AI can make you dramatically better at one task and confidently worse at the next.
The idea of a jagged technological frontier is one of the most useful findings in the emerging evidence on AI and knowledge work.
In a preregistered experiment with 758 BCG knowledge workers, AI users working on 18 tasks inside the model's capability frontier, the boundary between tasks AI handles well and tasks it does not completed
12.2 percent more tasks, worked 25.1 percent faster, and produced higher-quality solutions.
But on a selected complex managerial task outside the frontier, AI users were 19 percentage points less likely to reach the correct answer.
That is the whole problem in miniature.
The useful question is not whether AI is good at consulting, law, coding, grant writing, education, or strategy. Those categories are too broad. The useful question is which part of this workflow is inside today's frontier, which part is outside it, and how will I know the difference?
Frontier mapping is therefore a professional skill. Test tasks. Record failure patterns. Compare assisted and unassisted performance. Update the map as models change.
The person who knows where AI helps and where it quietly degrades performance has an advantage over both the enthusiast who trusts everything and the skeptic who refuses the tool.
Frontier awareness is therefore not a warning label attached to AI. It is a repeatable practice: learn where assistance raises performance, where it changes the nature of the task, and where confidence can outrun competence. The boundary itself becomes part of professional knowledge.
The skill is not trusting AI or distrusting AI. The skill is locating the boundary.
One of the strangest features of generative AI is that difficulty does not always behave the way human intuition expects. A model can perform impressively on a sophisticated analytical task and then fail on a neighboring task that appears simpler. The boundary is irregular.
Researchers studying 758 BCG knowledge workers gave this phenomenon a memorable name: the jagged technological frontier. On 18 realistic tasks inside the tested frontier, AI users completed more tasks, worked faster, and produced higher-quality solutions. On a selected task outside the frontier, reliance on AI reduced the likelihood of reaching the correct answer.
The finding destroys two comforting stories at once. The first is the enthusiast's story: AI makes everything better. The second is the skeptic's story: because AI sometimes fails, serious professionals should avoid it. Both positions substitute ideology for task-level evidence.
The practical response is frontier mapping.
Start with a real workflow. Establish an unaided baseline. Then test AI assistance on discrete tasks. Measure speed, quality, error, confidence, and detectability of error. Some failures are cheap because they are obvious. Others are dangerous precisely because they look plausible.
A 2026 lab-in-the-field experiment with 128 knowledge workers reported a similarly task-contingent pattern: generative AI increased efficiency across several knowledge-work tasks, while quality improved for some forms of packaging and creation but declined for knowledge acquisition. The details will vary across models and settings, but the principle is durable: productivity is task-dependent.
The frontier also moves. A workflow tested six months ago may deserve retesting. Models improve, tools gain retrieval and verification features, organizations improve their context, and users become more skilled. A permanent policy based on one snapshot can become obsolete.
This creates a new managerial discipline. Organizations need evidence about where their own frontier lies. Vendor demonstrations are not enough. Generic benchmark scores are not enough. The relevant frontier is the one between your people, your data, your workflow, your risk, and the particular system you deploy.
The individual version is equally important. Know the categories in which you are capable of detecting a bad answer. Know the categories in which you are not. Confidence should be inversely related to your inability to verify.
The jagged frontier turns humility into an operating advantage.
The frontier is also personal. Two people using the same model may face different boundaries because one possesses the expertise to detect a subtle error and the other does not. The tool has the same capability; the human-machine system does not.
That means AI literacy cannot be reduced to feature knowledge. It includes calibration. Where does the system tend to help me? Where does it tempt me to accept plausible nonsense? Which tasks require external verification? Where does my own expertise compensate for the model, and where am I too inexperienced to recognize failure?
Mapping the frontier should be a recurring professional practice. Keep a record of successes and failures. Identify tasks where AI consistently creates leverage, tasks where it requires supervision, and tasks where the cost of error makes another method preferable. The frontier will move. Your map should move with it.