Splitting Work Among AI Helpers Rarely Pays Off, Study Finds

When a company wants an AI to research something big, it often splits the job: one AI hands pieces to several helpers, who hand pieces to more helpers, like a manager building a team. It sounds efficient, but it costs more money and, according to a new paper by researcher Rong He, it doesn't produce more results. Every layer down the chain loses some of what it found. Push that loss through five layers, and a mountain of findings becomes a handful. The author verified this with 20,000 randomly generated test structures — the math held every single time.

Worse, the helpers themselves are imperfect. Looking at real recordings of AI research sessions, roughly one instruction in sixteen came back off-target — the helper answered a slightly different question than asked. Each extra layer multiplies that small error. Meanwhile, the layers do cost real money: one company's flat AI setup billed 39% more than a simple estimate predicted, and running two layers only beat one layer once a job grew past roughly 400 findings.

The one genuine benefit is control, not results. If you keep all the information in one place, the AI's 'memory' fills up until it starts forgetting things. Splitting work into layers keeps that top-level memory cleaner. So depth buys order and tidiness — 'integrity,' as the paper calls it — but not more answers.

The paper's most practical finding is about habit. By studying over 740,000 real AI tool calls, the author found that people delegate at the very start of a task, when nothing is wrong yet, rather than when the AI's memory genuinely fills up. That means most delegation is autopilot, not necessity. The study estimates only 0.7% to 11.3% of real sessions were worth splitting — yet about 7.8% actually did it. If you're paying for multi-AI setups, the advice is simple: check whether the job is genuinely huge before adding helpers.