The Abstraction Layer

AI will create new work. "When" matters more than "what".

There’s a false debate running through the center of our economic conversation.

On one side, you have the advocates of technological displacement who warn of a jobless future, pointing to the surgical cuts in tech, finance, and professional services as the first signs of a structural collapse. On the other side, you have the cheerleaders of augmentation who point to near-historic low unemployment and promise AI will simply make us more productive, creating millions of new roles we cannot yet conceive.

Both positions are right, and both are missing the point.

The debate is false because it treats a temporal gap as a structural choice. The reality is that we’re about to enter a valley. If there’s a gap between the moment the old roles dissolve and the moment the new roles form, that gap is a valley. And that valley is where we will live for the next few years.

To understand why this is happening, we have to stop looking at the headlines and start looking at the structure of how technology actually changes work.

The three propositions of the “Not Yet”

Yesterday I laid out three counterintuitive propositions that will guide our work over the next few months:

1.I do not believe there has been any meaningful job displacement by AI.

2.I do not believe there has been any meaningful job creation by AI.

3.I do not believe there have been any meaningful productivity gains from AI.

For each of these propositions, there’s an implied yet. I believe all three will happen, and I have seen no compelling arguments to the contrary. The question guiding me in each has been when, and the answers for each of those three propositions have meaningful stories arising from the differences between them.

If the AI jobs displacement article was about the mechanics of the displacement—the quiet, surgical freeze happening beneath the surface of the BLS numbers—this piece is about the other side of the ledger. It is about what happens when the old roles are gone, where the value goes when it leaves, and what the new layer of human contribution actually looks like.

To see that picture clearly, we have to look past the temporary scaffolding we are building today and understand the difference between a bridge and a destination.

The scaffolding and the destination

If you open any mainstream business report today, you will see a list of “AI jobs” that are supposedly the vanguard of the new economy. They are roles like prompt engineers, AI safety auditors, model fine-tuners, and vector database specialists.

These are not the future of work. These are the scaffolding.

When a building is under construction, you see steel pipes, wooden planks, and safety netting clinging to the exterior. The scaffolding is necessary. It’s also temporary. Once the building is finished, the scaffolding is dismantled and carted away. No one mistakes the scaffolding for the architecture.

Yet we’re making exactly that mistake with the current crop of AI-adjacent roles. Prompt engineering is a classic bridge job. It exists only because our current models are crude, temperamental, and require specialized human translation to produce reliable outputs. Within eighteen to twenty-four months, as agentic architectures mature, the models will handle their own prompting. They’ll write their own system instructions, optimize their own context windows, and audit their own outputs.

The bridge jobs will evaporate on the exact same timeline as the roles they were built to replace.

The real destination is not a set of specialized technical roles. It’s a new cognitive layer—what we define as Orchestration Work.

The historical pattern of the layer

This is not the first time we have built a new layer. In fact, it is the only way our economy has ever adapted to major technological transitions.

Technology doesn’t simply destroy or create tasks. It abstracts the tedious, mechanistic layers of work into a utility foundation, forcing human labor to migrate upward to a new cognitive layer.

Consider the Second Industrial Revolution. Before factory electrification in the late nineteenth century, manufacturing relied on manual workshop assembly, localized steam power, and the proprietary, localized knowledge of master craftsmen. Electrification and the assembly line abstracted those mechanistic tasks into the physical infrastructure of the factory.

The result wasn’t the permanent end of work. It was the invention of the Professional-Managerial Class.

Suddenly, the economy needed industrial engineers, cost accountants, corporate compliance officers, personnel managers, and systemic planners. These people did not touch the physical product. They did not turn the screws or shovel the coal. Their job was to design, coordinate, and optimize the systems that did.

The mechanistic execution had been abstracted into the utility. The human contribution had been elevated to orchestration.

We saw the same pattern during the computing revolution of the late twentieth century. Before the mainframe and the database, corporate offices were filled with typist pools, manual ledger bookkeepers, and rooms of clerks performing physical calculations. The spreadsheet and the relational database abstracted those tasks into software.

The new roles that emerged were systems analysts, database administrators, network engineers, and digital operations managers. The work shifted from executing the calculation to managing the system that ran the calculation.

