The Jagged Frontier: Where AI Helps Your People and Where It Quietly Hurts

In 2023, researchers from Harvard Business School ran one of the most rigorous field experiments yet conducted on generative AI at work. Partnering with Boston Consulting Group, they gave 758 consultants, around 7% of the firm's individual contributors, a set of realistic tasks and randomly assigned access to GPT-4. The results, published by Dell'Acqua and colleagues, should reshape how every leader thinks about AI adoption.

For tasks that sat inside the model's capabilities, the effect was dramatic. Consultants using AI completed 12.2% more tasks, worked over 25% faster, and produced work rated more than 40% higher in quality. If the story ended there, the advice would be simple: give everyone the tool. But the story does not end there.

For one carefully designed task that sat outside the model's capabilities, consultants using AI were 19 percentage points less likely to reach the correct answer than colleagues working without it. The AI did not merely fail to help. It actively pulled capable professionals towards confident, plausible and wrong conclusions.

The frontier is real, and it is invisible

The researchers called this boundary the jagged technological frontier. Some tasks fall inside it and AI handles them beautifully. Others, often superficially similar, fall outside it, and AI degrades human performance. The problem for leaders is that the frontier is jagged and largely invisible. You cannot tell from the surface of a task which side of the line it sits on.

What separated the people who were harmed by AI from those who were helped was not seniority or intelligence. It was engagement. The consultants whose performance dropped tended to accept AI output uncritically, interrogating it less and deferring to it more. The researchers described this as unengaged interaction. The lesson is uncomfortable: the danger is not that AI is wrong, it is that AI is wrong in a fluent, persuasive voice, and tired humans stop checking.

The great equaliser, and the homogenisation risk

There was a second striking finding. Among tasks inside the frontier, the lowest performers gained the most. Below-average consultants improved by 43%, while above-average consultants improved by 17%. AI compressed the gap between weak and strong performers. For organisations, this is genuinely good news: capability can be levelled up quickly.

But there was a cost. The ideas produced by AI-assisted consultants were markedly less varied than those produced without it. When everyone draws on the same model, outputs converge. Over time, the homogenisation of thinking is a strategic risk for any organisation that competes on originality.

What this means for leaders

The job is no longer to decide whether your people use AI. They already do. The job is to help them work the frontier deliberately. That means three things. First, teach interrogation as a discipline, not a nicety: every AI output is a draft to be challenged, not an answer to be accepted. Second, map your own frontier by identifying which of your team's recurring tasks AI does well and which it does badly, then make that map explicit. Third, protect diversity of thought by treating AI as a starting point for human judgement rather than a replacement for it.

Try this

Take three recurring tasks from your team and run a simple test. Complete each one with AI and without it, then have a colleague who did not know which was which rate the quality. You will quickly learn where your frontier sits, and that map is more valuable than any vendor demonstration.

If you want a structured way to assess how well your team captures and challenges its own reasoning, the Decision Trace Readiness diagnostic is a useful starting point.