Leadership in Practice
The Jagged Frontier: Why AI Helps Some of Your Team’s Work and Hurts the Rest
A field experiment with 758 consultants found AI users completed 12.2% more tasks, 25.1% faster, at 40% higher quality — on the work inside its range. On one task just outside it, they were 19 percentage points more likely to be wrong.
The experiment worth quoting
In 2023, researchers from Harvard Business School, Wharton and BCG ran a field experiment with 758 consultants. Half got access to GPT-4. Both groups worked on the same set of realistic consulting tasks.
Inside the tool’s range, the results were striking: the AI group completed 12.2% more tasks, 25.1% faster, with 40% higher quality.
Then the researchers added a task designed to sit just outside what the tool could reliably do. On that task, consultants using AI were 19 percentage points less likely to produce the correct answer than those working without it.
Same people. Same tool. Opposite results. The researchers called the boundary the "jagged technological frontier" — and the jaggedness is the whole story.
Why the frontier is jagged
The capability boundary isn’t a clean line by job title or seniority. It runs through a single role, sometimes through a single afternoon. The same analyst is inside the frontier writing a competitive summary and outside it judging whether the market data means what it appears to mean.
The tool doesn’t know where the line is. It answers with the same fluency on both sides. And that fluency is what makes the outside-of-frontier work dangerous: an answer that is confidently wrong reads exactly like one that is confidently right.
Why your rollout looks uneven
This explains the pattern most leaders see and can’t account for. Two teams, same tool, same training, opposite outcomes.
The team that got better was doing work inside the frontier and used AI to remove the drafting tax. The team that got worse was doing work outside it — synthesis, judgment, novel situations — and used AI to skip the part where they would have noticed the problem.
Usage statistics won’t tell you which team is which. Both are using it heavily.
Three questions that map the line for a role
You don’t need a framework. You need a conversation about the actual work.
- What do we already know how to do? That work is usually inside the frontier. AI should make it faster, and you should expect the gain.
- Where is the work ambiguous, novel or high-consequence? That’s outside. AI can support it — generating options, surfacing precedent — but a person still does the reasoning and owns the call.
- What does verification cost? If checking the output takes as long as doing the work, you haven’t saved anything, and you’ve lost the reps.
The part most teams get wrong
Inside the frontier, the people who gain the most are usually the least experienced. In the short term that’s good news.
In the longer term it means the junior work that used to build judgment no longer exists on its own. If a first-year analyst’s drafting is now handled by a tool, that analyst never practices the thing that would have made them a strong second-year. The reps disappear quietly, and the shortage shows up years later as a thin bench.
So the leader’s job isn’t to ban the tool. It’s to rebuild practice deliberately where the work used to build it by accident: explaining reasoning out loud, reviewing generated work line by line, and keeping some genuinely outside-frontier work with the people who need the reps.
The leadership move
Map the boundary before you scale the tool. Put verification where the stakes are. And count what people can do without it, not how often they use it.
That’s the difference between a productivity gain and a capability loss that takes a year to notice.