In every case, the transition followed a predictable sequence:

Mechanistic Execution ─→ Abstraction into Utility ─→ Elevation to Orchestration

This transition is never instant. There’s always a lag—a period where the old work has been automated but the new organizational structures have not yet been designed to utilize the newly freed human capacity. Historically, this lag has lasted decades.

Figure 1: The Layering Pattern. In previous transitions, the gap between infrastructure deployment and the measurable emergence of the new labor layer spanned 15-50 years. The agentic transition is compressing this cycle, but the structural lag remains.

The active infrastructure

I’ve used the metaphor of home plumbing to describe this transition, and it’s a useful way to understand the invisibility of the change. Once indoor plumbing works, you stop thinking about it. The work of carrying water disappears into the walls. The mechanistic task becomes a background utility.

But plumbing is a passive metaphor. You turn the tap, and the water flows. The AI layer we’re building isn’t passive. It’s executing, generating, and making low-stakes decisions continuously.

A better analogy is the electrical grid.

The grid is infrastructure, but it’s also active. It’s constantly balancing load, rerouting power around failures, and responding to fluctuating demand in real time. The people who work with the grid aren’t plumbers. They’re grid operators, load forecasters, and systems engineers. Their job is not to make electricity flow—the grid does that—but to direct, balance, and account for the flow.

Or think of the highway system.

Before the interstate highways, moving goods across the country required constant, localized human attention at every stage: the local driver, the route-planner, the toll collector, the physical cargo transfer crew. The highway system abstracted all of that into a national utility.

The new jobs that emerged were not “highway operators.” They were logistics coordinators, supply chain managers, and distribution network designers. The infrastructure absorbed the physical movement, which allowed human labor to elevate to thinking about what moves where and why, rather than how to make it move.

This is the transition knowledge work is making today. The writing of the code, the drafting of the contract, the formatting of the slides, the reconciliation of the ledger—these are the local roads. They’re being abstracted into the active highway of agentic AI.

We can see this active infrastructure operating at scale today. Look at logistics giant C.H. Robinson. By implementing agentic systems to handle Less-Than-Truckload bookings, they are now managing 29 percent more freight volume than they did in 2019, while employing 30 percent fewer people. Roughly half of their carrier bookings are now generated and routed entirely by AI agents. The physical movement of freight remains, but the cognitive coordination has been abstracted into the utility.

The new work isn’t about making the document or booking the truck. It’s about deciding where the document goes, what it means for the organization, and who owns the accountability for its contents.

The architecture of orchestration

What does this orchestration work actually look like on a normal workday?

I want to suggest a preliminary framework for how this new layer is structured. This isn’t a settled taxonomy, but rather a working model of the three core capabilities that cannot be delegated to a machine:

1. Context Engineering

An AI agent can execute a workflow, but it can’t understand the organizational context in which that workflow exists. It doesn’t know that the client who is complaining about a billing error is also the CEO’s brother-in-law, or that a sudden spike in server latency is related to a marketing campaign launching next week.

Context engineering is the work of feeding the right organizational realities, constraints, and relationships into the active systems so the output matches the human environment.

2. Algorithmic Auditing and Exception Handling

When an active system runs thousands of processes simultaneously—reconciling invoices, provisioning IT access, or drafting compliance reviews—it will inevitably hit exceptions. An edge case that doesn’t fit the training data, a conflicting data point, or a subtle hallucination.

The human orchestrator doesn’t run the process. They manage the exceptions. They are the air traffic controllers of the cognitive flow, stepping in only when the system flags a deviation from the flight plan.

3. Accountability Ownership

An AI agent can generate a medical diagnosis, a legal contract, or a financial projection. It can’t own the liability for the outcome.

If the diagnosis is wrong, the machine doesn’t get sued. If the contract is flawed, the machine doesn’t face regulatory fines.

Accountability can’t be abstracted. The ultimate human contribution in the orchestration era is the willingness to sign your name to the output of the system and say: I review this, I understand this, and I own the consequences.

The measurement blind spot

If this new layer is forming, why can’t we see it in the economic data? Why are the BLS and JOLTS numbers silent on the creation of orchestration work?

Because our statistical instruments are designed to measure a world that’s disappearing.

The BLS Establishment Survey is built to track stable, W2 headcount within established corporate hierarchies. It’s an instrument designed for the industrial and computing eras. But orchestration work doesn’t form as traditional W2 headcount.

