Working the jagged frontier: judgement as the core AI skill
A 2023 field experiment with 758 consultants showed AI lifts performance inside its capability frontier and degrades it outside, and that workers cannot reliably tell which side they are on.
TL;DR
- In a field experiment with 758 BCG consultants, AI raised performance sharply on tasks inside its frontier and lowered it on tasks outside.
- Inside the frontier, consultants completed more tasks, faster, with markedly higher quality, and weaker performers gained the most.
- Outside the frontier, consultants using AI did worse than those without it, and could not reliably tell where the line was.
The jagged frontier study, conducted by researchers including Fabrizio Dell'Acqua, Ethan Mollick, and Karim Lakhani with Boston Consulting Group, tested AI on professional knowledge work. Across 758 consultants, AI did not raise or lower performance uniformly. It created a jagged boundary: some tasks fall inside the model's capability and some fall just outside, even when they look equally hard.
AI's competence is jagged: it can ace a complex task and fail one that looks easy.
Inside and outside the frontier
Inside the frontier, the gains were large: consultants completed about 12 percent more tasks, worked over 25 percent faster, and produced work rated more than 40 percent higher in quality. The benefit was largest for below-average performers, who improved far more than top performers, which suggests AI compresses the gap between them.
Outside the frontier, the picture inverted. Consultants using AI on a task designed to sit beyond its reach performed about 19 percentage points worse than those working without it. Worse, the researchers found miscalibrated trust: people over-relied on AI exactly where it was weakest. The follow-up qualitative work described two effective working styles, often called centaur and cyborg, that divide labour between human and model deliberately rather than by default.
Confidently unreliable
This is the most important finding for any adoption programme, and it is routinely ignored. The danger is that AI is confidently unreliable, and that capable professionals cannot intuit where its competence ends. Productivity gains are real, and so is the cliff, which stays invisible from the inside.
That is why "give everyone a licence and encourage usage" is not a strategy. Usage without calibration produces confident, well-structured, wrong work, and people who are less likely to catch it. The skill that matters is knowing where the frontier runs for your specific tasks, and that is learned rather than assumed.
Building the skill
- Map the frontier for your own work. Identify, by task type, the places AI reliably helps and the places it quietly degrades quality.
- Teach calibration as well as prompting. The core capability is judging when to trust output and when to override it.
- Keep human review heaviest precisely where AI feels most fluent, because confident output is where miscalibrated trust does its damage.
The promise of AI in knowledge work is genuine. Capturing it depends mainly on whether your people know which side of the jagged line they are standing on.
Related reading
- Discovering AI's jagged frontier, and what we have learned since
- Hiring in the age of AI: the entry-level squeeze
Source: Navigating the Jagged Technological Frontier, Harvard Business School and Boston Consulting Group.