Almost every responsible guide to AI automation, including ours, recommends keeping a person in the loop: the system drafts, routes or classifies, and a human approves. The advice is sound. But it is often treated as a box to tick rather than a piece of work to design, and recent research suggests that a poorly designed approval step can give a false sense of safety while wearing out the people asked to perform it.

This article looks at what goes wrong with human review of AI output and how a small business can build approval steps that hold up.

Why oversight is not automatic

The first problem is trust that grows faster than accuracy. In EY's 2026 survey of 604 respondents at Swiss companies, 40% said they considered AI answers more reliable than manually researched information in some situations, particularly under time pressure. Time pressure is exactly when a reviewer is most likely to approve without reading.

The second problem is fatigue. A Harvard survey of more than 1,400 US workers, reported by Euronews in March 2026, found that people whose job involved monitoring AI outputs reported 12% more mental fatigue than those who did not. Productivity began to fall once employees juggled more than three AI tools, and those reporting the strongest symptoms also reported more major mistakes. Reviewing machine output turns out to be demanding work, not a light extra duty.

The third problem is that errors cluster where they are hardest to spot. An Anthropic analysis covered by Euronews Next found that its model's success rate falls as tasks get longer or harder, and that once the cost of human checking and correction was included, the estimated productivity gains shrank. Long, complex outputs are both more likely to contain mistakes and more tiring to review.

Review exceptions, not everything

One answer is to change what people review. Speaking at a Google Cloud event in Doha, Alex Rutter, who leads the company's AI business in Europe, the Middle East and Africa, said Google currently advocates keeping "a human in the loop to verify for the business the results", with people looking at the exceptions rather than the entire dataset. He added that how much oversight is needed depends on the task and the organisation's appetite for risk.

In practice, that means designing the automation so that it sorts its own output into what clearly passes a defined check and what does not. A person then spends their attention on the second group, where it is needed, instead of skimming hundreds of routine items and missing the one that matters.

Not everyone in the industry favours the framing. Cisco's president, Jeetu Patel, has questioned the human-in-the-loop model and talks instead of AI being present in every loop, with agents treated as digital co-workers. For a large technology company with extensive security infrastructure, that may be a reasonable direction. For a twenty-person firm in Switzerland, a clear human decision point remains the more prudent design.

What a good approval step looks like

A useful approval step has five properties:

  1. It is placed before the irreversible action. Sending, paying, publishing, signing or deleting. Review after the fact is an audit, not an approval.
  2. It shows the evidence. The reviewer sees the source next to the output: the original invoice beside the extracted figures, the customer's question beside the drafted reply.
  3. It asks a specific question. "Do the amount and supplier match the purchase order?" is answerable in seconds. "Is this OK?" invites a reflexive yes.
  4. It is sized to human attention. If someone must approve two hundred items a day, the step will be skipped in spirit even if not in fact. Reduce volume through automatic checks, and route only exceptions.
  5. It leaves a record. Who approved what, when and on what basis. This protects the reviewer as much as the business.

A checklist for your own workflows

  • List every point where AI output leaves your business or changes a system of record.
  • For each, name the person who approves it and the question they must answer.
  • Estimate daily review volume. If it exceeds what one person can genuinely read, add automatic checks first.
  • Limit how many AI tools any one person is expected to supervise.
  • Sample approved items each month and check them independently, to see whether reviews are real.
  • Rotate review duties where possible, and treat review as part of the job, not something squeezed in between other work.
  • Revisit the design when error rates change, in either direction.

The bottom line

A human in the loop is only as good as the loop. Approval steps work when they sit before the consequential action, show the reviewer what they need, ask a precise question and keep the volume within human limits. Designed that way, they are what allows a small business to automate with confidence. Designed carelessly, they become a signature on work nobody read.

Want help designing review steps for your own processes? The free AI Readiness Audit maps your recurring tasks and ranks what is worth automating first.

Sources

  • EY Switzerland, "Artificial intelligence widely established in Swiss companies – but many are still in the early stages of scaling", 27 May 2026. link
  • Euronews, "'AI brain fry': Why your brain feels fatigued after using AI chatbots at work", 10 March 2026. link
  • Euronews Next, "Will AI replace jobs? Anthropic report finds the answer is not so straightforward", 19 January 2026. link
  • Euronews, "From answering questions to carrying out complex tasks: The next phase of AI", 28 September 2026. link
  • Euronews, "Cisco president warns AI agents need 'background checks' like human employees", 11 February 2026. link