The jagged frontier: why your most confident AI users fail silently

In a now well-known field experiment, researchers from BCG and Harvard Business School (Dell'Acqua et al., later published in Organization Science) gave hundreds of consultants access to a frontier AI model. On tasks that sat inside the model's capability, the AI users were faster and produced markedly higher-quality work. On a task deliberately designed to sit just outside it, AI users were more likely to be wrong, because the model handed them a confident, polished answer that was subtly off.

The researchers called this the jagged frontier. AI is excellent on one side of an invisible line and unreliable on the other, and the line is uneven: it moves from task to task, and it is not marked. Nobody gets a warning that they have crossed it.

The intuition is that heavy users are your safest users. With the jagged frontier, the opposite can be true.

• Confidence outruns calibration. Someone who uses AI all day trusts it more, so they verify it less, which is exactly the wrong instinct near the edge of what it can do.
• The failures are silent. A wrong answer that looks right does not announce itself. It gets pasted into the deck, sent to the client, or shipped in the code, and the cost surfaces later, somewhere else, detached from its cause.
• Speed disguises it. AI makes the wrong answer arrive faster and cleaner than a human would produce it, and polish reads as competence.

This is why a headcount of people using ChatGPT tells you almost nothing about whether your team uses it well. The person generating the most output can be the one quietly shipping the most subtly wrong work.

Most learning dashboards measure inputs, and every input metric is blind to the frontier.

• Licences bought measure access, not skill. A seat is not a habit.
• Course completions measure attendance, not whether anyone changed how they work the following Monday.
• Self-report surveys measure confidence, and confidence is the very thing the jagged frontier turns against you.

The training itself is not the missing piece. Around 81% of Indian GCCs already run internal generative-AI training programmes (EY GCC Pulse Survey), and EY upskilled 44,000 of its own people before taking an AI academy to the market. Volume is not the gap. Knowing who has actually crossed the frontier and who is stuck below it is the gap.

You cannot fix an invisible line by running more generic training at it. You make it visible by measuring behaviour instead of opinions. Three steps.

• Score real fluency, not a quiz. A short conversational diagnostic reveals how a person actually works with AI: what they use it for, where they stop, whether they check the output, and whether they can tell a good answer from a plausible wrong one.
• Place everyone on a five-level map, from absent to multiplier. Now the jagged frontier stops being an abstract idea and becomes a picture of your specific team, with the exposed people visible.
• Price the gap in rupees. Map each level to the hours AI can realistically save and a sober productivity lift, and moving people up the map becomes an investment decision rather than a hope.

Discernment, not prompting, is the skill that separates a multiplier from a confident guesser.

India is expected to need roughly a million more AI-skilled professionals in 2026 (Nasscom), and AI-enablement roles are among the fastest-growing categories on Indian job boards. The spend is coming whether or not anyone can prove it works. The only question your CFO will ask, a quarter later, is whether it did.

A team that can name who sits where on the frontier, and what closing that gap is worth, walks into that review with an answer instead of an anecdote. That is exactly what our AI Litmus diagnostic builds in two weeks: a team maturity heatmap, the top three gaps holding you back, and a board-ready rupee ROI figure. We are taking 10 founding pilots this month at a founding rate. If proving AI ROI is landing on your desk this quarter, that is the conversation to have now.