transformation Agency published technical guidance for government AI agents on 4 June 2026. It calls for oversight across an agent's working life, including monitoring, intervention and the ability to pause or stop a system.

The National AI Centre's May 2026 guidance for businesses makes a related point: the same AI tool can carry different risks depending on its use. It recommends clear owners, testing before deployment and monitoring after launch.

In research published on 10 August 2026, the Australian AI Safety Institute warned that several individually safe agents do not necessarily form a safe system when they start interacting. Connecting more agents or outside services changes what needs to be tested.

Government guidance is not a legal checklist that every small business must follow. But the practical question is familiar: what can this system do, who authorised it, and who can change that decision?

An agent should not give itself a promotion

This question is shaping the practical AI learning I am developing through The Remix. The planned everyday-automation program asks people to take one real activity, test an agent-style process and decide where review belongs before handing over more work.

I want to see what the agent actually does, what happens when the information is wrong and whether someone can stop or correct it. A smoother demonstration is not enough evidence to grant wider access.

Then there is a sensible decision to make: keep the current limit, extend it or narrow it.

An agent can take on a bigger role. The authority to give it that role stays with the business.