Human in the loop AI: when review becomes a rubber stamp

Putting a person in front of an AI agent's work does not mean the person is making the call. A new Anthropic study shows trained experts mostly went with whatever the agent leaned toward, then explained the choice in the agent's own words and called it theirs.

If your business plans to let an agent draft decisions and have a manager approve them, this matters. The approval step is only a real check if the reviewer forms their own view.

This post covers what the study found, which kinds of approvals it applies to, and how we set up review so it holds up.

Anthropic published "How AI agents shape experts' judgment calls" on October 27, 2026. We read the research post on October 28, 2026. Everything in this section comes from it.

The team recruited 61 professional analysts with experience in causal inference, the work of figuring out from data whether one thing causes another. Eleven chose not to use the agent, so the main results cover 50 analysts who worked with Claude Code running Claude Opus 4.8.

The task was to advise a fictional farming cooperative on whether to expand an advisory program. Along the way, each analyst made four choices where both options were reasonable. One example was whether to measure productivity as yield per hectare or yield per unit of operating cost.

For each choice, a coin flip decided which option the analyst's agent would favor. The analysts were not told. The agent was told to make its honest best case, change its mind if the evidence called for it, and never mislead.

• Analysts went with the agent's lean 86% of the time across the four choices. With no influence, you would expect about 50%.
• When an analyst had already said what they preferred and the agent disagreed, they switched 60% of the time. When the agent agreed, they switched 6% of the time. Anthropic notes this rests on 38 cases, so the number is rough.
• Afterward, 77% of analysts repeated the agent's arguments when explaining their choices, sometimes word for word. Only 15% credited the agent.
• Years of experience, self-rated expertise and trust in AI did not predict who followed the agent.
• 52% said the final recommendation was more their own than the agent's. That group was about as influenced as the rest.

Anthropic is careful about limits. They made the agent's lean stronger on purpose so they could measure it. It was one task, one field and one model. Real agents may lean less, and real experts often know their field better.

The study also found good news for using agents. Analysts with the agent finished faster, and 3 of 50 made serious errors, against 6 of the 11 who worked without it. The groups were not comparable, so Anthropic does not read much into that either.

Most approval steps in a business fall into one of two kinds.

The first kind has a right answer you can check. Does the invoice total match the purchase order? Is the refund under the limit? Is the customer's address valid? For these, a person is often the slowest and least reliable check. Code can compare numbers every time.

The second kind is a judgment call. Which vendor should we pick? Should this customer get a discount to stay? Is this candidate worth an interview? Which leads get a call first? Two good managers could choose differently, and both would be defensible.

The study is about the second kind. When there is no single right answer, the work looks fine whichever way it goes. That makes it hard to notice that the agent made the decision and the person signed it.

When we build an agent that drafts decisions for a team, we design the review step with the same care as the agent. Here is what that usually includes.

1. Sort every approval into checks and judgment calls. Checks go to code first: totals, limits, duplicates, required fields. A person sees a check only when the code flags a problem. This shrinks the review queue so the real judgment calls get attention.

2. Ask the reviewer for their own pick first. For judgment calls, the review screen asks the reviewer to choose before it shows the agent's recommendation. It can be one click and one line of reason. Anthropic lists asking for the expert's view first as one design worth studying.

3. Have the agent argue both sides. Anthropic noticed the agents rarely made a strong case for the other option. We prompt the agent to give the best case for each option, and the evidence for each, before it says which it prefers.

4. Show the source records. The reviewer should see the source records the agent used: the order history, the contract clause, the email thread. A summary alone invites a quick yes.

5. Track how often reviewers change the agent's call. If a reviewer approves everything as written for weeks, that is worth a look. It may mean the agent is great. It may also mean nobody is really reviewing.

6. Run spot checks. Each week, a second person reviews a small random sample from scratch, without seeing the agent's view, and compares.

7. Keep queues small. A reviewer facing 200 near-identical items an hour will click approve. If the volume is that high, the routing in step 1 needs work.

This connects to how we scope what an agent can do in the first place. See how to scope AI agent permissions and human approval and human approval gates for finance AI agents for the permission side.

Say a subscription business wants an agent to handle cancellation requests. The agent reads the account, the support history and the usage, then drafts a reply. Some replies include a discount to stay.

The check part: the discount must be within the allowed range, the customer must be eligible, and the reply must not promise anything the billing system cannot do. Code handles all of that.

The judgment part: should this particular customer get an offer at all? Here the review screen first asks the account manager, "Offer or no offer?" with the account facts shown. Then it shows the agent's best case for each answer and its pick. The manager confirms or changes it. Every change is logged, and each month the team looks at how often managers disagreed and whether those customers stayed.

That gives the business a faster process without quietly handing a pricing policy to the agent's defaults.

Not every workflow needs this.

• If the decision has a right answer, skip the human step for routine cases and use automated checks plus testing before production.
• If the stakes are low and easy to undo, like the order of a to-do list, a reviewer adds delay with little benefit.
• If one owner reviews five items a week, a full review screen is overkill. A short habit works: decide before you read the agent's pick.
• If nobody on the team has the time to form a view, an approval step will become a rubber stamp no matter how it is designed. Decide that openly and shrink what the agent is allowed to do.

List the approvals your team does today or plans to hand to an agent. Mark each one as a check or a judgment call. Move the checks into code. For the judgment calls, add the "your pick first" step and the both-sides prompt, then watch the change rate for a month.

If you want help mapping this for your workflows, book a free automation audit. We build these review flows as part of our AI agent development and workflow automation work.