The Future of Work Is Finding New Problems to Solve

I believe this more than I believe almost anything else about the next decade of work, so I want to lay the argument out carefully, including the parts that are uncomfortable for someone who runs an automation business to say.

Here's the uncomfortable part first: most of what knowledge workers do all day is now automatable. Not eventually. Now. I know because automating it is my job, and the machines I point at execution work keep winning. The reports write themselves. The monitoring monitors itself. The follow-ups follow up. If your value at work is "I complete tasks that are handed to me, accurately," you are competing with software that completes them faster, at three in the morning, for cents.

That sentence has been written a thousand times as a threat. I think it's the wrong reading. Because there's a step before every task that the software cannot do, and almost nobody is trained to do it either.

Every solved problem was found first

Problems don't arrive pre-defined. Someone formulates them first, and the formulation is where the value concentrates. The classic Getzels and Csikszentmihalyi study of art students found that problem-finding behavior, not technical rendering skill, was what predicted long-term creative success years later.

We talk about problem-solving as if problems arrive pre-packaged: defined, scoped, waiting in a queue. School trains this explicitly. Twelve-plus years of "here is the problem, find the answer," where the problem is given and the answer already exists in the back of the book. Most jobs are organized the same way: the ticket, the brief, the backlog. Someone upstream decided what the problem was. Your job starts after that decision.

But the deciding was the valuable part. In the 1960s, researchers Jacob Getzels and Mihaly Csikszentmihalyi ran a now-classic study of art students at the School of the Art Institute of Chicago. They watched how students approached a still-life drawing task. One group treated it as a given problem: arrange objects, draw them well. The other group spent their time differently: handling the objects, rearranging, exploring, in effect discovering what the drawing was about before committing. Years later, the second group (the problem-finders, as the researchers called them) had dramatically more successful creative careers. The skill that predicted long-term creative success wasn't rendering. It was formulating.

The same hierarchy shows up everywhere once you look. Einstein said the formulation of a problem is often more essential than its solution. Every significant company is a found problem: someone noticed friction the world had normalized so completely that it didn't register as a problem at all. People didn't walk around saying "I wish I could rent a stranger's spare bedroom." The problem (trust between strangers at scale) had to be found before it could be solved.

What machines actually can't do

AI out-executes professionals on well-posed problems, but it cannot notice that an unnamed problem exists. Unfound problems are absent from training data by definition, and choosing the objective sits outside the optimization loop. The blind spot is structural, not a temporary capability gap.

I want to be precise here, because "AI can't be creative" is a comfort blanket and mostly false. The models are startlingly good at solving well-posed problems and remixing within a frame. Give a frontier model a defined problem and it will often out-execute a competent professional.

What the machine cannot do is care that something is wrong. Problem-finding starts with an itch: this process feels heavier than it should be; this number doesn't smell right; everyone in this industry does X and nobody remembers why. That itch is not in the data, because the defining feature of an unfound problem is that nobody has written about it yet. A model trained on the world's text inherits the world's blind spots. The unnamed problem is, by definition, in the blind spot.

There's a structural version of this argument too. An optimization process needs an objective function. Execution is optimization. But choosing the objective (deciding what's worth optimizing, noticing the objective everyone has is subtly wrong) sits outside the loop. McKinsey's 2026 enterprise research keeps finding the same thing from the other direction: 77% of AI pilots fail to scale, and the dominant cause cited is not model capability but inadequate direction. Companies have rented infinite solving capacity and discovered they're starving for problem definition. The constraint moved up the stack.

The execution cliff, and what's left standing on it

As AI collapses the cost of executing defined work, economic value migrates upstream to problem definition. When solving trends toward free, the premium moves to knowing which work is worth doing, and to the people who notice the problems that never made it onto a ticket.

Here's the economic shape of it. Execution cost is collapsing. AI-related automated tasks grew 760% in two years on Zapier alone; Gartner expects 40% of enterprise apps to ship with task-specific agents by the end of 2026. When the cost of solving a defined problem trends toward zero, the value of the definition itself absorbs the difference. This isn't a metaphor; it's just where the margin goes. The scarce input gets the premium, and the scarce input is no longer the ability to do the work. It's knowing which work is worth doing.

You can see who's standing on the far side of this cliff. Not the fastest executors. The people who walk into a business and notice the problem that isn't on any ticket: the intake process everyone tolerates, the report nobody reads but everyone produces, the customer behavior that contradicts the company's whole model of itself. In our own practice, the engagements that mattered most never started from the brief we were handed. They started from something we noticed while doing the brief. The brief was the door. The found problem was the house.

Problem-finding is a practice, not a gift

Problem-finding is learnable behavior, not innate talent. Stay in contact with the real situation longer before accepting a problem definition, collect frictions without solving them, interrogate what everyone has normalized, and automate your own execution so your attention is free for noticing.

The Getzels finding would be depressing if problem-finding were innate. It isn't. It's a set of habits, and they're learnable:

Sit with the objects longer. The problem-finders in the Chicago study physically handled more objects for more time before drawing. The work equivalent: resist the brief for an hour. Walk the actual process. Watch a real user. The problem that matters is usually visible in the first contact with reality and invisible in the document describing it.

Collect frictions, not solutions. Keep a list of things that feel heavier than they should be, in your work, your industry, your tools. No solving allowed at capture time. Most entries are nothing. The list is how you find the one that isn't.

Interrogate the normalized. The biggest problems don't feel like problems; they feel like Tuesday. Ask "why does everyone in this industry do X" until you hit an answer that's only "because we always have." That's ore.

Automate your own execution first. This one is self-serving coming from me, and it's still true. You cannot find problems while drowning in tasks. Every hour of execution you automate is an hour of attention that becomes available for noticing. The machines don't just do the work; they buy back the conditions under which thinking happens.

What this means if you run a business

Hire and evaluate for judgment upstream of execution. The valuable question for any partner or agency is no longer "can you build this?" but "what would you have noticed that we didn't?" Roles defined by completing assigned tasks depreciate; roles defined by finding the task appreciate.

It means the org chart inverts slowly and then all at once. You'll need fewer people who complete defined tasks and more people who can stand in the mess upstream of definition, and the second kind doesn't show up on a resume keyword scan. It means "we don't know what we need" is becoming the most honest and most valuable opening line of a client engagement. And it means the right question to ask any partner, vendor, or agency is no longer "can you build this?" (everyone can build it now) but "what would you have noticed that we didn't?"

That conviction is the entire reason our practice is structured the way it is: the automaton agency model exists because production had to be automated before human attention could live full-time at the layer where problems get found. It's the top of the five-layer stack for a reason: it's the only layer that appreciates.

School taught us to be answer machines, and then we built actual answer machines. What's left is the part school skipped: learning to find the question. I think that's not a loss. I think it's the most interesting job description humans have ever been handed.

FAQ

Published: July 2026.

Related: The agentic website: when your site runs itself, your job moves up a layer · What an AI SEO agency actually does · What is an automaton agency? · The five-layer framework