𝗧𝗵𝗲 𝗜𝗻𝘃𝗶𝘀𝗶𝗯𝗹𝗲 𝗪𝗮𝗹𝗹: 𝗧𝗵𝗲 𝗠𝗼𝘀𝘁 𝗘𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗠𝗶𝘀𝘁𝗮𝗸𝗲 𝗪𝗲 𝗠𝗮𝗸𝗲 𝗪𝗵𝗲𝗻 𝗔𝗱𝗼𝗽𝘁
Sometimes the best way to understand a new problem is to observe how we get it wrong when we try to solve it with old tools.
A few weeks ago, I was sitting with the management committee of an industrial company. They had spent three months trying to integrate artificial intelligence into their business processes. They had bought licences, run basic training sessions, and asked their teams to start using it. The verdict was unanimous and disappointing: the tool did not do what they needed.
"We asked it to cross the quarterly sales data with current inventory to predict stock-outs," the operations director told me, visibly frustrated. "It makes up numbers. It produces calculations that do not add up. We cannot trust it. And yet the marketing team says it writes their campaign emails perfectly in five seconds. It is a nice toy for drafting, but it does not work for serious business."
As I listened to him, I realised he was evaluating artificial intelligence using the same mental framework he had been applying to traditional software for twenty years. When you buy an ERP or a CRM, you assume its capabilities are linear, predictable, and documented. If the system can process one invoice, you assume it can process a thousand. If it can calculate VAT, you assume it can calculate income tax. Traditional software has clear limits: it does what it was programmed to do, and it fails predictably when you ask it to do something it was not designed for.
But generative artificial intelligence is not traditional software. It is not programmed with explicit rules; it is trained on probabilistic patterns. And that technical difference produces a practical effect that throws most executives off balance: its capabilities are not linear. They are, in the words of Wharton professor Ethan Mollick, a "jagged frontier."
This Saturday morning, free from the urgency of the weekly agenda, I would like to pause on that concept. Because understanding the jagged frontier is not a theoretical exercise. It is, arguably, the most valuable idea an executive can take away today about how to govern AI adoption in their organisation.
𝗧𝗵𝗲 𝗠𝗮𝗽 𝗧𝗵𝗮𝘁 𝗗𝗼𝗲𝘀 𝗡𝗼𝘁 𝗘𝘅𝗶𝘀𝘁
In his book Co-Intelligence: Living and Working with AI (2024), Ethan Mollick describes the jagged frontier as the invisible, unpredictable line that separates what artificial intelligence can do extraordinarily well from what it does disastrously badly.
Imagine the wall of an ancient castle seen from above. It is not a straight line. It has towers that jut outward, sections that retreat inward, asymmetric bastions, and deep recesses.
Everything inside the wall represents tasks that AI can perform at a level equal to or better than a human expert. Everything outside the wall represents tasks where AI fails, hallucinates, or produces mediocre results.
The problem — and this is where the brilliance of Mollick's metaphor lies — is that the wall is invisible. Worse still: its shape runs completely counter to our human intuition about what is hard and what is easy.
For a human, writing a love sonnet in perfect Shakespearean metre is a cognitively demanding task, while counting the number of words in that same sonnet is trivial. For a large language model, the frontier is exactly the reverse. Writing the sonnet sits deep inside the wall (it does it in three seconds with astonishing quality). Counting the words sits outside the wall (it will get it wrong almost every time, because it does not "read" words — it predicts tokens).
This cognitive asymmetry is what frustrated the operations director in my meeting. In his human mind, merging two Excel tables is a mechanical, "easy" task, while drafting a persuasive sales email is a creative, "difficult" one. Therefore, if the machine can do the hard thing, it should be able to do the easy thing. But AI does not think. It predicts. And the jagged frontier dictates that generating persuasive language sits within its native capabilities, while exact mathematical calculation without external tools does not.
𝗙𝗮𝗹𝗹𝗶𝗻𝗴 𝗔𝘀𝗹𝗲𝗲𝗽 𝗮𝘁 𝘁𝗵𝗲 𝗪𝗵𝗲𝗲𝗹
The danger of not seeing the jagged frontier is not merely frustration. The real danger is what Mollick calls "falling asleep at the wheel."
In one of the most revealing empirical studies cited in the book, Mollick and a team of researchers from Harvard and MIT worked with nearly 800 consultants at Boston Consulting Group (BCG). They split the consultants into two groups: some used AI (GPT-4) and some did not. They were assigned real consulting tasks.
For tasks that fell inside the jagged frontier — ideation, drafting, qualitative analysis — the consultants who used AI were significantly faster and produced work that was rated as higher quality, more creative, and better written than that of their colleagues without AI.
But the researchers set a trap. They designed a task that appeared to fall within AI's capabilities, but which in reality required a subtle statistical judgement that language models typically fail at. It was a task deliberately placed outside the jagged frontier.
The results were striking. Consultants who did not use AI got the correct answer 84% of the time. Consultants who used AI — the same AI that had made them look like geniuses on the previous tasks — only got it right between 60% and 70% of the time.
