The risk in a well-argued no isn't that it's wrong. It's that you stop checking if it is.

Friday afternoon, someone hands you the stack you asked for: three vendor bids, a hiring recommendation, a pass on a second location. Your ops lead has already written up the reasoning on each one — a paragraph or two, clean and specific, explaining why to say no. You read it, it holds together, and you initial it. You have four other things due today. That thirty seconds of reading felt like due diligence. It wasn't.

A field experiment out of Harvard Business School, the University of Washington's Foster School of Business, and MIT Solve — The Narrative AI Advantage? A Field Experiment on AI-Augmented Evaluations of Early-Stage Innovations, by Jacqueline N. Lane, Léonard Boussioux, Charles Ayoubi, Ying Hao Chen, Camila Lin, Rebecca Spens, Pooja Wagh, and Pei-Hsin Wang — tested exactly this mechanism, using an AI stand-in for the person writing the memo. Two hundred twenty-eight people screened 48 real applications to MIT Solve's 2024 Global Health Equity Challenge, generating 3,002 individual screening decisions. Some screened with no help. Some saw an AI's pass-or-fail call with no explanation attached. Some saw the identical call with a written rationale — a paragraph explaining why. A separate panel of four experts scored all 48 submissions against the same rubric, giving the researchers a benchmark for which calls were actually right.

Worth knowing before you lean on this: only 72 of the 228 screeners were MIT Solve staff or contest judges who do this kind of evaluation professionally. The other 156 were university students — 113 undergraduate business majors and 43 participants in a leadership program for first-generation college students — recruited for course credit or a small stipend. It's one organization, one kind of decision — whether to advance a proposal — and a working paper that hasn't yet cleared peer review, though it's been in circulation and under revision since 2024.

Here is what the expert benchmark showed. Screeners who got a bare AI recommendation, with no explanation attached, matched the expert panel's calls 69.5 percent of the time — up from 66.0 percent with no AI help at all. That's a real gain. Screeners who got the identical recommendation, but with a written rationale attached, matched the experts only 64.3 percent of the time — worse than getting no help at all. The rationale didn't make people more accurate. It made them more compliant, and only on one kind of call. Left alone, people are skeptical of a rejection more than an approval: baseline agreement with a "fail" call was only 38 percent. A bare AI "fail" recommendation, with no reasoning attached, pushed that to 60 percent. The same "fail" call with a written rationale attached pushed it to 65 percent — and drove the increase in wrongly-rejected good proposals up by 14.9 percentage points, almost double the 8.4-point increase from the bare recommendation alone. The explanation didn't inform the override. It replaced it.

You have to know when to question AI-generated evaluations.Jacqueline N. Lane, lead author, quoted in Harvard Business School's Working Knowledge, March 12, 2025

In a large organization, one screener's bad call gets absorbed — averaged against a hundred other screeners, caught by a second committee, corrected next quarter. You don't have that buffer. In a company like yours, the "no" someone writes up for you most likely is the process. There is no second panel quietly scoring the same candidate, bid, or plan behind your back. If the reasoning reads clean and you are behind on four other things, checking it against the actual resume, the actual bid, the actual numbers feels redundant — the writing already did the work of feeling checked.

The part the research doesn't say, but that I'll say plainly: this study measured a machine's paragraph, not a person's. But nothing about the mechanism it isolated — a written rationale substituting for verification, concentrated specifically on the "no" answers — depends on who or what wrote the paragraph. A well-argued internal memo recommending you pass on a hire, a deal, or a location does the same job a fluent AI explanation did in this experiment: it makes agreement feel like it required less checking than it actually did. That's not something the study tested. It's something the mechanism it found makes obvious.

This week, pick one rejection sitting on your desk that already has a writeup attached — a pass on a candidate, a vendor, a location, a deal. Before you sign off on it, go find one fact in the underlying material that the writeup didn't mention, and ask whether it changes the call. Not an audit of the whole file. One fact the memo left out.

• If the writeup cites three reasons, find a fourth one it left out.
• If it's confident, that confidence is not evidence — check it anyway.
• If you can't find the underlying material to check, that's the finding worth acting on first.

Whether your instinct under a full inbox runs toward verifying or toward deferring is exactly the kind of pattern the Leadership Profile is built to surface — not to tell you your decision-making is broken, but to show you which of the "no's" crossing your desk are worth a second look before you initial them.