The Jagged Frontier

If you take one idea from this course about what AI can and cannot do, make it this: AI capability is unevenly good across tasks, and the boundary between what it does well and what it does badly is invisible.

We instinctively grade tools the way we grade people — "this one is strong" or "this one is weak," a single dial of competence. AI does not behave that way. On one task it produces work that rivals your best people. On a neighboring task that looks just as easy, it fails — and, crucially, it fails with the same fluent confidence it shows when it succeeds. Researchers named this shape the jagged frontier: a capability map with sharp peaks and deep valleys sitting right next to each other, with no warning label marking where one ends and the other begins.

The risk is not weak AI. The risk is over-trusting AI exactly where it is silently bad.

That single reframe changes the leadership question. The naive debate — "should we use AI or not?" — is the wrong altitude. The real question is "on which tasks does AI lift us, on which does it quietly drag us down, and how would we even tell the difference?" A leader who understands the jagged frontier stops issuing blanket mandates and starts governing trust task by task.

The frontier moved from intuition to evidence in a 2023 field experiment run by researchers at Harvard Business School and BCG (Dell'Acqua, Mollick, Lakhani and colleagues). They gave 758 BCG consultants real consulting-style tasks, with some randomly given access to a frontier AI model and some not. Then they split the tasks into two groups: ones that sat inside the AI's capability frontier, and ones engineered to fall outside it.

The results are the cleanest illustration of the jagged frontier we have:

Same people. Same tool. Same apparent difficulty. Opposite effect — depending only on which side of an invisible line the task fell. Inside the frontier, AI was a powerful accelerant. Outside it, consultants using AI were about 19 percentage points less likely to reach the correct answer than colleagues working without it — because the AI gave fluent, plausible, wrong answers, and capable professionals trusted them.

These figures are time-stamped findings from one landmark study, not permanent laws — re-verify them at the live source before you quote them in a board meeting.

You don't need the math to govern this, but you do need the intuition, because it explains why the failures are confident rather than obvious.

The text engine under most generative AI is a large language model — think of it as autocomplete on steroids. It is trained to predict "what plausibly comes next," which is exactly why it is so fluent and articulate. But it is predicting plausible text, not retrieving verified fact. So when a task drifts outside what it reliably knows, it does not stop and say "I'm not sure." It keeps producing confident, well-formed output — it simply makes things up. The industry word for this is hallucination, and it is an inherent property of how these systems work, not a bug awaiting a patch.

A useful analogy for the boardroom:

Generative AI is a brilliant, fast, eager intern who has read almost everything, has no memory of your company, never says "I don't know," and occasionally invents things with total confidence. You delegate to it — and you check its work.

That single image carries the whole governance lesson. You would never let a first-week intern send a regulatory filing or approve a loan unsupervised, no matter how articulate they sounded. AI deserves the same posture: delegate generously, verify deliberately, and put a human check wherever a wrong answer is load-bearing.

The opposite of a blanket AI mandate is not caution — it is calibrated trust: deciding, task by task, how much to rely on AI based on two things you can assess even without technical depth — how reliable AI is on this kind of task, and how costly a wrong answer would be.

A simple 2x2 makes it usable on Monday morning. Sort any candidate task on how load-bearing a wrong answer is (low to high stakes) against how well-suited the task is to AI (a fluent-but-bounded drafting job vs. novel judgment):

The enemy of calibrated trust is automation bias — the well-documented human tendency to over-trust confident, automated output and switch off our own scrutiny. The more fluent the answer, the stronger the pull. Calibrated trust is the deliberate counterweight: you decide in advance where the human check sits, rather than discovering after the fact that everyone quietly deferred to the machine.

Use AI to expand your options, not to make your choice. Final judgment and accountability stay human — not as a courtesy, but because the tool cannot be held accountable and cannot tell you when it is wrong.

The same researchers observed that the people who got the most from AI did not all work the same way. They fell into two patterns — useful labels for any leader designing how their teams collaborate with AI.

The leadership takeaway is not to crown one pattern the winner. It is to recognize that "how should humans and AI work together here?" is a real design decision — and to make it deliberately, task by task, instead of defaulting to "everyone just uses the chatbot however." Centaur thinking suits tasks where you can confidently map the frontier and hand AI a bounded slice. Cyborg thinking suits messier tasks where the frontier is uncertain and a human staying interwoven is your error-catching mechanism.

Either way, the human is present by design — dividing or interweaving, never absent. That presence is what turns the jagged frontier from a hidden liability into a managed one.

Calibrated trust only becomes real when it shows up in how your organization talks about AI. Three habits turn the concept into practice.

1. Interrogate every accuracy claim. When a vendor or a team says "it's 90% accurate," never accept the headline. Ask the questions that reveal the shape of the risk:

• "Accurate in what way — and what do the failing 10% actually look like?"
• "Where is a wrong answer load-bearing, and what human check sits before it acts?"
• "Who is accountable for the failures?"

The shape of the errors matters far more than the headline number. Ten percent of harmless typos is nothing; ten percent of confidently wrong medical, legal, or financial conclusions is a crisis.

2. Use AI to expand options, not make the choice. AI is superb at generating drafts, alternatives, and syntheses — and weak at novel, accountable judgment. Research has found that when asked for strategic advice, models tend to return generic, fashionable "trendslop" rather than genuine insight, and can subtly flatter the person asking. Treat AI as a tireless option-generator and keep the decision human.

3. Keep accountability human and explicit. Every consequential AI-assisted output needs a named human owner who can stand behind it. The tool cannot be held to account, cannot explain itself to a regulator, and cannot feel the weight of a wrong call. That is — and should remain — a person's job.