When a company automates its procurement department, it doesn’t hire five “AI Orchestrators” to replace the fifty clerks it let go. Instead, it signs a contract with an AI integration firm, hires a freelance consultant to design the workflow, and trains one remaining manager to run the system.

The new work is forming on the margins. It’s showing up as freelance platform transaction volume, as micro-consultancies, as project-based integration contracts, and as “silent” skill upgrades within existing roles that don’t change the job title.

This structural split has created what economists are calling the “Big Freeze.”

Figure 2: The Big Freeze. US hiring rates have collapsed to levels last seen in 2010—when unemployment was near 10 percent—yet current unemployment remains near historic lows. Companies aren’t firing, but they’ve quietly closed the door to new hiring.

This freeze doesn’t fall equally. It’s concentrated at the entry point of the professional career. While the overall labor market looks stable, recent research from Stanford’s Digital Economy Lab shows a 16 percent decline in early-career employment across the most AI-exposed occupations since the release of ChatGPT (~1300 days ago).

Unemployment among recent college graduates has climbed to nearly 6 percent, rising twice as fast as the rest of the workforce. The door to entry-level knowledge work is swinging shut. And because entry-level roles have historically been the training ground where workers develop the contextual wisdom required for senior roles, this freeze threatens the long-term pipeline of human talent.

If you want to see the new work before the BLS reports it in three years, you have to look at different instruments:

Intensive vs. extensive margin: We should track Gusto’s cohort data on startup headcount. We’re seeing a record high in new business applications, but a persistent decline in first-year headcount. AI is making firms smaller (intensive margin) but enabling more people to start them (extensive margin). The work is being created, but it is distributed across more, smaller entities.

Freelance transaction value: We should look at aggregate transaction volume on platforms like Upwork and Fiverr, specifically tracking the shift from “execution” contracts (writing, basic design) to “orchestration” contracts (systems integration, workflow design). The execution contracts are falling; the orchestration contracts are rising.

Skill adoption curves: We should monitor LinkedIn’s Economic Graph data for “role re-titling” velocity—the rate at which traditional roles append systems coordination and algorithmic auditing to their profiles without changing their official W2 titles.

We are trying to measure a basketball game using a baseball scorecard. The game is moving too fast for the instrument.

The valley and the social license

This brings us to the timing projection. If the displacement is happening at corporate speed (Q4 2026 / Q1 2027) and the new orchestration layer requires organizational trust and legal clearance to mature, how long is the valley?

My best estimate is eighteen to thirty-six months.

This is a significant gap. If massive white-collar displacement occurs in the first half of next year, we cannot wait three years for the market to organically form the orchestration layer. The social friction will be too high. The political backlash will be too severe.

This is where the concept of social license comes in.

The frontier AI labs—the companies spending $700 billion on capex—are highly sensitive to the political environment. They know that if their technology is seen as a pure job destroyer, they’ll face aggressive regulatory crackdowns, robot taxes, or outright bans.

To protect their regulatory freedom, they will have to fund the bridge.

We’re likely to see the emergence of a privately funded, modern equivalent of the Works Progress Administration (WPA). Not a government program—which would require a political consensus that does not exist—but a multi-billion-dollar retraining and placement initiative funded directly by the cash-rich beneficiaries of the transition.

They’ll partner with major enterprise platforms like Workday, ServiceNow, and Salesforce to build massive, accelerated “Orchestration Academies,” paying displaced workers to retrain as the systems engineers of the active infrastructure.

They’ll do this not out of charity, but out of survival. It’s the price of their license to operate.

The open door

The May employment print was solid, but it was a rear-view mirror. The consensus is still looking at the headline and feeling relieved.

But if you look at the structure, the house is already divided. The entry-level on-ramp is freezing over, the execution layer is beginning to dissolve, and the active infrastructure is being installed in the walls.

This is happening—no question about that. The uncertainty is in how deep and wide the valley is before the new structures take hold.

The signal to watch over the next two quarters isn’t the unemployment rate. It’s the spread between entry-level white-collar vacancies and senior coordination roles. When that gap begins to widen dramatically, it means the valley has arrived. Those who understand the structure of the active infrastructure before the headline changes will have had a few quarters to act.

That’s the window. It’s not closing dramatically. It’s just closing.

---

If you want the underlying analyses and data dashboard tracking these narratives, send me a message and I can share the details and sourcing.