Ask any AI to explain a framework and it will do a beautiful job. MEDDIC, Shape Up, Getting Things Done, the Eisenhower matrix — name it, and you'll get a clean, correct, well-structured summary in seconds. The information is free now. Genuinely free, and genuinely good.
So here's the question that actually matters: if every framework ever published is now a free instant explanation away, why does the work feel exactly as hard as it always did?
Because explanation was never the bottleneck. Judgment was. And judgment is a completely different thing — one that didn't come free with the model, and won't.
Information became free. Judgment is what the work actually requires.
The cleanest evidence I've seen
I'm wary of AI studies — most are either hype or doom. But there's one I keep coming back to, because it doesn't tell a tidy story. It tells the real one.
Researchers at Harvard Business School and Boston Consulting Group put 758 management consultants in front of GPT-4 and gave them real tasks. On the tasks that sat inside the AI's wheelhouse — synthesis, analysis, drafting, explanation — the people with AI were dramatically better.
That's a serious lift, and it's the number everyone quotes. But it's the second finding that should change how you think. On a different set of tasks — ones that needed judgment rather than recall — the consultants using AI were 19 percentage points less likely to get the right answer than the ones working without it.
Read that again. On the tasks that needed judgment, the AI didn't just fail to help. It actively made good professionals worse — confidently, plausibly worse. The researchers called the dividing line the "jagged technological frontier." Some tasks that look hard are easy for AI. Some that look easy sit beyond it. And the thing that decides which side a task falls on isn't difficulty. It's judgment.
That study published in 2023. The models have improved enormously since. The frontier has moved — but it hasn't changed shape. Bigger models know more. They still don't know which things matter in your specific, messy situation. That's not a knowledge problem you can train away. It's a different category of thing.
So what is judgment, actually?
It's a slippery word, so let me make it concrete. When I've pulled apart what an expert does that a brilliant generalist can't, it comes down to four things. Not vibes — four specific, nameable capabilities.
Calibrated thresholds. An expert knows what "good enough" actually looks like — not in theory, but the real bar, set by having seen a thousand examples. "A 7 out of 10 here is failing." A generalist gives you a reasonable-sounding standard. An expert gives you the standard.
Decision trees. The expert asks specific questions in a specific order, because the order is part of the method. First check X; if X, then look at Y; only then Z. That sequence is how they avoid solving the wrong problem. It rarely survives into a summary, because it lives in how they work, not in what they wrote.
Prescribed formats. Not "here are some ways to present this." The one way the expert has found actually lands. The exact shape of the output. Generalists offer options. Experts prescribe.
Pattern recognition. The thing they notice that you walked straight past — "every one of these shapes fails the same requirement, so your problem is upstream." That's not in the book as a rule. It's the residue of years of doing the work.
Base AI has read about all four. It can describe a threshold, recite a decision tree, name a format. What it can't do is apply them the way the person who built them does. The distance between describing and applying is the entire gap — and it's exactly the gap the Harvard study measured.
Why bigger models don't close it
The natural hope is that this is temporary — that the next model, or the one after, finally crosses the frontier. I don't think it does, and the reason is structural, not pessimistic.
A larger model is trained on more of what's written down. But judgment is mostly the part that never got written down. The calibration an expert can't quite articulate. The pattern they recognise without consciously checking. The sequence they follow on instinct. You can scrape every book on Shape Up ever published and still not capture the thing Ryan Singer does when he looks at a stuck bet and says "the requirements are the problem, not the design." That move isn't in the text. It's in the practitioner.
So the failure mode isn't "AI is bad." AI is an extraordinary generalist. The failure mode is asking a generalist to do specialist work and being surprised when you get generalist output — competent, plausible, and subtly wrong in exactly the way that costs you.
What actually closes it
There are two answers on offer. One is to wait for the model to get good enough. The Harvard evidence says don't hold your breath — and worse, that while you wait, the confident-but-wrong failures are the dangerous kind, because they don't look like failures.
The other is to give the AI the judgment it's missing. Not by writing longer prompts or pasting in more context — that just feeds the generalist more to average over. By extracting the actual method from the actual expert — their thresholds, their sequences, their formats, their pattern recognition — and packaging it so the model applies it faithfully instead of approximating it.
That's what a skill is, and I've watched the difference it makes show up in testing. The same methodology, properly extracted, moved a real product-strategy task from a 26.6 to a 31.4 out of 40 against base Claude — an 18% lift, concentrated precisely in the judgment dimensions. (How that plays out on real problems is its own piece — The End of Generic AI Advice walks three of them.)
The headline is simple, and it's the thing I'd want every knowledge worker to internalise. The information ceiling has already been hit; everyone has the same free explanations now. The next advantage isn't a better model. It's better methodology — the judgment that turns a correct answer into a useful one.
Cheers, Adam