Skip to content
finalstate.aiThe Jagged Frontier: The Edge You Can't See Until You're Past It
113Atlas

113TWO TASKS, ONE AFTERNOON

The Jagged Frontier: The Edge You Can't See Until You're Past It

A jagged frontier is an uneven boundary of model capability. Similar-looking tasks can produce opposite effects from AI assistance.

Two similar-looking illustrative documents; a magnifier reveals an error in one.Task A and Task B have the same visual finish. A source check reveals a miss in B. The scene is illustrative, not an observed pair of experimental outcomes. Task A Task B
Illustrative outputs, not two observed participants. A check reveals the miss.

NOT A FENCE, A COASTLINE

Similar appearance can conceal different capability

A schematic coast puts two equally drawn task markers on opposite sides.A and B share one horizontal comparison line. There is no difficulty axis, scale or measured contour. The authors' metaphor illustrates task-dependent performance; these locations do not represent measured task abilities. AB Schematic InsideOutside
Schematic only. The line does not map a model's measured abilities.

The coast is a metaphor for an uneven boundary. It has no difficulty axis or measured distances: A and B stand for similar-looking tasks whose results differ. The study's own frontier diagram is also explicitly conceptual.

SEVEN HUNDRED AND FIFTY-EIGHT

Selected tasks improved with GPT-4

Human-rated quality on selected product-development tasks: 4.38 without AI, 5.68 with GPT-4 and 5.86 with GPT-4 plus an overview.The three fixed bars show means on the study's 1–10 human rating scale, with a zero chart baseline. Table 2 of the 2026 publication reports n=130, 129 and 126. Time and completion are separate measures stated in the nearby prose.QualityHuman rating, 1–10 100 4.38 Without AI 5.68 GPT-4 5.86 GPT-4+ overview
Dell'Acqua et al., Organization Science (2026), Table 2 and §4.1. Human quality means on a 1–10 scale; n=130, 129 and 126, respectively. These are group comparisons, not a time series.

The June 2023 experiment involved 758 BCG consultants in two separate task arms. Here, 385 worked on 18 product-development tasks, with no AI, GPT-4, or GPT-4 plus a prompt-engineering overview. The 2026 published analysis reports these distinct measures; they are gains tied to a tested workflow.

  • Human-rated quality: about 32% higher across the two AI conditions
  • Time to reach the final question: 25.1% lower on average
  • Tasks completed: 12.2% more on average

THE PLANTED TASK

A separate case produced fewer correct answers

Correct strategy recommendation: 84.5% without AI, 70.6% with GPT-4, 60.0% with GPT-4 plus the overview.The bars share a zero-to-100-percent scale and retain fixed values. These rates concern 373 consultants in one selected outside-frontier task in June 2023. The reported combined gap is about 19 percentage points, not a relative risk of 0.81.Correct recommendationWithout AI 84.5%GPT-4 70.6%GPT-4 + overview 60.0%0100%
Dell'Acqua et al. (2026), Figure 5 and Table 7. One selected strategy case; n=129, 119 and 125. Lower correctness is established; a single cause is not isolated.

Another 373 consultants had to recommend a brand using a spreadsheet and interview evidence. Correctness was 84.5% without AI, 70.6% with GPT-4, and 60.0% with GPT-4 plus the overview. The paper's §4.2 describes the combined gap as about 19 percentage points. That is a gap between percentages, not a 19% relative reduction.

A polished answer can still be wrong.

How sure the output sounds need not reveal which side of the boundary a task occupies. Fluency is not an external check.

CENTAURS AND CYBORGS

Two reported ways of sharing the work

Separate subtask allocation and intertwined work both lead to an unresolved check.Circles denote a person and squares denote AI. The centaur splits a work document; the cyborg alternates contributions within it. A proposed external check ends in a question mark, with no automatic success in either pattern.CentaurCyborg ?Check
Circle: person. Square: AI. The open check is an added requirement, not a success conferred by either workflow.
  • Centaur: divide and delegate subtasks between person and AI
  • Cyborg: integrate AI throughout the work

The research team's 2023 account described these patterns. They were not randomized tests of two safe workflows. Requiring an external check is the prescription here, and its result can remain unresolved. The reverse centaur asks a different question: who controls the plan?

AN UNEVEN EDGE

AI capabilities cover an uneven set of knowledge work

Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” Organization Science, 2026, §1

The published paper describes an uneven boundary. It does not establish that every later release improves every task. A model change calls for another test of the work being delegated.

WHAT THE STUDY DOES NOT MAP

A result is not a diagnosis

A measured result record stops before an open bracket for untested explanations.The gap separates the established task outcome from a single causal account or an out-of-distribution diagnosis. It does not imply those explanations are false, only that this result does not establish them. Result Cause?OOD? Untested
The result record stops short of the untested explanation. OOD means out of distribution; it is a hypothesis here.
  • Demonstrated: lower correctness with AI on one selected outside-frontier task
  • Not demonstrated: that the task was out of distribution for GPT-4
  • Not established by this result: a single cause, or that judging was harder than doing

The authors discuss over-reliance as an explanation. The performance contrast alone cannot settle it. Keep those hypotheses open and choose a cheap external check that can reveal a wrong answer.

