AI is very good at some tasks and poor at others, and the line between them does not follow how hard the task is for a person. A model can solve a maths olympiad problem and fail to read an analogue clock. So where AI fails cannot be judged by intuition: it is mapped with real cases, written down, and revisited from time to time, because the line moves as models improve.
The jagged frontier
The term comes from a study with BCG consultants, described by Ethan Mollick in his book Co-Intelligence. Inside what AI does well, those who used it did more, faster and with higher quality. On a task that looked similar but sat outside, those who used AI made more mistakes than those who did not.
The AI Index 2026 from Stanford University gives recent examples of the same unevenness. A model won gold at the International Mathematical Olympiad, while the best model reads an analogue clock correctly about half the time, against 90% for people. In a hallucination test across 26 leading models, the share of made-up answers ranges from 22% to 94%.
Business documents are no different. In the same report’s finance tests, the 15 best models sit within 3 to 5 points of each other, and none gets above 70% at extracting and calculating values from scanned tax certificates.
Why the frontier is dangerous
The biggest risk lies in people’s trust. When a system is right almost every time, reviewers stop checking carefully, and miss exactly the cases where the system is wrong, which Mollick calls falling asleep at the wheel.
This has two practical consequences. First, “the system is right 95% of the time” says little unless you know which 5% and what each one costs. Second, asking the team to “check everything carefully” does not work for long, so the team needs to know where to look.
How to build the map
1. Test with real cases
For each type of task in the process, gather a sample of real cases with the correct result, including the hard ones: badly scanned documents, rare formats, ambiguous requests. The system runs on the sample, and you measure where it gets things right and where it fails. The shadow mode period is the way to run this test at real volume.
2. Write down three zones
The result is a simple map, by type of task:
- Light review: tasks where the system is consistently right and an error can be caught later at no cost. Reviewed by sampling.
- Review by exception: tasks where the system is right in most cases, but not all. The system gives each result a confidence level, and only cases below a threshold go to a person, with the reason explained.
- Done by people: tasks outside the frontier, or where an error costs too much. The system can prepare the case, but a person decides.
3. Use confidence to choose what to review
The accounting study cited in the AI Index 2026 shows the effect of this design. Accountants served 55% more clients per week with AI, and the ones who gained most were the experienced ones, who used the system’s confidence levels to decide where to review. Review time went to the cases that needed it.
In the systems we build, every result comes with a confidence level, and anything below the threshold goes to a person with the original document alongside and the reason in plain language. The full design of the exception queue is in rule-based tasks and judgment tasks.
4. Revisit the map
The frontier moves. What is outside today may be inside in a few months, with a new model or better rules. The opposite also happens, when documents arrive in a new format or a tax rule changes. That is why we recommend revisiting the map at least once a quarter for high-volume processes, using the errors logged since the last review.
Review time counts too
A system that is right most of the time but forces you to review everything may save nothing. The real gain is the time the process used to take minus the time now spent supervising and correcting. That is why review time goes into the price and into the project’s numbers from the start, as we explain in what a managed process costs.
The review map is one of the four pieces that decide whether the team uses the system, described in why your team isn’t using the AI system.
Frequently asked questions
What is AI’s jagged frontier?
The idea that AI is excellent at some tasks and poor at others, and that the line between them does not follow how hard the task is for a person. Inside the frontier, AI improves the work. Outside it, it makes the work worse, because people trust the answer.
How do I know which tasks I can trust AI with?
By testing with real cases from your process, including the hard ones, and measuring where the system is right and where it fails by type of case. The result is written as a map that separates what gets light review, what gets review by exception and what stays with people.
What is a confidence level?
A value the system gives each result, indicating how sure it is. Results below a threshold go to human review, and the rest flow through untouched.
How often should the map be revisited?
At least once a quarter for high-volume processes, and whenever the model, the document format or a business rule changes.
Want to know where your process sits against the frontier?
Book a call and we will test a sample of real cases with you before any proposal.
