AI agents: what to check before they act

Part of the guide What is AI governance, in plain words?

What makes an AI agent different?

An AI agent can work through a task and take action for its user. That might mean changing a customer record or making a payment. Its access to data and tools is therefore part of the decision about how to use it.

Oliver Patel highlights a practical question for organisations: how to stop agents accessing systems and sharing data in ways they should not. His observation puts the focus on the agent's reach.

What should an agent be allowed to do?

IMDA recommends choosing suitable tasks and limiting the agent's freedom to act. That includes which tools and data it can access. A useful starting question is: what does this agent need to do its job?

The framework includes an example from Dayos, which handles IT support requests. Dayos assessed each type of request for its impact, whether an action could be undone, and whether a person could oversee it.

Dayos then used different levels of permission. Some actions can run automatically, others need approval, and some are off limits. The diagram shows that company's approach. It is an example to discuss, not a set of permissions to copy into every firm.

What does useful human approval look like?

Useful human approval starts with clear checkpoints, as IMDA recommends. The guidance also describes automation bias. That means trusting a system too much because it has worked well before.

In IMDA's Tencent example, actions such as editing files need human approval by default. Complex computer commands are explained in plain English to help the user decide. Suspicious commands still need approval even if they were approved in advance.

For a buying or setup discussion, ask the team to show the approval screen and explain which actions trigger it. The Tencent example gives useful questions to ask.

How can you test before giving wider access?

Test the agent and control which services it can reach, as IMDA recommends. Its GovTech Singapore example shows a small start before wider use.

GovTech Singapore first limited its coding tool to employees working on low risk systems, with no external tools. It prepared shared records of activity, monitoring and rules for connecting approved external tools. It also tested potential attacks to check the controls worked.

That example offers questions for a trial. Who can use the agent? Which systems can it reach? How will the team check its behaviour? It describes one rollout, not a promise that the same setup suits every firm.

What should people using the agent know?

People using an agent need clear information and training, under IMDA's approach. Its Workday example says users are told upfront that the tool uses AI.

Bring those points into the rollout discussion. How will people know they are using an AI agent? What training will they receive? How will approval points be explained?

Does this apply if you buy the agent?

Yes, the framework covers agents you buy as well as those you build. Its questions about access, human approval, testing and user training cover both routes.

Use those topics to structure a conversation with the supplier or your own team. Keep the country and purpose clear: this is Singapore's practical governance guidance. It is not proof that a tool meets UK law.

What to do

1. List the task, tools and data the agent needs, then discuss the limits on its access.
2. Identify the points where a person must approve an action.
3. Ask how the agent will be tested and kept within its approved services.
4. Check how users will be told about the agent and trained to use it.

Common questions

Must a person approve every AI agent action?

IMDA recommends meaningful approval points. Its Dayos example allows some actions to run automatically, requires approval for others and blocks some altogether. Those are choices in that example, not universal settings.

Is this a UK legal checklist?

No. The framework used here is Singapore's guidance on responsible agent use. This article explains its practical approach; it does not set out the laws that apply to a particular UK firm.

Who remains accountable when the agent acts?

IMDA's framework emphasises that humans remain accountable. It recommends clear checkpoints where human approval is needed.

Where next

• What is AI governance, in plain words?
• What to ask an AI supplier before you buy
• A DPIA for an AI tool: when and how

This is an independent editorial guide, not legal advice.