Most of the conversation about AI and work is about process. Which tools to use, which workflows to adopt, how to integrate AI into existing routines. It makes sense. Process is what fills the day. Process is what’s visibly changing. And process is where the anxiety concentrates, because if AI can handle the steps you used to do manually, the natural question is: what’s left?
But I think the fixation on process misses the structural shift underneath it.
Work has three layers. At the front, there’s an outcome: what should exist that doesn’t yet? A legal strategy, a product architecture, a market entry plan, a piece of writing with a specific purpose. At the back, there’s validation: does this output actually achieve what we intended? And in the middle, there’s the production process: research, drafting, analysis, coordination, iteration. The sequence of steps that turns a specified outcome into a deliverable output.
For most of the history of knowledge work, these three layers were bundled together inside the same person. A lawyer didn’t just specify the legal strategy and then validate the brief. She produced it. Her expertise lived in the process as much as in the judgment that framed the outcome. The specification, the execution, and the validation were inseparable, because the process was expensive enough and complex enough that only someone who understood all three could do the work.
That bundling is coming apart.
When the cost of producing an output approaches zero, or at least drops by an order of magnitude, the value in the chain migrates to the endpoints. The person who specified the right outcome owns more value than the person who executed flawlessly on the wrong one. And the person who can validate whether the output actually achieved the intended outcome, who can look at a deliverable and say “this is right” or “this misses the point,” holds the quality gate that determines whether the cheap production was worth anything at all.
The middle, the production process, becomes commodity infrastructure. Not worthless. Still necessary. But no longer the place where differentiation lives and no longer the primary justification for a role.
The bookends
I want to be specific about what “specification” means, because it’s easy to confuse with prompting. Prompting is a tactic. Specification is an exercising of judgment and the clear setting out of what makes a “good” output in service of outcomes.
A good specification isn’t a detailed instruction set. It’s a clear articulation of the outcome that matters, framed with enough context that the system (whether AI or human) can make reasonable decisions about the hundreds of micro-choices involved in production. It requires understanding the purpose of the work, the audience, the constraints, the trade-offs. It requires knowing what good looks like before it exists.
Validation is the mirror image. It’s the ability to look at a finished output and assess whether it achieved the intended outcome, not whether it followed the process correctly. These are different evaluations. An output can follow every step perfectly and still miss the point. An output can take an unexpected path and nail the outcome. Validation requires the same judgment as specification: understanding what good looks like, but applied retrospectively rather than prospectively.
Together, specification and validation form the bookends of a production cycle. The human sits at both ends. The middle is where AI is most naturally suited, and where the cost collapse is most dramatic.
The uncomfortable part
Here’s what I think deserves honest acknowledgment: most knowledge workers were never trained to separate specification from execution. The two skills were bundled together, and for good reason. A consultant’s ability to specify the right analytical framework was developed through years of building analyses. A designer’s ability to specify what a product should feel like emerged from years of producing designs. The specification skill wasn’t taught independently. It was a byproduct of mastering the process.
So when someone says, for example, “all code gets commoditized to near zero production cost,” the execution isn’t about tools and workflows. It’s about whether the specification and validation skills can be unbundled from the process skills, and whether they’re sufficient on their own to justify a continued human role.
Sometimes the answer is clearly yes. A senior architect who has spent twenty years understanding how buildings work, how people move through spaces, how materials behave under stress, can specify an outcome and validate the result regardless of whether she’s the one producing the drawings. Her judgment was always the valuable part. The drawings were evidence of that judgment, not the source of it.
Sometimes the answer is less clear. A mid-career financial analyst whose primary skill is building models in Excel has specification knowledge (what the model should capture) that’s deeply entangled with process knowledge (how to structure the spreadsheet, how to handle edge cases, how to make the formulas robust). If AI handles the model construction, does the analyst’s specification skill survive the unbundling? Maybe. Maybe not. It depends on whether the judgment was genuinely independent of the process, or whether the process was where the judgment lived.
I don’t think there’s a universal answer. The honest position is that the unbundling will reveal, for each role and each person, whether the specification and validation skills were the substance or the shadow of the process skills. For some, the process was a vehicle for judgment. For others, the judgment was a vehicle for the process. The distinction only becomes visible when the process layer drops away.
