The Wicked Frontier
Suppose a city is considering a new bridge. Once engineers know the span, expected traffic, budget, and safety standards, they can test possible designs against those constraints. The work may be extraordinarily difficult. It may require new materials, complex simulations, and years of engineering. But progress is legible. A design carries the required load or it does not. It meets the safety standard or it does not. Its cost can be estimated and compared with the alternatives.
A second set of consequential questions exist in parallel. Should the bridge be built at all? Where should it land? Which journeys should it shorten? What happens to the neighborhoods it changes? Who should pay, and whose idea of improvement counts? Analysis can inform those choices, but it cannot make them objective. Each possible answer gives different weight to different interests, trade-offs, and preferences.
In 1973, planning theorists Horst Rittel and Melvin Webber gave this second set a name: wicked problems. Wicked did not mean negative, or even exceptionally complicated - but a problem which resists formal resolution. There is no single, definitive account of the problem, because different people are experiencing different parts of it. There is no agreed stopping rule. Its possible answers are not simply true or false, but better or worse depending on what matters, who is judging, and who will live with the result. The opposite of wicked is (poetically), tame, where the result of a problem is knowable and objective.
This distinction helps explain something important about AI: it is not advancing through both kinds of problem at the same rate. It moves fastest through tame problems, where an answer can be checked and every attempt provides feedback for the next.
A Frontier With Two Fronts A proof can be verified. Code can be run against a test suite. A structural design can be checked against stated limits. These problems may be ferociously hard, but their difficulty has a shape. An answer can be checked without reopening the goal each time. Where answers are cheap to check and new problems cheap to create, progress can compound quickly. I call this the tame frontier. Tasks related to software, engineering, physics, and materials science sit inside it, especially where strong simulations or automated experiments exist.
However, most knowledge work does not have that shape. Knowledge work is wicked work. The wicked frontier begins where the target cannot be inferred from the problem itself: which tradeoff matters here, what a good outcome means in this situation, and who has the standing to decide. A model can help frame the choice, surface the tradeoffs, or even make the call when authorized. What today's AI capabilities cannot establish is standing — the right to make the call and the responsibility to own what follows: whose interests should prevail, who had the authority to define the objective, who remains accountable afterward.
That gap doesn't close as models get more capable. It relocates, to wherever the next consequential definition hasn't been made yet.
The Sweet Spot No real job sits perfectly inside these two extremes. A single week moves through work that's genuinely tame and worth automating outright, some that's wicked all the way down, where the judgment must remain visible and someone must own it, and some that's wicked-looking but tame underneath and worth decomposing. For AI, that's often the sweet spot.
Take a request that looks wicked at first glance: prepare a view on whether to enter a new market. On its face, this is loaded with judgment — appetite for risk, competitive response, what success even means a year out. But look at what the request actually requires to get moving, and much of it decomposes into pieces that are individually plain.
Pull the relevant market data. Model a few pricing scenarios. Summarize how three comparable companies handled a similar entry. Draft two or three framings of the decision so a leader can react to something concrete instead of a blank page. None of those steps require anyone to have already decided what the company should do. They require competent execution against a target that becomes clear the moment it's isolated from the larger question sitting on top of it — and that is exactly what capable AI, at the frontier and increasingly well behind it, is good at, and is getting cheaper at doing every quarter.
To see how large this category might be, think through an ordinary week of knowledge work: research and synthesis, drafting and redrafting, checking numbers, coordinating logistics, testing ideas, and preparing materials for someone else's decision. Set beside that the moments where the outcome depends on which interests are weighed, what precedent is being set, or what the organization is prepared to live with if it is wrong. The exact balance differs by role, but it is rarely a sliver. It could be large enough to change the economics of most jobs.
Adoptioneering This is why treating "AI adoption" as one uniform push isn't precise enough to actually work. The tame zone wants automation. The sweet spot wants decomposition and the tooling to do it well. And the wicked zone wants something different from either: people empowered to keep making the calls nothing upstream can make for them, and increasingly empowered as the other two zones improve and free up the time and attention to do it properly. This isn't a story about which jobs are at risk. It's a story about where different kinds of effort will pay off, and which decisions will remain ours to own — a direction of travel worth planning around now, rather than a threat to react to later.
That leaves one practical question: how does the sweet spot actually get found, in a specific team, on a specific piece of work, rather than just asserted to exist? The answer is smaller and more repeatable than it sounds. Take a request that arrives wicked-looking. Find the substeps inside it that are actually tame — the parts with a clear, checkable target once isolated. Do those with AI. Keep the residue: the part that still needs a person to decide something. That act of separating — this part can just be done, this part still needs a call — is itself a small piece of authorship, and it happens at the size of a single task, not once at the top of a strategy document. It isn't a single decision made by a leader and rolled out. It happens constantly, at every size of task, performed by whoever is closest to the work. And whatever doesn't decompose, no matter how far down the attempt is pushed, doesn't disappear. It rises to whoever is holding the next layer of judgment, which is exactly where it belongs.
The tame frontier will keep advancing, turning more work into infrastructure. The sweet spot will widen with it, creating real acceleration for anyone willing to decompose the work. The wicked frontier will not move simply because capability does. It will come into sharper focus as the tame execution around it recedes into the background.
Focusing on building those decomposition skills will be rewarded over and over, even as the capabilities of AI change.