Study: AI Teams Work Better Without a Boss Checking Their Work

Companies are increasingly building 'AI teams' — several AI assistants working together on one job, like writing a market report. Most of these setups include a boss: one AI that reviews what the others produce and can send work back for another try. It feels sensible. Humans work this way. So researchers set out to test whether the boss actually helps.

They ran a clean experiment. Five AI agents, same roles, same instructions, same tools, same underlying model. The only thing that changed was whether the manager AI was allowed to reject a draft and demand a rewrite. They did this task 86 times and had a panel of AI judges score every report, plus an automated check for factual accuracy. The result surprised them: the flat teams — no boss — won. Their reports were rated more useful and clearer to read. The reports were the same length, but the hierarchical team's writing hedged 53% more, meaning it padded sentences with wishy-washy caveats like 'it could be argued that.' Every rewrite round nudged clarity down. And the extra management layer burned 51.5% more computing cost for no quality gain.

The most telling detail: the writer's first draft was just as good under both systems. The damage happened inside the revision loop. The boss wasn't fixing anything — it was smudging a good draft.

The lesson is simple and it applies well beyond AI. A supervisor earns their keep when they can check facts and catch real errors. When they can only offer opinions — 'make it stronger,' 'add more nuance' — they add cost, blur the message, and slow everything down. If your company is paying for AI tools that 'review' other AI output, this study suggests you may be paying more for a worse result. Ask whether the checker can verify anything. If not, skip the middleman.