Where Should AI Stop and a Person Take Over?

I start with two questions when deciding how much authority to give an AI system: what would a mistake cost, and how easily could we undo it?

Those questions make the discussion specific. They help distinguish a task the system can handle independently from one it should prepare for approval.

The boundary belongs around the action and its consequences. It should be agreed before the system starts taking those actions.

Give different actions different levels of authority

Sorting routine incoming emails into queues is an example where I might allow automatic action. A wrong classification can be corrected, provided the queue design and checks prevent important messages from disappearing unnoticed.

Sending a payment, changing a contract, or making a binding commitment to a customer calls for a different approach. I would require approval because correcting the result could be difficult or expensive.

Between those situations are tasks where AI can prepare the work: gather the relevant information, draft a response, or propose a change. A person then checks the evidence and authorises the action.

The decision should also consider the information involved, the scale of activity, and whether an error could spread before someone notices it. A reversible mistake repeated many times can still become a serious problem.

Make human review practical

Asking someone to approve everything creates its own weakness. If a queue is full of routine recommendations, people may begin clicking through it.

I would rather direct attention to the cases that need judgement and give the reviewer enough information to make that judgement.

That means showing the evidence beside the recommendation. It means making uncertainty and exceptions visible. The reviewer also needs permission to reject the suggestion or pause the process, rather than being treated as an obstacle to throughput.

A review stage should have a clear purpose. What is the person checking that the system cannot be trusted to decide independently?

Check whether oversight is working

One useful signal is how often reviewers change or reject the system’s recommendation.

If nobody ever does, there are several possible explanations. The output may be strong. Reviewers may lack useful context. They may also be approving without checking.

The number alone cannot settle which explanation is right. It gives the owner a reason to examine the review process and check examples of accepted work.

Agree the response before the first mistake

A named person needs responsibility for coordinating the response when something goes wrong.

Before launch, the business should agree who can stop the system, who investigates errors, who communicates with affected people, and who authorises a restart or process change. Technical teams need clear responsibilities within that arrangement.

Otherwise, the first incident becomes a discussion about whose problem it is while the affected process waits.

Our AI Governance service helps establish responsibilities, approval boundaries, policies, and review arrangements. Controls relevant to a particular implementation also belong within that delivery scope.

If your teams are introducing AI into operational work, talk to us about deciding where it can act and where a person needs to take responsibility.

Back to Resources

Which of these sounds like your business?

Pick the closest, or send an enquiry and tell us the decision you are trying to make. We will tell you plainly whether we can help.

What happens next: we reply, then have an initial conversation. If an engagement is appropriate, a written scope sets out what you receive and what we need from you before any work begins.