“My general advice for using AI for any purpose is to ask yourself: ‘How do I make what I am doing even more ambitious?’ You do this over and over until you either hit a wall and thus learn the shape of the jagged frontier (useful!) or you find out that the AI can do more than you think (common!)” – Ethan Mollick – Associate Professor at The Wharton School of the University of Pennsylvania
The practical value of ambitious AI use lies in testing the boundary between what a system can plausibly extend and what it can quietly distort. That boundary is not a clean line separating easy work from difficult work. It shifts across tasks, domains and forms of judgement, which is why deliberate escalation can be more informative than cautious experimentation. The original advice places that escalation at the centre of the user’s method: ask how the work could become more ambitious, then test the answer until failure reveals a limit or success reveals unused capacity 1.
This approach changes the role of prompting. Instead of treating an AI system as a tool for polishing an already defined assignment, the user treats it as a partner in discovering the larger problem. A routine market summary might become a scenario analysis; a draft lesson might become a complete learning sequence; a rough product idea might become a testable launch plan. The point is not to inflate every task. It is to expose capabilities that remain invisible when the request is narrowly specified.
The jagged frontier is a problem of task design
Research associated with Ethan Mollick gives this principle an empirical foundation. In a preregistered experiment involving 758 consultants, participants with GPT-4 access completed 12,2% more tasks, worked 25,1% faster and produced higher-quality results on 18 realistic consulting assignments 2,3. Those gains did not mean that the system was uniformly reliable. On a complex managerial task placed outside its capability boundary, AI-assisted participants were 19 percentage points less likely to produce a correct solution than participants without AI access 2,8.
That contrast explains the jagged frontier. Tasks that look similar to a human observer can sit on opposite sides of an AI system’s effective capabilities. Idea generation may be unusually strong, while apparently basic calculations or nuanced managerial judgement may fail in unexpected ways 8. Difficulty, therefore, is not enough to predict suitability. Users need to learn which components the system can handle, which require verification and which should remain under direct human control.
Repeated ambition is useful because it creates a form of capability mapping. A first request establishes a baseline. A more demanding request tests whether the system can generalise. A further request may expose a weakness in factual accuracy, reasoning, context retention or execution. The resulting failure is not merely a disappointing output. It identifies a boundary that can inform future delegation. Conversely, a strong response may reveal that the original workflow was constrained by habit rather than by genuine necessity.
From prompting to organisational learning
The larger strategic implication is that AI adoption should be organised around experiments rather than declarations. Organisations often ask whether a model can perform a job, but jobs are bundles of tasks with different risk profiles. A better question is which parts of a workflow benefit from assistance, which parts demand expert review and how responsibility should move between human and machine. The jagged-frontier research supports this task-level approach because benefits and harms can occur within the same broad category of knowledge work 4,9.
Ambition also matters because conservative use can conceal the most valuable applications. If employees use AI only to shorten emails or summarise meetings, leaders may conclude that the technology offers modest efficiency gains. More ambitious trials might reveal support for research synthesis, option generation, process redesign or cross-functional coordination. A later field experiment with 791 professionals found that AI-assisted individuals matched the performance of unaided teams on certain product-innovation tasks, while also helping bridge differences between technical and commercial perspectives 7. Such findings suggest that the important unit of change may be the structure of collaboration, not simply the speed of an existing task.
Yet escalation without controls can turn curiosity into operational risk. A system that produces an impressive strategy may still invent evidence, miss a legal constraint or recommend an action that fails under local conditions. The correct response is not to avoid ambitious prompts, but to pair them with explicit verification. Users can request assumptions, alternatives, source trails, uncertainty labels and tests that would falsify the recommendation. They can separate ideation from approval, draft from publication and analysis from decision-making. Ambition expands the search space; governance determines what is allowed to leave it.
Why the method remains contested
Several objections deserve attention. First, the frontier moves. A limitation observed in one model version may disappear in another, while a new strength may create fresh risks. Recent discussion suggests that some early examples of obvious AI incompetence are becoming less useful as systems improve, leaving more technical and domain-specific edge cases 6. This makes fixed rules fragile. Capability assessments need dates, model names, task definitions and evaluation criteria rather than vague claims about what AI can or cannot do.
Second, productivity gains can be unevenly distributed. Faster completion does not automatically mean better judgement, deeper expertise or greater organisational resilience. If employees rely on plausible outputs without developing the ability to inspect them, apparent efficiency may weaken the skills required for supervision. There is also a risk that ambitious experimentation benefits workers who already possess strong domain knowledge while exposing less experienced users to confident errors. Training should therefore cover not only prompting, but also evaluation, provenance, confidentiality and escalation routes.
Third, the ambition principle can be misunderstood as a demand to pursue maximum automation. Its stronger interpretation is more disciplined: keep increasing the value and scope of the question until the system’s contribution becomes legible. Sometimes that process ends with a sophisticated AI-assisted workflow. Sometimes it ends with a clear refusal to delegate a judgement. Both outcomes are useful because both improve the map of responsibility.
A durable practice for AI work
A practical cycle follows from this analysis. Start with a meaningful objective rather than a request for generic output. Ask the system to propose a more ambitious version of the objective, then require it to identify assumptions, risks and evidence needs. Break the expanded task into components and test each one separately. Compare the results with a human baseline. Record where the system adds quality or speed, where it introduces errors and what review is required. Finally, repeat the exercise when the model, data, workflow or stakes change.
The value of the method is ultimately epistemic. It turns uncertainty about AI into an evidence-gathering process. Success shows where capacity has been underused; failure shows where confidence should stop. Neither outcome justifies blind optimism or blanket scepticism. The most capable users are those who make the assignment ambitious enough to discover new possibilities and the evaluation rigorous enough to detect when the machine has crossed its boundary.
References
1. https://lnkd.in/p/dqtHxGmJ – https://lnkd.in/p/dqtHxGmJ
2. dell-acqua-et-al-2026-navigating-the-jagged-technological … – https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf
3. Navigating the Jagged Technological Frontier: Field Experimental … – https://doi.org/10.2139/ssrn.4573321
4. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality – https://econpapers.repec.org/article/inmororsc/v_3a37_3ay_3a2026_3ai_3a2_3ap_3a403-423.htm
5. Ethan Mollick’s Post – LinkedIn – 2026-03-13 – https://www.linkedin.com/posts/emollick_two-and-a-half-years-ago-we-put-out-the-working-activity-7438342548135411712-Jxr0
6. The jagged frontier of generative AI: A conversation with … – 2026-03-04 – https://www.insightpartners.com/ideas/generative-ai-ethan-mollick/
7. A Field Experiment on Generative AI and Teamwork – 2026-06-12 – https://dl.acm.org/doi/abs/10.1287/orsc.2025.20702
8. Navigating the Jagged Technological Frontier – https://mitsloan.mit.edu/sites/default/files/2023-10/SSRN-id4573321.pdf
9. What is the jagged AI frontier? – MIT Sloan – 2026-02-03 – https://mitsloan.mit.edu/ideas-made-to-matter/working-definitions/what-is-jagged-ai-frontier
10. Centaurs and Cyborgs on the Jagged Frontier – 2023-09-16 – https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged
11. Generative AI and Productivity in the Global South – 2026-07-13 – https://www.socialscienceregistry.org/trials/18778
