Sudakov: AI recommendations filter out good ideas

22.08.2026

5 min read

Whoever judges first in a selection round frames everyone else’s thinking. In the Harvard Business Review field experiment, a language model judged first. 228 experienced evaluators followed.

Key takeaways

A machine’s rejection sways the group. 228 evaluators followed its rejection even when four experts wanted to keep the idea.
Sequence matters. When the recommendation comes first, the group must argue against the machine.
Rejections need a named owner. High-stakes decisions get human judgment first. Rejected cases face later review.

What does this mean for a selection round? The experiment compared three approaches and shows where the group abandons its own judgment.

228 evaluators, three approaches, one benchmark

Writing in Harvard Business Review on 19 August, Leonid Sudakov and Nathan Furr describe a field experiment assessing early-stage ideas. Their finding: the machine’s recommendation shifts judgment before people have assessed ideas independently. The setting was a real selection process.

228 experienced evaluators reviewed 48 actual submissions to a global MIT social impact competition. Three approaches competed: humans alone; recommendations from a large language model with narrative explanations; the same recommendations without further context. Four experts associated with the competition provided the benchmark.

Approach	What the group sees	What this means for the group
Humans alone	The submission without an automated preliminary decision	Slower, but with independent reasoning
Model without context	A bare recommendation without a narrative explanation	Rapid acceptance, little dissent
Model plus narrative	The same recommendation with a polished story	Agreement may increase because the story persuades

The underlying research appears in the Harvard Business School (HBS) working paper The Narrative AI Advantage? by J. N. Lane and co-authors. The researchers describe the same three approaches: human assessment, recommendations without explanations, and recommendations plus narrative. HBR draws out the leadership question. The implications determine the sequence for the next round.

Rejection has more influence than approval

The tool encouraged evaluators to reject submissions the expert panel approved. People largely followed. Promising projects fell through the filter.

The finding extends beyond innovation labs. The same mechanism operates in investment committees, product committees, supplier selection and internal investment rounds. A model that sounds fast and confident lowers the threshold for rejection. Dissent requires a better argument than the machine’s and usually more time than the meeting allows.

That makes sequence crucial. When the machine’s recommendation comes first, the group searches for reasons to challenge the machine’s judgment. The decision comes before anyone opens the figures.

BCG describes the same effect as lock-in

Boston Consulting Group (BCG) published a related finding in early August. It was a separate study. When rules, decision processes and operational knowledge become embedded in a model stack, cognitive lock-in develops. The organization believes it still makes decisions, yet does so within an architecture it does not fully control.

BCG identifies signs visible without examining the technology. Teams stop explaining why because they follow the model’s reasoning. Business units complain that the tool misses their actual needs. And strategy bends around what the model does well.

The HBR experiment makes the same point measurable. The narrative version influences evaluators at least as strongly as the recommendation without explanation. A polished story can even increase reliance.

Who owns the rejected idea?

In Germany and the European Union (EU), responsibility remains with the individual and the company. Since early August, the AI Act has required transparency for certain systems. It does not make selection decisions for anyone. A log entry reading “Model rejected” provides no defensible justification before regulators, a works council or in the event of subsequent harm.

Once assessments affect careers, sites or major investments, employee participation becomes mandatory. The works council must know which criteria the tool considers, which it skips and who can still halt an automated rejection. Otherwise, the dispute surfaces in later proceedings. By then, the group has already lost ideas.

The rule is broad and robust. Low-stakes decisions can use automated screening; high-stakes decisions get human judgment first. The machine contributes a second opinion or a challenge. Every approval and rejection has a named owner responsible for reviewing rejected cases later.

Three questions for the next round

Three questions turn the finding into an agenda item. First: does the group see the idea before the machine’s recommendation appears? Second: can someone vote against it without losing face? Is that recorded in the minutes? Third: does an independent reviewer check a sample of rejections, as the experiment’s expert panel did?

If any question remains unanswered, the process lacks oversight. Sudakov and Furr insist that independent judgment must remain strong enough to overrule a persuasive system. That sounds soft. It becomes hard-edged once budgets, liability and a pipeline of internal bets depend on it.

The question for the next discussion among peers: is a list of recently rejected proposals available, showing which rejections originated with the model? Without that list, recommendations are already in control. Nobody decided to make it so.

Frequently asked questions

Can a model make the initial rejection?

Yes, for low-stakes screening. For investments, staffing and strategic bets, human judgment comes first; the model then provides a challenge.

Is the model’s explanation sufficient documentary proof?

No. The narrative explains how the system reasons. It replaces neither personal sign-off nor subsequent review of rejections.

How does the AI Act affect this process?

It requires transparency for certain systems. Selection and liability remain with the individual and company. A log entry alone cannot justify approval.

When must the works council be involved?

As soon as assessments affect careers, sites or major staffing decisions. Criteria, the right to challenge decisions and sample reviews belong in the agreement.

Which first step makes the biggest difference?

In the next round, show the machine’s recommendation only after the initial human vote and assign every rejection to a named owner.