What AI Actually Breaks When You Sell Thinking for a Living

Assemblers, architects, and the collapse of the formulation premium

Most of what we call knowledge work is formulation.

Not content production. Not information processing. Formulation — the act of applying individual knowledge to a specific context in order to turn messy reality into structured action.

A lawyer does not produce documents. A lawyer formulates: taking facts, constraints, and risk posture and converting them into an argument, a contract, or a compliance structure that commits real people to real consequences. A software architect does not write code. They formulate: converting a business goal into a set of design decisions that make the thing buildable. A strategist does not make decks. They formulate: converting an ambiguous situation into a plan that can survive contact with the organization.

Every knowledge profession, at its core, is a formulation profession. The artifacts are outputs. The formulation is the work.

And it is exactly this — the formulation premium — that AI is now attacking. Not uniformly. Not everywhere. But in ways that most people are not tracking, because they are looking at artifacts instead of the logic underneath them.

Assemblers and architects

Here is the split most people miss: Most of what gets sold as knowledgework, what gets paid for, is formulation.

But often, what feels like formulation to the person doing it, is actually assembly. It is the recombination of known patterns, known structures, known rhetorical moves, and known frameworks, applied to a recognizable situation. It is skilled work. It takes experience. But it operates on materials that already exist in recognizable form.

AI can do assembly...in fact, one could argue as I do, that it is pretty much the only thing that it does. The patterns are in the training data.

What AI cannot do reliably is architecture — the kind of formulation that requires reading the actual situation and building something that fits it, rather than fitting the situation to templates. Architecture demands three things AI does not have.

The first is contextualization. Not the generic version of the problem, but the real one — the politics nobody wrote down, the client's actual risk tolerance as opposed to their stated risk tolerance, the history of what has already been tried and why it failed.

The second is genuine novelty. AI recombines what has been done before. It cannot reliably create structures that do not yet exist in recognizable form, because it has no way to evaluate whether something truly new is good — only whether it resembles things that were.

The third is tacit knowledge detection. Recognizing what is not in the brief. The unstated assumptions. The unwritten rules. The things that would make the plan fail that nobody has surfaced yet. AI does not know what it does not know. An experienced architect does.

Assemblers operate on known patterns. Architects operate on real conditions. AI commoditizes the first. The second is where value concentrates.

The Architect – Assembler – Architect Loop

In practice, the two work together. Architects set constraints and direction. Assemblers — increasingly AI-assisted or AI-performed — generate within those constraints. Architects review the output against reality, catch what the assembly missed, and redirect.

That loop — architect, assembler, architect — is becoming the fundamental unit of knowledge work production. And it does not happen just once. Within any segment of a value chain, the loop may iterate multiple times as the formulation is tested and refined. Or it may happen sequentially, each cycle building on the last — a constructive process where every judgment pass adds a layer to the final output. A legal strategy is not formulated in a single pass. Neither is an organizational redesign, a software architecture, or a campaign. Each loop is additive. Each one requires the architect to bring real judgment to bear on what the last cycle produced.

Which means AI does not reduce the need for architect work. It accelerates the demand for it. When assembly happens at machine speed, the judgment layer has to move faster too. The architect becomes the bottleneck in a much faster cycle, and the organizations that cannot speed up their judgment layer will find that faster production just means faster error.

Two axes, four worlds

To see where this plays out, you need two dimensions.

AI substitutability of production. How much of the core output can AI generate in usable or near-usable form? Low means AI cannot really do the work. High means AI can produce something that ships, or that can be quickly brought to shippable form.

Judgment burden and error cost. How expensive is it to be wrong? Low means mistakes are cheap and reversible. High means mistakes are costly, regulated, or reputationally dangerous.

These two axes create four distinct worlds, each with different implications for assemblers and architects.

Collapse — High substitutability, low judgment burden

This is where assembly dies fastest.

When AI can generate usable output and the cost of being wrong is low, the market floods with "good enough." The flood does not just increase competition. It changes the pricing logic of the work. The median output becomes abundant, so the median price falls.

