Part of the Pristone Academy AI Technical Track
Version 1.0 — a framework
AI doesn't replace jobs. It changes the economics of the layers within them.
As implementation gets cheap, work is redistributed rather than eliminated — and the safe zones are the top and the edges, not the middle.
AI doesn't move people up. It moves the filter.
The most common question about AI and work is the wrong one. “Will AI replace this job?” treats a job as one indivisible thing. But a job is a stack of layers — levels of abstraction at which decisions are made and value is created. AI does not swallow the whole stack. It penetrates some layers, leaves others, and redistributes human effort toward the layers where people keep the greatest comparative advantage.
Layers are not titles. A founder may spend the morning writing code; a senior engineer may make product decisions; a nurse may exercise executive-level judgment during an emergency. The org chart and the layer map are different things. Your exposure to AI depends on the layers your work actually occupies, which may span several boxes or none of the ones your title suggests.
The mechanism runs in one direction. AI lowers the cost of implementation. Implementation becomes abundant. Scarcity shifts toward judgment, trust, customer understanding, architecture, distribution, and accountability. Organizations reorganize around the new scarcity. Individuals redistribute across layers.
AI does not simply replace jobs or automate tasks. It changes the economics of the layers of work. As implementation becomes cheap and abundant, people move up, move sideways, stay and multiply their output, or leave — and where they land depends on which layers AI penetrates, and whether their capabilities match the layers that remain.
Four laws govern how the shift plays out.
Law 1 — Uneven penetration
AI penetrates work according to its characteristics, not its prestige. Digital, observable, testable, cheaply correctable work changes first — software is the leading edge because most of it is text, most of it is repeatable, and mistakes are cheap to catch. Physical, relational, tacit, regulated, and high-consequence work changes later. AI can draft a nurse's charting, but it cannot start the IV, read the family in the hallway, or hold the liability at 3 a.m. Ethan Mollick calls this uneven boundary the jagged frontier; the framework applies it one level up, to organizational layers. The variable underneath it all is accountability: a human still bears responsibility for the outcome, which is why many roles persist long after the technical capability to automate them arrives.
Law 2 — Middle compression
AI compresses any layer whose core function is information transformation. That is not only the bottom of the org chart — status reporting, coordination, routine analysis, and synthesis sit in the middle layers, and AI compresses them too. This creates the framework's central tension: the layers people are being promoted into are the same layers being thinned. More people gain access to fewer seats. The middle becomes a squeeze zone; safety lives at the top and at the edges.
Law 3 — Filter relocation
The old system filtered people before they could try — credentials, hiring committees, promotion cycles, budgets. AI weakens those filters but does not remove filtering. It relocates it. Customers, users, retention, and trust now filter people after they try. Early filtering meant failing privately on someone else's payroll. Late filtering means failing publicly with your own money and your own name. AI democratizes access and privatizes risk in the same motion — extending Jacob Hacker's Great Risk Shift into the structure of work itself.
Law 4 — Layer fitness functions
Each layer rewards a different combination of skills, so success at one layer does not predict success at the next. Engineer, architect, product leader, founder: four different fitness functions. Mayor, governor, senator, president: the same pattern. AI lowers the barrier to reaching the next layer. It does not supply the capabilities that layer selects for. Access is not competence.
Daron Acemoglu and David Autor's task-based model established that automation operates on tasks, not jobs, and that displaced labor reallocates toward tasks where humans keep comparative advantage. This framework applies that lens to whole layers of an organization — and adds the mobility paths people use to cross them.
Individuals answer the four laws with four moves.
Each move carries a signature risk. Choosing a move means accepting its risk — which is exactly what makes this a decision tool, not a taxonomy.
Move into architecture, product, management, strategy, or founding. Signature risk: capability mismatch — the fitness function above you may not match your strengths, and you may not find out until you're deep into a role you can't stand.
Shift into implementation where human advantage holds: legacy modernization, incident response, production reliability, regulated systems, field engineering, customer-facing technical work. Signature risk: the frontier moves — the resistant domain may be next. Sideways moves buy time, not permanence.
Remain an implementer and become an AI-assisted specialist whose output multiplies. Signature risk: commoditization — everyone has the same models. If the edge comes only from the tool, the edge is temporary and the rate is falling.
Change occupations or industries when the other three moves aren't viable. Signature risk: starting over at zero — reputation, network, and tacit knowledge don't transfer at full value. It should be a decision, not a drift.
The mastery case of Up is the multi-layer operator — someone who moves up but keeps implementation fluency, traversing strategy, architecture, code, and customer conversations in the same week. Not a separate move; the upward move executed with retained range.
But range has its own signature risk: diffusion. When everything looks worth doing because you're genuinely good at all of it, effort spreads across layers and nothing compounds in any one of them. Range without a compounding position is motion, not advantage. The fix is to point the range at a single constraint — usually distribution — rather than spreading it evenly across the map.
Three questions decide which move is yours.
Digital and observable means the clock is running. Physical, relational, and regulated means there is more time.
The fitness-function test. Honest answers prevent capability mismatch and wasted sideways moves.
The relocated filter rewards people who take responsibility for results, not people who complete tasks. Every viable move runs through this.
To apply the framework to a whole profession rather than a person, ask five parallel questions: What are its major layers of work? Which is AI compressing, and how fast? Where is human comparative advantage increasing? How does accountability shape adoption here? And which moves are open to the people in each layer?
A framework that predicts nothing is vocabulary. This one predicts five checkable things: organizations get flatter; multi-layer operators command a premium, but only when their range is concentrated on a single constraint; careers become non-linear as sideways and diagonal moves grow; adoption stays uneven in the order Law 1 predicts, regardless of prestige or pay; and trust and distribution become the scarce assets as execution commoditizes. Where the framework fails, this is where the failure shows up first.
Four things to keep about the layer shift.
Cheap implementation doesn't delete work — it moves human effort toward the layers where people keep comparative advantage.
Information-transformation layers compress. Safety lives at the top, where you own outcomes, and at the edges, where work is hard to observe.
Access democratized; risk privatized. Customers and trust now filter people that credentials and committees used to.
AI lowers the barrier to the next layer without supplying what that layer selects for. Every layer has its own fitness function.
Stop asking “Will AI replace this job?” Start asking “How will AI change the economics of the layers within this profession, and where will humans hold the greatest comparative advantage?”
This is an explanatory model, not a universal law. It applies most strongly to knowledge-intensive work, and should be revised as embodied AI, robotics, regulation, and liability frameworks mature. If you're trying to place your own work on this map — which layer you sit in, which move fits, and what to build next — that's the kind of thinking a 1:1 session is built for.
Frequently asked questions
More free lessons
What AI actually is, and how learning from data differs from following rules.
How a machine sorts a pile of points into clean groups with no labels.
How a machine keeps the few directions that matter and throws away the rest.
The self-attention idea behind every modern transformer and LLM.
Curious who's behind these lessons? See the proof of work.