Jagged frontier AI: what product managers should hand to AI and what to keep

The jagged frontier is the idea that AI's abilities do not rise evenly with task difficulty. It is excellent at some tasks a person finds hard and fails at others that look just as easy, and you cannot tell which is which from the outside. The HBS AI Institute's summary of the study puts it this way: "The study introduces the concept of a 'jagged technological frontier,' where AI excels in some tasks but falls short in others."

The evidence comes from a 2023 field experiment run by Harvard Business School researchers with Boston Consulting Group, covering 758 consultants, which the HBS AI Institute notes was 7% of the firm's individual contributor workforce. Co-author Ethan Mollick reported the headline numbers in Centaurs and Cyborgs on the Jagged Frontier: on tasks inside the frontier, consultants using GPT-4 "finished 12.2% more tasks on average, completed tasks 25.1% more quickly, and produced 40% higher quality results than those without."

What happens when a task sits outside the frontier?

The researchers also built one task designed to sit just outside what GPT-4 could do: a business problem where the AI would give a wrong but convincing answer. Mollick reports that consultants "got the problem right 84% of the time without AI help," while those using AI were "only getting it right 60-70% of the time."

That result is the part a product manager should remember. The AI did not refuse or hedge. It produced a polished answer, the consultants trusted it, and the write-ups were still better than the ones written without AI. Quality of prose and correctness of the answer came apart.

Two limits apply. The study used GPT-4 in 2023 on consulting tasks for a fictional shoe company, so the exact boundary has moved since then. And Mollick himself warned that "the technological frontier is not just jagged, it is expanding," which means a task outside it last year may be inside it now. What still holds is the shape: capability stays uneven, and the uneven parts are not labelled.

Which product manager tasks sit inside the frontier?

Frontier fit is a property of a task, not of a job title. Four questions place a PM task on one side or the other. Answer each one honestly for a specific, recurring task:

1. Is there a lot of public writing about tasks like this? Summarising interviews, drafting release notes and writing test cases are all heavily represented in what models learn from. A decision about your company's specific pricing is not.
2. Can you check the output faster than you could produce it? A clustered list of 40 feedback comments takes minutes to spot-check against the originals. A market-sizing number built on assumptions you cannot see takes longer to verify than to build yourself.
3. Does the right answer depend on facts only your team has? Internal politics, what a key account said on last week's call and why a feature was cut last quarter are invisible to the model unless you supply them.
4. Would a plausible but wrong answer be caught before it does damage? A rough draft reviewed by you is safe. A number pasted straight into a board deck is not.

A task that clears all four is inside the frontier for practical purposes. Fail question 3 or 4 and it is outside, however easy it looks.

PM task	Likely side	Why
First draft of release notes from merged tickets	Inside	Common format, easy to check against the tickets
Clustering 40 interview notes into themes	Inside	Checkable against the notes; misses show up quickly
Rewriting a technical spec for a sales audience	Inside	Common transformation; you know both audiences
Generating edge-case test scenarios for a new flow	Inside	Easy to review, cheap if one is wrong
Choosing which customer segment to prioritise next year	Outside	Depends on strategy, economics and context the model lacks
Reading what a VP's objection in a review really means	Outside	Relationship and history, not text
Final pricing recommendation	Outside	A confident wrong answer is expensive and hard to spot

The inside tasks are where AI for customer feedback analysis and similar workflows pay off. The outside tasks still benefit from AI as preparation (gathering the data, listing the options, steelmanning each one) as long as you make the call.

When should a PM work as a Centaur and when as a Cyborg?

The study found two working styles among the consultants who got both kinds of task right. Mollick describes Centaurs as having "a strategic division of labor, switching between AI and human tasks," and Cyborgs as people who "intertwine their efforts with AI, moving back and forth over the jagged frontier."

Work as a Centaur when the task splits cleanly and at least one part is outside the frontier. A quarterly planning memo is a good example: you decide the bets and the trade-offs, then hand the AI the job of turning your notes into a structured draft and checking it for gaps. The line between your part and its part stays visible.

Work as a Cyborg when the whole task is inside the frontier and you will iterate many times. Writing a PRD section, tightening onboarding copy or shaping a survey are tasks where you might trade a sentence, a critique and a rewrite a dozen times in ten minutes. The risk is drift: after enough back and forth you stop noticing which claims came from you and which from the model, which is exactly how a wrong answer slips through.

A simple rule: the more a wrong answer would cost, the more Centaur you should be.

How do you map your own frontier?

Mollick's advice is blunt: "Just use AI enough for work tasks and you will start to see the shape of the jagged frontier." Turn that into a habit rather than a vague intention. Pick your five most frequent tasks, run each with AI for two weeks, and write one line per run: time saved, errors found, and whether you would have caught them without looking. The tasks where you found errors you would have missed are your outside tasks, whatever the four questions said. The full audit-then-build method is in how to use AI as a product manager.

Repeat the exercise when you switch models. The difference between reasoning and fast models is one of the places the frontier shifts most, and a confident wrong answer (see AI hallucination) is what a task outside the frontier usually looks like.

For a structured practice version of this, the AI productivity practice exercise walks through a PM's week and asks which tasks to hand over. For picking tools once you know your inside tasks, see best AI tools for product managers.

Where can you build a frontier-aware AI workflow?

10x Productivity with AI is a 1 week Builders Camp bootcamp with 2 live sessions, taught by Andre Albuquerque and part of the AI Agentic Builders Expert Track and the Product Delivery Specialist Track. Its published topics include AI workflow design (choosing the right tasks, inputs and tools so productivity gains are real and measurable), research and synthesis, and safe usage habits with quality checks. Its practical challenge puts you in the seat of a PM asked to audit his own workflow in 48 hours and show where AI actually helps and where it does not.

See the 10x Productivity with AI bootcamp