In this quadrant, assembly gets commoditized directly. Basic marketing copy. Routine web pages. Standard internal decks. Boilerplate emails. First-pass analysis. Simple code scripts. Social media calendars. Proposal shells. The same structural patterns can now be generated endlessly, and the differentiator shifts away from who can produce them.

The important thing to understand is not that this eliminates jobs. It eliminates the pricing logic that supported those jobs. When a client can generate five passable versions of something in twenty minutes, the conversation about what they will pay for the sixth version changes permanently. That is a structural repricing.

Stratification — High substitutability, high judgment burden

This is where people get confused.

When AI output is usable but the cost of error is high, the work does not collapse. It stratifies. Assembly accelerates. The architect-assembler-architect loop spins faster. And the judgment layer — not the production layer — becomes the binding constraint.

Legal drafting is the clearest example. AI can produce documents. But the real cost is not generating the draft. It is living with the consequences of what the draft commits you to. A contract clause that sounds plausible but creates an unintended obligation is not a productivity gain. It is a liability event. So the economic center of gravity moves upward — into review, governance, and risk allocation.

Financial models, clinical documentation, and security-sensitive code all behave the same way. The drafting gets faster but the oversight requirement grows. The faster the drafting, the more review surface is created, and the harder it becomes to tell the good output from the dangerous output at speed.

The winners in this quadrant are not the fastest generators. They are the people who design the constraint systems, the review architectures, and the validation gates that keep speed from becoming disaster. The bottleneck is not production. The bottleneck is trustworthy review — and it is growing.

Cognitive High Ground — Low substitutability, high judgment burden

Some work is not easily substituted because it is not primarily about producing an artifact. It is about designing the system that produces artifacts. It is about integrating conflicting constraints across stakeholders. It is about navigating ambiguity, risk, politics, and the physics of real organizations.

Complex architecture. Systems integration. Organizational transformation. High-stakes decision design.

I spend most of my time in this quadrant. When I work with an agency leadership team on restructuring how they deliver, the "deliverable" is not a document. It is a set of decisions about team structure, governance, incentive alignment, and workflow sequencing that have to survive contact with dozens of people who each have different priorities and different definitions of success. No AI generates that. It requires reading rooms, absorbing contradictions, making tradeoffs explicit, and carrying the accountability when some of those tradeoffs do not work.

This was always architect work. Here, AI changes the surface area — it accelerates research, generates options, speeds up documentation — but fluency with AI becomes assumed, not differentiating. The premium is synthesis, accountability, and the ability to connect formulation to reality.

Physical High Ground — Low substitutability, low judgment burden

This quadrant changes slower, but not because the work is simple. Much of what lives here — skilled trades, physical-world services, localized operations, bespoke craft — involves constant real-time judgment that operates on a fundamentally different kind of knowledge.

The electrician diagnosing a problem in an old building is not just using their hands. They are reading a situation that includes visual cues, spatial relationships, material conditions, and the accumulated knowledge of how buildings of that era were typically wired versus how this specific one actually was. That information is situated — it exists in the encounter between the person and the environment. It is not textual. It was never written down. It cannot be passed into a context window. Add tacit knowledge on top of that — the feel for when something is wrong that comes from years of embodied practice — and you have a domain where AI's limitations are not about compute or model size. They are about what kind of knowledge AI can even access.

AI becomes a smart handbook here. A better reference tool. A faster diagnostic aide. But the architecture happens in the act of doing, and the information it depends on is not the kind AI can reach. This is high ground of a different kind — defended not by complexity of thought but by complexity of presence.

How AI disrupts the traditional value chain

The framework becomes concrete when you apply it to a value chain most professional services people know: the one between an agency and its client.

Historically, clients paid agencies for three things bundled together. They paid for output, because output was expensive to produce. They paid for translation — turning vague business intent into workable creative and execution plans. And they paid for judgment, because they wanted someone to take responsibility for what was said, made, shipped, and risked.

AI unbundles that.

