Human in the loop. Attention out of the loop.

Originally published in AI in Practice on LinkedIn on 25 October 2026.

“Don’t worry. There’s a human in the loop.”

Ok, righto… which human? And what are we taking off their plate so they can do that properly?

If it’s the same person already dealing with customers, keeping projects moving and covering someone else’s workload, we should have that convo before connecting another agent.

Putting their name in an approval box doesn’t create another hour in their day.

An AI agent that needs your attention all afternoon has an interesting definition of “taking work off your plate”. It can produce something quickly, then hand you the job of working out whether it’s correct, what it missed and whether it should do what it’s proposing next.

Ask permission for every tiny step and you can end up turning the human into a full-time approval clerk. Repeated prompts risk becoming something they click through to keep things moving.

Before you count APPROVED as effective oversight, check what the reviewer actually assessed.

I use Claude, Codex and Hermes, among other tools and workflows in my own custom setup. I review their work and make decisions about what they can do. I want that arrangement to give me more capacity to do the work that matters.

These are issues I’m seeing and working through with businesses and their teams. The technical setup is part of it. So is deciding who supervises the work and how that fits into the job they already have.

Count the attention the whole job takes

When we talk about productivity, I’m interested in what happens after the agent produces an answer. How much attention did it take to get that answer over the line?

That includes:

• Explaining the job and finding the context it needs.

• Checking the result against the source, including what it left out.

• Correcting mistakes and dealing with exceptions.

• Picking the task back up after another interruption and remembering what the hell was happening.

Those minutes belong in the calculation.

For a hypothetical example, say an agent prepares 20 customer replies. At three minutes just to check each one, that’s an hour before corrections or exceptions. Whether that’s an improvement depends on what the same work took before, and whether the replies meet the standard you need.

Now imagine that hour lands with the person everyone already goes to when something gets complicated.

You might have saved preparation time while adding work to your busiest, most experienced employee. That can still be a worthwhile trade. Price it in, and decide what else they can realistically do.

Ask WHOSE time you’re saving.

The AI cowboy selling you a “digital workforce” should be able to explain this part too. Who’s supervising it, how much attention does it require, and what happens when that person is unavailable?

“Someone will check it” leaves an awful lot for the customer to sort out.

If you scale AI output without considering the capacity to review it, you can build a very impressive queue of unfinished work. Then the pressure lands on someone to clear it faster.

Give the reviewer time and context

Human attention is a finite resource. Good AI governance needs to budget for it.

That means looking at the decisions you’re putting in front of people, how frequently they arrive and what it takes to assess them. A routine draft and a request involving private customer information might appear through much the same interface. Have you made the difference obvious?

Can the reviewer inspect the evidence? Do they understand the consequences? Can they say no without being treated as the person holding everything up?

NIST’s AI Risk Management Framework calls for defined oversight responsibilities, appropriate training and executive responsibility for AI risk decisions. Putting an employee at the end of the workflow shouldn’t become an excuse to leave those decisions with them.

“But a human approved it” deserves a follow-up: what did you give that human to work with?

Please don’t dump a transcript of the agent’s adventures onto their screen and call that transparency. Show the proposed action and the evidence supporting it. Make the affected customer, cost or commitment clear where relevant, along with material uncertainty and whether the action can be reversed.

Then give them a practical way to reject, correct or escalate it.

Choose the pauses that matter

Removing prompts might make the experience smoother while allowing a mistake to travel further.

There’s relevant research from before the current generative AI wave. In a 2021 controlled decision-support experiment, researchers found that designs encouraging more deliberate engagement reduced overreliance compared with explanations alone. Participants also rated the most effective designs less favourably. An easier experience wasn’t necessarily producing better decisions in that study.

Now consider the approval queues we’re building around agents. AWS’s current Agentic AI Lens warns about reviewer fatigue and rubber-stamping when every action goes through human review regardless of risk. It recommends matching approval requirements to the consequences and giving reviewers enough context to decide.

The study examined a controlled decision task; AWS provides practical guidance for agent workflows. My takeaway is that we need to distinguish a pause that helps someone assess a decision from another interruption they’re expected to clear. Some friction earns its place. We should be able to explain what it helps the person catch.

I’d start by separating the work like this:

• Work within agreed authority. Reading approved sources or preparing a private draft using approved tools may proceed without permission at every step, within tested limits on access and activity.

• Work that needs a meaningful decision. A customer message, price change or financial commitment may warrant a checkpoint before the action happens. The reviewer should see the proposed action clearly.

• Work that crosses a boundary. A recipient outside the agreed scope, additional data access, materially conflicting evidence or spending beyond an agreed limit should trigger a pause or escalation. The agent shouldn’t quietly expand its own authority to finish the task.

Those boundaries need to be enforced through permissions and checks. Instructions alone aren’t enough to establish that the controls work.

And “reversible” deserves a closer look. You might be able to delete a draft, but that won’t undo private information already sent to an unapproved service. A series of small actions can also add up to a significant commitment. Consider the total, not just each click.

Start with one recurring task

There’s room to remove unnecessary decisions too.

Take invoice processing. AI might help extract information from an awkward document. Once that information has been validated, an agreed spending threshold can be applied through rules-based automation.

You don’t need a model having a little think about your spending policy every time an invoice arrives. Test the rule and give someone responsibility for keeping it current.

To work out where to make these changes, sit with the person reviewing one recurring task. Watch where they leave the approval screen to find something out. Ask which checks they repeat and which decisions genuinely require their judgement.

Use what you learn to:

• Remove duplicate checks and provide missing context at the point of decision.

• Limit outstanding work to what the team can responsibly supervise.

• Define what pauses or escalates when the reviewer is unavailable. For actions requiring approval, silence shouldn’t become permission.

• Compare the whole task with the previous process, including checking, corrections and quality. Keep track of whose workload changed.

Use the delivery routines you already have to make those calls. Before adding another dashboard, find out whether the existing one can show where work is waiting and who has capacity.

If your business case only stacks up because their time is missing from it, go back and do the bloody calculation again.

I’m James Mark Lewis, founder of James Mark Digital. I work hands-on as an AI advisor and fractional digital leader, helping businesses work through decisions like these. That includes pressure-testing the business case, understanding the impact on the team and getting clear about who owns what once the technology is running.

If this sounds familiar, DM me on LinkedIn or email james.lewis@jamesmark.com.au. Tell me what you’re trying to get done and where it’s getting stuck. We can work through whether I’m the right person to help.