01 · TWO TASKS, ONE AFTERNOON
It is the same afternoon. The same operator, the same machine, two jobs across the desk an hour apart. The first comes back sharper than the work would have been alone, and the day feels lighter for it. The second looks no harder. A memo, a read of a market, a call on a number. It comes back just as quick, just as clean, just as sure of itself. Only this one is wrong. Nothing on the page says so. The voice that got the first one right is the voice that got this one wrong, and between them it never changed its tone.
02 · NOT A FENCE, A COASTLINE
The trouble starts with what we expect. We expect a tool to have a smooth edge. Easy things it does, hard things it fails, and the failures announce themselves. This tool does not work that way. Its edge is jagged. A coastline, not a fence. Two tasks that look equally hard to a person can sit on opposite sides of it. One is inland, well inside what the machine can do. The other is a single step into the water, and there the machine drowns without a splash. From where you stand, the two look identical. The difficulty you can see is simply not the boundary that decides.
03 · SEVEN HUNDRED AND FIFTY-EIGHT
None of this is a fable. In 2023, Harvard Business School and Boston Consulting Group ran a preregistered experiment with seven hundred and fifty-eight consultants. After a baseline, participants entered one of three randomized conditions: no AI, GPT-4, or GPT-4 with a prompt-engineering overview. Across eighteen realistic tasks selected to sit inside the model's frontier, the gains were not small. Those using AI completed twelve point two percent more tasks, worked twenty-five point one percent faster, and produced work rated about forty percent higher in quality. This is the part the headlines carried, and it is true. But the experiment also included a task selected to sit outside the frontier. That is the result worth remembering.
04 · THE PLANTED TASK
The researchers selected one task a single pace past the edge. To the eye, it looked no harder than the rest: reconcile quantitative data against interview evidence and recommend which brand deserved investment. The machine took the bait and walked its users off the cliff. Consultants using AI were nineteen percent less likely to produce a correct solution than consultants without AI. The paper's phrase for the over-reliance is unkind and accurate: falling asleep at the wheel. Fluent output can look finished enough to suppress the checking the task still needs.
05 · THE VOICE HOLDS
Here is the danger in a single line. Capability can fall away while the confident tone holds. Step across the edge, and the machine may sound as certain as it did on a task it handled well. No tremor, no hedge, no tell you can safely treat as a verifier. How sure the output sounds need not reveal which shore you are on. Fluency is not an external check. The frontier is hard to see from the inside because the voice need not bend where the ground gives way.
06 · CENTAURS AND CYBORGS
And yet some people worked that ground more safely. The consultants who did best were not simply the ones who leaned hardest on the machine. The study described two habits among high performers. The centaur draws a clean line down the middle: the person keeps the work they can judge and verify, the machine takes the work it appears to do well, and the seam stays visible. The cyborg blends the two but keeps checking, trading lines back and forth, catching slips as they appear. The evidence is narrower than the metaphor: in this experiment, these were successful patterns for keeping person and model from merging into unchecked automation. What they share is a person who never assumes which side of the edge a task is sitting on.
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 paper gave the shape its name. Dell'Acqua, Mollick and seven others, spread across business schools and the Boston Consulting Group, called it a jagged technological frontier. Listen to the word carrying the weight. Not expanding, though it is. Uneven. The reach of these models grows with every version, but it never grows into a tidy circle you could draw a fence around. It grows the way a coastline grows: in fingers and inlets, with bays of failure reaching far inside ground that looks perfectly solid. That is why no one hands you the map. The edge is real, and it is ragged, and it will not hold still.
09 · WHAT THE STUDY DOES NOT MAP
Keep the evidence boundary visible. The experiment showed that participants using AI were nineteen percent less likely to solve one selected outside-frontier task correctly. It did not demonstrate that this task was out of distribution relative to GPT-4's training or deployment distribution. It also did not isolate a single causal mechanism, or establish that judging the answer was harder than doing the work. Those may be hypotheses for later tests, not findings to import into this result. What the result supports is operational caution: validate the task externally instead of treating fluency as a capability signal.
10 · THE COAST THAT MOVES
One last thing, and it is the cruel one. The frontier moves. A new model can redraw the coastline, often pushing it outward, but sometimes changing the shape in ways that make an old handoff unsafe. What you safely handed the machine in the spring should be retested by autumn, because the same flat, confident voice may hide a new failure. You should not count on a release note for that. So the frontier is not a fact you learn once. It is a thing you re-earn every time the tool changes. Before you hand it over again, ask one operator's question: what external check would catch a confident miss here?