What this changes about the cost of mistakes
There’s a secondary effect that I think matters as much as the structural one. When production is expensive, the cost of specifying the wrong outcome is partially hidden by the production timeline. You won’t discover the mistake for weeks or months, by which time the sunk cost creates its own momentum. The wrong strategy gets executed thoroughly because the execution itself took long enough to feel like progress. There’s also a chance that lots of human micro-evaluations along the way actually get you to a decent outcome.
When production is cheap and fast, the cost of specifying the wrong outcome gets paid immediately. You get the output in hours, not months. If it’s wrong, you know it’s wrong right away. There’s no sunk-cost illusion to hide behind. Nor, cruicially is there any chance that it’ll get course corrected along the way. In an automated production paradigm: garbage in gets garbage out.
This accelerates the feedback loop between specification and validation. Specify, produce, validate, respecify. The cycle that used to take a quarter now takes a day. That’s powerful, but it also means that weak specification is exposed immediately and repeatedly, rather than concealed by a long production timeline. The person who can specify well becomes dramatically more productive. The person who can’t becomes visibly, measurably stuck, cycling through outputs that never quite achieve the intended outcome.
The gap between strong specifiers and weak specifiers, which was always present but muted by the cost of production, gets amplified. Fast, cheap production is a multiplier for good judgment and an accelerator for bad judgment.
There was a popular theory early in the AI adoption curve: cheap production would let junior people perform like seniors. Give a first-year analyst AI tools and she’d produce the output of a ten-year veteran. The leverage would flow downward. Organizations could hire cheaper, train less, and let the tooling close the gap.
What’s actually emerging in the labor market is closer to the opposite. Seniors got superpowers. Juniors got barista jobs.
This makes sense once you see it through the specification-validation lens. A senior operator’s advantage was never primarily in the speed of production. It was in knowing what to produce, recognizing when the output is wrong, how it might not serve the desired outcome, and understanding the “why” in ways that inform the next specification. Those are the bookend skills. When you hand that person a tool that compresses the middle, they don’t just get faster at the same work. They get access to a cycle speed that was previously impossible: specify, produce, validate, adjust, re-produce, all within a single sitting. Suddenly, everyone involved in the process is out of their way and they can fly. Their judgment, which used to be bottlenecked by the production timeline, now operates at something closer to its natural clock speed.
A junior, by contrast, was still building the judgment. The process was the training ground. The years of manually constructing analyses, drafting briefs, building models: that wasn’t just production. It was the accumulation of pattern recognition that eventually becomes specification skill. Remove the process, and you don’t get a junior who performs like a senior. You get a junior who can produce outputs quickly but can’t tell whether they’re good. The feedback loop that the senior uses to iterate toward the right answer becomes, for the junior, a loop with no compass. Fast iteration without strong validation is just fast drift.
The result is a labor market that confounds the original prediction. Instead of AI compressing the seniority premium, it’s widening it. The experienced operator with strong specification and validation skills becomes dramatically more productive, because the constraint was never their judgment but the cost of acting on it. The junior who hasn’t yet developed those skills loses the primary mechanism through which they were developed. The tool that was supposed to democratize capability is, at least in this early phase, concentrating it. The downstream consequences on human capital development are stark- something I might pick up in another blog one of these days.
Not a prescription
I’m not offering career advice. I don’t know which roles will survive the unbundling and which won’t, and I’m skeptical of anyone who claims to know. The structural observation is narrower than that: human leverage in an AI-augmented workflow concentrates at specification and validation. The middle compresses. That’s a description of where value migrates, not a prescription for what any individual should do about it.
What I will say is that the people I’ve watched navigate this transition most effectively are the ones who already thought of themselves as outcome-oriented rather than process-oriented. They weren’t attached to the steps. They were attached to the result. When the steps changed, they adapted without an identity crisis, because their identity wasn’t in the process.
The people who struggle are the ones whose expertise was expressed entirely through a process they’d mastered over years. When that process gets compressed or automated, they don’t just lose efficiency. They lose the framework through which they understood their own contribution. That’s not a skills question. For many, it’s something closer to an existential one.
I think the distinction between these two responses will define a significant portion of how knowledge work reorganises over the next several years. Not the technology. Not the tools. The question of whether the humans in the system know what they’re here to do, independent of the process they used to do.