The jagged frontier: why AI helps on one task and hurts on the next

In 2023, researchers working with Boston Consulting Group ran one of the largest controlled experiments yet on knowledge work and AI. Seven hundred and fifty-eight consultants were randomly assigned to work with or without GPT-4 on realistic business tasks.

On tasks the model handled well, the result was emphatic. Consultants with AI completed about 12% more tasks, around 25% faster, and their output was graded roughly 40% higher in quality. The weakest performers improved the most — the gap between the top and bottom of the cohort narrowed sharply.

Then the researchers introduced a task deliberately designed to sit just beyond the model's competence: one where the obvious-looking answer, the one the data invited, was wrong. On that task, consultants using AI were 19 percentage points less likely to reach the correct solution than those working without it.

Nothing about the second task announced itself as harder. That is the whole problem.

The frontier is jagged, and it is not shaped like human difficulty

We are used to capability boundaries that behave sensibly. A junior analyst can do simple things and not complex ones; as they learn, the boundary expands outwards more or less evenly. We estimate what someone can handle by asking how hard the task feels to us.

That intuition fails completely with machine capability. A language model's competence was shaped by what it was trained on and how it was built, not by anything resembling human cognitive load. So the boundary is not a smooth circle expanding outwards. It is jagged: deep salients where the model is superhuman sit directly beside notches where it is confidently, fluently wrong.

Drafting a segmentation from messy interview notes may be inside. Reconciling two numbers whose discrepancy only matters if you know how the business books revenue may be outside. Both take a competent human about the same effort. Nothing on the surface distinguishes them.

The question is never "can AI do this?" It is "can we tell, cheaply, when it got this wrong?"

Why people do not notice they have crossed it

If the failure were obvious, the cost would be small. It is not, for three reasons that show up repeatedly in the research.

Fluency is not accuracy, and we read it as accuracy. The output is well-organised, appropriately hedged and written in the register of an expert. Every surface cue we normally use to judge reliability points the wrong way.

Self-assessment is unreliable in exactly the wrong direction. In a 2025 study, METR recruited sixteen experienced open-source developers and had them complete 246 real tasks on repositories they already knew well, with and without AI assistance. The developers expected AI to make them faster. Afterwards, they reported that it had — by about 20%. The measured result was the opposite: they were 19% slower with it. If experienced engineers cannot judge the effect on their own work while doing it, an adoption survey certainly cannot.

Combining human and machine does not automatically beat either one. A 2024 meta-analysis in Nature Human Behaviour pooled 106 experiments on human–AI collaboration. On average, the combination performed worse than the better of the human alone or the AI alone. The pattern underneath was instructive: combinations tended to help on creation tasks and to hurt on decision tasks, and they helped most when the human was already better than the AI at the job. Synergy is not a property of putting the two together. It is a property of designing the handoff.

How to map your own frontier

The frontier cannot be reasoned about from first principles, because it is an empirical fact about a specific model doing a specific task on your specific data. It has to be measured. In practice this takes about two weeks per workflow.

1. Break the work into tasks small enough to test. "Handle a claim" is not testable. "Extract the policy number and incident date from the submitted document" is.
2. Assemble a held-out set of real cases with known answers. Thirty to fifty per task, drawn from work already completed, including the awkward ones your team remembers. Synthetic examples will flatter the system.
3. Score accuracy and the cost of detection separately. A task where the model is right 80% of the time and a human can spot the other 20% in seconds is a good candidate. A task where it is right 95% of the time and catching the remaining 5% requires re-doing the work is not. This single distinction reorders most roadmaps.
4. Record where the failures cluster. Random errors are tolerable. Systematic errors — always the same edge case, always the same kind of reasoning — define the notch you must design around.
5. Re-run it quarterly. The frontier moves with each model release, usually outwards but not uniformly. A map from last year is a historical document.

What this changes about the roadmap

Once you have a map, three decisions get easier and one gets harder.

Easier: what to build first — the deep salients where the machine is strong and verification is cheap. Where to place humans — not "in the loop" as a blanket policy, which is expensive and quickly becomes rubber-stamping, but at the specific checkpoints where the notches are. What to measure — outcomes against your recorded baseline, never adoption or self-reported time saved.

Harder: telling a sponsor that the workflow they were most excited about sits in a notch, and that the honest recommendation is not to automate it this year. That conversation is uncomfortable, and it is the most valuable thing an advisor can do. The BCG study's outside-the-frontier result is what happens when nobody has it.

The organizational version of the same lesson

There is a reason the gains were largest for the weakest performers in both the BCG study and Brynjolfsson, Li and Raymond's work on customer support agents, where novices gained around 34% and the most experienced gained close to nothing. Augmentation distributes existing expertise before it creates any new expertise.

That is genuinely valuable — it is the fastest route to raising the floor across a large organization. But it also means the value of your senior people goes up, not down: they are the ones who can recognise a notch when they are standing in one. An augmentation programme that quietly erodes that judgment has spent its gains and borrowed against the future.

Which is the whole argument for treating this as system design rather than tool adoption. The model is one component. The map, the checkpoints, the measurement and the people who can read them are the rest of it.