THE COAST THAT MOVES

Each release can redraw the coast

A previous task check sits beside a new check whose result remains open after a change.The before record contains an illustrative completed check. The after record requires a retest and ends with a question mark. There are no observed release versions, projected gains or automatic approvals. BeforeAfter ?Retest
Illustrative retest, not observed model releases. The new check remains open.
  • Retest a delegated task when the model or workflow changes
  • Keep responsibility for the decision with someone who can act on a failed check

An earlier success is evidence about the setup that produced it. After a change, ask what external check would catch a confident miss, then return to the value of an honest "I don't know".

Read the transcript

01 · TWO TASKS, ONE AFTERNOON

Two jobs arrive at the same desk on the same afternoon. In this illustrative scene, a person uses one model for both. The first answer helps. The second looks equally finished, but checking its source reveals a mistake. A successful answer to the first job gives the person some evidence about that job. It does not settle whether the model can handle the next one. The surface of the answer can stay familiar while the demands of the task change.

02 · NOT A FENCE, A COASTLINE

The jagged frontier names that unevenness. Tasks that look similarly difficult to a person can receive quite different help from the same model. A coastline gives the idea a shape: two nearby positions may lie on opposite sides. The drawing is a schematic. Its turns and distances are not measurements, and its task markers do not locate any specific model capability. To find out whether assistance helps on a particular piece of work, someone has to test that work. Neither apparent difficulty nor the last successful answer supplies the missing result.

03 · SEVEN HUNDRED AND FIFTY-EIGHT

The BCG experiment took place in June twenty twenty-three. It involved seven hundred and fifty-eight consultants divided between two separate task arms. After a baseline assessment, each arm had three randomized conditions: no AI, GPT-4, or GPT-4 with a prompt-engineering overview. Everyone with AI used the same model version, available at the end of April that year. In the product-development arm, three hundred and eighty-five consultants tackled eighteen selected tasks, including developing a footwear concept. The twenty twenty-six published analysis found the AI groups' average human quality ratings about thirty-two percent higher. Those ratings assessed the responses on a one-to-ten scale; they were not a count of correct answers. Two other measures improved: consultants reached the final question in twenty-five point one percent less time on average, and completed twelve point two percent more tasks. These measures describe selected work under the study conditions.

04 · THE PLANTED TASK

The separate strategy case required a recommendation about which brand had the best growth potential. Participants had to interpret a spreadsheet using details in company interviews. The researchers had selected this case because the model struggled to solve it from the supplied material and instructions. Among three hundred and seventy-three consultants, the share giving a correct recommendation was eighty-four point five percent without AI. With GPT-4 it was seventy point six percent; with GPT-4 and the overview, sixty percent. The published paper describes the combined decrease as about nineteen percentage points. That is an absolute gap in correctness, not a nineteen percent relative reduction. The experiment supports the contrast between these groups on this case. It does not isolate every step that led a participant to accept a wrong recommendation.

05 · THE VOICE HOLDS

A polished answer can still be wrong. A recommendation may read smoothly while depending on a mistaken interpretation of an interview or a number. How sure the output sounds need not reveal which side of the boundary a task occupies. Fluency is not an external check. For the strategy case, the relevant check would return to the spreadsheet and interview evidence, where the recommendation had to be justified.

06 · CENTAURS AND CYBORGS

The research team's twenty twenty-three account described two ways people shared work with AI. A centaur divides the work, assigning different subtasks to the person and the model. A cyborg intertwines the work, moving between personal contributions and AI assistance within a task. These names describe patterns; participants were not randomly assigned to prove that either pattern was safe. An external check is an additional requirement proposed here. It may confirm the answer, reveal a mistake, or leave the handoff unresolved. Dividing or intertwining work does not by itself produce that evidence. Nor does either label answer the question raised by the reverse centaur: who gets to control the plan?

07 · Advertisement · 11x AI Growth Workers

Pipeline is not one task. It is a chain of small handoffs: find the buyer, research the account, respond fast, qualify cleanly, follow up on time. 11x gives sales, marketing, and RevOps teams digital workers for that background motion. The system keeps work moving, while people focus on the conversations that deserve them.

08 · AN UNEVEN EDGE

The published paper describes AI capabilities as covering an uneven set of knowledge work. The experiment gives evidence about selected tasks, one model version and a particular working arrangement. A later release can change those results, including through a regression on a task that previously worked. The case for another test does not depend on assuming that every release expands the frontier.

09 · WHAT THE STUDY DOES NOT MAP

Keep the evidence boundary visible. The result establishes lower correctness with AI on the selected strategy case. It does not establish that the task was out of distribution for GPT-4. That would require evidence about the relevant distribution. The authors discuss over-reliance, but this performance contrast alone does not isolate a single causal mechanism. It also does not establish that judging was harder than doing. Those explanations remain questions for further tests. An external check serves a narrower purpose: catching a wrong recommendation before someone acts on it, even while the full explanation of the failure remains open.

10 · THE COAST THAT MOVES

An earlier successful check belongs to the setup that produced it. When the model changes, retest the delegated task. The result may improve, remain useful, or expose a new problem; the study does not predict which will happen on your work. A release note can help identify changes, but it cannot supply a missing task result. Keep the new check open until there is evidence to close it. Before handing the work over again, ask what external check would catch a confident miss here. Someone still has to decide what to do when that check fails.

01 / 10 · TWO TASKS, ONE AFTERNOON0:00 / 6:38