The Jagged Frontier

The hardest thing to predict about AI is not how good it is — it is how unevenly good it is. It can pass a bar exam and then miscount the letters in "strawberry." Picture a coastline, not a fence: the boundary of what it does well is jagged, with brilliant inlets right beside blind coves.

Brilliant and clueless, side by side

The unevenness is the point. Tasks that feel hard to us — draft a contract clause, summarize a dense report — can sit inside the frontier, where the model shines. Tasks that feel trivial — count items, follow a precise spatial rule — can sit outside it, where it fails with full confidence. Difficulty for a person is a poor guide to difficulty for the model.

The over-trust trap

A controlled study of 758 consultants put numbers on this. On tasks inside the frontier, those using AI finished 12.2% more work, 25% faster, and at higher quality. On a task chosen to sit outside the frontier, the AI users were 19% more likely to get it wrong than those working without it — because the fluent, confident output pulled them over a cliff they could not see.

You cannot always tell which side you are on

If the frontier were marked, this would be a non-issue. It is not. The same prompt style produces a brilliant result and a confidently wrong one, with no shift in tone to warn you. So the working rule is to verify anything that matters — especially counts, dates, names, and citations, the places the coastline tends to cut inward.

The most useful model for a manager

You do not need to know how the model works to use this. You need to know the boundary is jagged, that the model will not flag when it crosses it, and that your judgment is what keeps work on the right side. That is the core of directing an agent well.

Source: Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (2023).