The headline from the BCG field experiment is hard to ignore. On tasks inside GPT-4’s capability frontier, consultants worked faster and produced stronger output. The paper is now peer-reviewed in Organization Science.

The part I keep thinking about is the reversal: on a task outside the frontier, access to AI made participants less likely to reach the correct solution.

The frontier is a management problem.

Two tasks that look similarly difficult to a human can sit on opposite sides of a model’s capability boundary. That creates two risks: hallucination and misplaced confidence about which tasks are safe to delegate.

AI governance should be designed around task boundaries, not around a generic instruction to “keep a human in the loop.”

A polished answer can narrow attention before anyone realizes an assumption is wrong. The human reviewer may be technically present and still be checking a frame they would not have chosen independently.

I would rather govern the workflow than the tool.

What can the model draft? What can it recommend? What evidence has to be checked independently? Which decisions require a fresh human analysis before the AI output is even shown?

I am bullish on AI. I just think the strongest implementations will look less like “give everyone a chatbot” and more like deliberate redesign of work, verification and accountability.

What I still do not know.

I still do not know how much of the productivity gain survives once teams redesign roles around the tool rather than simply adding it to existing work. That is the question I would want to follow over time.

REFERENCES / FURTHER READING

  1. Dell’Acqua et al. Navigating the Jagged Technological Frontier. Organization Science. 2026. Organization Science — Navigating the Jagged Technological Frontier ↗
  2. Harvard Business School AI Institute. Navigating the Jagged Technological Frontier. Harvard Business School AI Institute — Jagged Technological Frontier ↗
  3. Brynjolfsson, Li & Raymond. Generative AI at Work. NBER Working Paper 31161. NBER — Generative AI at Work ↗

These are personal research notes and commentary. I link the underlying sources so the evidence can be checked directly.