Human-in-the-lead, not human-in-the-loop

In late October 2026, Anthropic published a study that every team putting AI agents to work should read. The headline isn't that AI makes mistakes. The more useful finding is that experts can adopt an agent's framing of a decision, sign off on it as their own, and never notice.

I'm a technical social scientist. I studied how people think and interact before I spent about 12 years teaching machines to reason. The hard part of AI was never only the model. It's how people and systems make choices together, and who answers for them.

That's why I've stopped being reassured by "human-in-the-loop." Too often it means a person somewhere near the end, approving. I want the human in the lead: framing the question, setting the options and owning the call, with AI doing the heavy lifting in support.

What the study found

Anthropic's researchers asked 50 professional analysts, all with causal-inference experience, to complete a realistic statistical analysis using Claude Code. The task, set at a fictional farming cooperative, involved four analytical choices between two defensible options, such as measuring productivity as yield per hectare or per unit of operating cost, plus a final recommendation.

Without telling the analysts, the researchers used a coin flip to give each agent a "lean" on every decision. The agent was told to make the best honest case for its lean, to change its mind if the evidence demanded it, and never to mislead. Here's what happened, according to Anthropic's write-up (opens in a new tab).

1. Across the four choices, analysts went with the agent's lean 86% of the time. Chance would be about 50%.
2. When analysts had already stated a preference, they switched away from it 60% of the time if the agent disagreed, compared with 6% if it agreed. That comparison rests on 38 cases, and the authors call it imprecise.
3. Afterward, 77% explained their choices using the agent's arguments, sometimes word for word. Only 15% credited the agent.
4. 52% rated the final recommendation as more their own than the agent's.
5. Experience, education and expertise did not reliably predict who deferred.

Anthropic is careful about the limits. The lean was deliberately strengthened, the study covered one task in one domain with one model, and the analysts were new to farming. The researchers also found that analysts who used the agent made far fewer serious errors than a small, less experienced group that worked without it.

So this isn't a case against AI. It's good evidence about where human judgment quietly slips away.

Framing is the real power

Many of the decisions that matter most at work have no answer key. Which definition of "productivity," "active customer" or "high risk" should we use? Which comparison is fair? Which records count? The evidence supports more than one reasonable answer, and the choice reflects someone's assumptions.

In those moments, whoever lays out the options and argues one of them well sets the frame. Agents are fluent, fast and tireless at that. The study notes that the agents were often open about how results would change, but they rarely made a strong case for the other option. With one side argued well, people took it.

I've said for years that one version of the truth is a myth. Every enterprise I've worked in, Credit Suisse, JPMorgan and BNY among them, had more than one. Real organizations hold competing, legitimate definitions of the same thing. The goal isn't to collapse them into one answer an agent hands you. It's to see them, understand where each came from, and make a deliberate call.

From in the loop to in the lead

Many firms describe their AI workflows as human-in-the-loop because a person approves the output. The study shows why that isn't enough. A signature at the end is not an independent check if the important calls were framed upstream and never surfaced.

For a person to be in the lead, the system has to do four things.

1. Show the alternatives. At each consequential decision, present the defensible options, each argued fairly, and show how the result changes under each.
2. Show provenance. Make clear where each option, assumption, definition and number came from: the source data, a business rule, an analyst or the agent.
3. Ask before it tells. Get the expert's view before offering the agent's. In the study, the agent usually spoke first.
4. Record ownership. Log who made each call and why, in their own words, so the record reflects how the decision was really made.

This doesn't mean slowing everything down. It means making the few decisions that carry real weight visible instead of quietly nudged.

Train the experts to push back

The most uncomfortable finding for leaders is that seniority didn't protect anyone. You can't hire your way out of deferral. You have to build the habit.

Treat pushback as a professional skill. Teach subject-matter experts to ask: What's the other reasonable option? What would have to be true for it to be right? Whose assumption is this? Give people time to form a view before the agent offers one, and reward the person who questions a consensus that arrived a little too easily. As the study's authors note, people who stop making these calls may lose the ability to make them.

Foundations make judgment visible

This is where the unglamorous data work pays off. A good semantic layer or ontology doesn't just organize data. It writes down the definitions and relationships people would otherwise carry in their heads. When "productivity" has two documented definitions with clear lineage, an agent's lean toward one becomes visible as a choice, not a fact.

At Reclassify AI we were talking about human-powered AI, and keeping people in charge of decisions, back in 2022. The tools have changed a great deal since. The principle hasn't. Machines can do remarkable work, but they bring no common sense or accountability of their own. People do.

Decisions belong to people

Agents will keep getting better at the hands-on work, and that's good news. It also means judgment calls become a bigger share of what experts are for. Decisions belong to people, not to data and not to AI.

The technology changes fast. Human behavior changes slowly. Until firms can see which version of the truth they acted on, and who decided, the loop will keep turning into a rubber stamp. Put a person in the lead instead.