What had happened? The AI was so good at so many things that the consultants had lowered their guard. They stopped applying their own critical judgement. They assumed that if the machine had been brilliant drafting the previous report, it would also be infallible calculating profitability on this one. They fell asleep at the wheel precisely when the road crossed the invisible frontier.
This phenomenon, documented also in studies with HR recruiters, is the nightmare of any management committee. They do not fear that AI will not work. They fear that it will work so well that their teams stop thinking, delegate their professional judgement to a probabilistic algorithm, and the company ends up treating undetectable hallucinations as fact.
𝗘𝘅𝗽𝗹𝗼𝗿𝗲𝗿𝘀 𝗮𝘁 𝘁𝗵𝗲 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿
If the frontier is invisible and counterintuitive, how can we govern AI adoption in our organisations? How do we prevent our teams from crashing into the wall or falling asleep at the wheel?
The instinctive response of many organisations is control through restriction. If we do not know what it does well and what it does badly, we limit its use to hyper-controlled pilot cases, draft restrictive usage policies, and wait for the market to mature or for vendors to publish clear instruction manuals.
But Mollick argues — and my experience in the corporate trenches confirms — that this is the worst possible strategy. There will be no instruction manual. Model capabilities change every few months, shifting the wall and altering the frontier. What was outside yesterday is inside today.
The only way to discover the shape of the jagged frontier is to collide with it. And the only way to collide with it without wrecking the car is to foster a culture of systematic, safe, and distributed experimentation.
Mollick proposes an operating principle he calls "Always invite AI to the table." It does not mean delegating the work to it. It means asking your teams, for every task they do this week, to open a window with their corporate AI model and ask it to do the task in parallel.
If you are writing a report, ask AI to write a draft. If you are reviewing a contract, ask AI to find the penalty clauses. If you are designing a pricing strategy, ask AI to act as an aggressive competitor and try to dismantle it.
Most of the time, the result will not be perfect. Sometimes it will be useless. But in the process of trying to use it for everything, your teams will begin to map the invisible wall. They will discover, through pure trial and error, which recesses of the frontier apply to their specific work.
They will learn that AI is a disaster at summing Excel columns (outside the frontier), but extraordinary at writing a Python script that sums those same columns without error (inside the frontier). They will learn that it is useless for making ethical decisions about who to hire (outside the frontier), but brilliant at preparing specific interview questions based on a candidate's CV (inside the frontier).
That tacit knowledge — that intuition about where the machine's limits lie in the specific context of your business — cannot be bought from a consultancy or downloaded as a PDF. It has to be built internally.
𝗧𝗵𝗲 𝗡𝗼𝗻-𝗗𝗲𝗹𝗲𝗴𝗮𝗯𝗹𝗲 𝗥𝗲𝘀𝗽𝗼𝗻𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆 𝗼𝗳 𝗛𝘂𝗺𝗮𝗻 𝗝𝘂𝗱𝗴𝗲𝗺𝗲𝗻𝘁
Throughout Co-Intelligence, Mollick insists on an idea that is often lost in the technological noise: artificial intelligence is not an independent mind — it is a "co-intelligence." It feeds on our collective knowledge, reflects our biases, and, fundamentally, requires our direction.
When executives assume that AI is simply faster software, they try to manage it the way they manage software: seeking efficiency, cost reduction, and pure automation. They look to replace humans.
But when you understand the jagged frontier, you realise that pure automation is a dangerous game. Precisely because AI can do so many things well, we need expert humans more than ever — humans who know when the machine is wrong. We need what the industry calls "humans in the loop."
We do not need humans to type faster. We need humans to apply the critical judgement the machine does not have. We need experts who, when reading an AI-generated profitability analysis, feel that professional unease in their stomach that tells them: "This number does not make sense. The machine has crossed the frontier."
The real value of AI is not that it allows us to stop thinking. It is that it forces us to think at a higher level. It frees us from the mechanics of drafting the first version so that it can demand from us the responsibility of editing it, validating it, and owning the consequences of publishing it.
𝗔 𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗪𝗲𝗲𝗸
AI adoption in a company is not a technology project. It is an organisational learning project. It is the process of sending your teams to explore new territory, without a map, so that they can discover where the limits of the possible lie.
The operations director I was meeting with did not have a technology problem. He had an expectations problem. He expected AI to be a deterministic calculation engine, and he found a probabilistic reasoning engine instead. He was asking a poet to do his accounting, and getting angry when the numbers did not add up.
When he understood the concept of the jagged frontier, his frustration disappeared. Not because the tool had improved, but because his mental map had adjusted to reality. The following week, his team stopped asking AI to cross inventory data directly, and started asking it to write the SQL code that would allow the traditional system to cross that data. They had found the door in the wall.
The artificial intelligence you have in your company today is the worst artificial intelligence you will ever use for the rest of your life. The wall will move. The frontier will expand. But the principle will remain the same: only those who explore the limits will know how to use them to their advantage.
As you finish your coffee this morning, I leave you with a question to carry into the week ahead:
In your team, who is actively mapping the jagged frontier of your business — and what incentives do they have to share what they discover before they fall asleep at the wheel?
Saturday reading: "Co-Intelligence: Living and Working with AI" (2024), by Ethan Mollick