Output gets cheaper fast, which means clients increasingly feel they should not have to pay agency prices for "the thing" — especially for Collapse-quadrant work. This is where clients start pushing for lower fees, more revisions, more volume, and shorter timelines, while sincerely believing they are being reasonable.

At the same time, formulation becomes double-edged. Clients gain the ability to generate plausible plans, briefs, strategies, and creative directions on their own. In doing so, they move into the Collapse quadrant themselves — inhabiting the assembly layer — and this increases their confidence. It does not increase their judgment. So a new pattern emerges: clients arrive with stronger opinions and more drafts, but also more hidden error, more incoherence across channels, and more internally conflicting constraints they do not see. They are generating Collapse-quadrant output that actually requires Stratification or Cognitive High Ground review to use safely.

I see this constantly now. A client shows up with an AI-generated brief that reads well on the surface. Clear objectives, defined audience, even a creative direction. But the brief contradicts their brand positioning from six months ago. Or it promises outcomes across three channels without accounting for the fact that those channels have different conversion mechanics and different timelines. The assembly looks right. The architecture is missing.

What changes in the value chain is not just price pressure. It is who is doing the formulation, and who is carrying the error cost.

In the old world, the agency quietly absorbed a huge amount of that error cost. You fixed bad briefs. You reinterpreted conflicting stakeholder input. You smoothed incoherence. You protected the client from their own internal fragmentation. That was architecture — and it was bundled into the price of production. Clients never paid for it explicitly. They got it for free.

Who Moved My Margin? The Repricing Problem

When production gets cheap — or clients do it themselves — the architecture that was bundled in disappears with it.

This is the hardest commercial problem in the shift. Agencies cannot just "start selling judgment." They have to reprice it from zero. They have to convince clients to pay for something those clients never knew they were receiving.

What does that look like in practice? It looks like an agency that gets paid to run a constraint-setting workshop before any creative begins — where the real brand boundaries, risk tolerances, and channel tradeoffs are made explicit, in writing, with sign-off. It looks like an agency that owns a defined review gate, with the authority to reject work that passes the AI plausibility test but fails the brand coherence test. It looks like an agency whose scope includes "we will tell you when your brief is wrong and show you why" — and that charges for that, rather than absorbing it as unpriced overhead.

That requires swimming upstream. The agency-client boundary has to move so the agency is involved earlier, in the problem definition itself, not just in the execution of a brief someone else wrote. The agencies that thrive will be the ones that reposition from "we make the thing" to "we own the thinking that makes the thing safe to make."

Clients, meanwhile, will increasingly want a relationship that feels like speed and abundance. They will ask for "just one more version" because versions feel free now. The agency has to counter with a different promise: not more output, but better consequence control.

As output becomes abundant, clients will try to buy volume. Agencies will have to learn to sell consequence.

This is not unique to agencies. The same dynamic is disrupting every professional services value chain where AI makes production cheap — law firms, consultancies, software shops, architecture practices. Professional services may in fact be the most exposed category, precisely because they are formulation businesses by definition. The agency version just makes it visible first because the unbundling is already happening in real time.

What this means

Two things follow from all of this, and they apply to firms and to individuals alike.

The first is that the judgment layer is not shrinking. It is speeding up. When assembly happens at machine speed, architects do not get to slow-walk their decisions. The architect-assembler-architect loop is getting faster, and each pass still requires real judgment applied to real conditions. The organizations and professionals who can compress that loop without sacrificing the quality of judgment will pull ahead. The ones who cannot will produce more output, faster, with more embedded error — and they will not understand why things are getting worse.

The second is that the market is about to reprice the difference between assembly and architecture in every knowledge profession. For a long time, both were bundled together and paid for as one thing. AI is unbundling them. Assembly is heading toward commodity pricing. Architecture — real formulation, connected to reality, accountable for consequences — is heading toward scarcity pricing.

Every professional services firm, and every professional, now faces the same question: are you assembling, or are you architecting?

The market will not wait long for the answer.