What large studies actually show about AI at work

Background

Most AI marketing quotes one number from one study. We have no Boteam customer results yet, so this page does the honest version instead: what the best-known research found, in whom, and what it does not say about your business.

The support-agent study (NBER)

Researchers studied customer support agents at a large firm given an AI assistant trained on the patterns of successful colleagues. Productivity rose, with the largest gains for less-experienced agents — the assistant effectively distributed the know-how of experienced staff.

What it does not say: that any business saves a fixed percentage. The mechanism — captured knowledge lifting the whole team — is the transferable part, and it is why the Company Brain area comes first in our playbook.

The consultant study (Harvard Business School / BCG)

Consultants using AI performed clearly better on tasks inside the model’s capability — and worse on tasks outside it, because they trusted the output anyway. The researchers called this boundary the “jagged frontier”.

What it does not say: that AI is good or bad at “work” in general. For an owner, the lesson is task selection plus review: pick jobs on the right side of the frontier and keep a person checking.

The product-team experiment (Harvard Business School / P&G)

In a field experiment with product development teams, individuals working with AI matched some outputs that previously took a small team — on specific, well-defined development tasks.

What it does not say: that teams are obsolete. It suggests a small business can take on work that used to need more hands, where the task is well-defined.

Limits, together

All three studies observed large organisations with resources a small business does not have. None measured a trades, clinic or agency business. They are the best available evidence for mechanisms — captured knowledge, the capability boundary, task-level leverage — not promises of outcomes.

Our approach: use these as reasons to run your own two-week pilot with a baseline and review, which costs little and produces the only number that matters — yours.