Why AI is the first general-purpose technology whose flagship tool can teach the user how to use it — the way a person would
Most futurism about artificial intelligence and work names an ideal positive outcome of widespread AI deployment — an economy in which AI has absorbed the friction layer of every institution — but treats the route there as a problem somebody else will solve. This is the gap that makes the general public understandably anxious. Neither the executives whose firms will host the pivot nor the workers whose lives will be reshaped by it have a coherent vision of the future they want, let alone a method for crossing into it.
In an attempt to steer the AI disruption in a positive direction, and head off public backlash before it becomes more organized, we suggest the following formulation: The pivot at the level of the firm should be the refactor-not-replace mode of AI deployment implied by the substrate law — Mode 2, the one that strips friction without firing the human and redirects freed time to the parts of the role the substrate of a chemical body that has lived through time is structurally qualified to perform.
The pivot experience at the level of the worker is the question this paper takes up. The argument is that AI is a category of general-purpose technology whose flagship tool can teach the user how to use it — with no additional cost or complexity. Earlier general-purpose technologies tried (the steam engine taught nothing through the tool itself; the personal computer tried with the manual, the help system, Clippy, and mostly relied on human teachers). The frontier model is the instance where the self-onboard property actually works at a level the worker can use at scale. The transition mechanism is not external to the technology; the technology is the vehicle of the transition mechanism itself. The paper sketches what that might look like in three institutions where it has to work — a regional manufacturer running on Excel, a county clerk's office, and a regional bank — and closes with what the distribution of benefits across the rest of the economy depends on.
Futurism about artificial intelligence has produced a great many destinations. Some of them are dystopian — the hollowing of knowledge work, the concentration of capital in the hands of whoever owns the compute, the breakdown of the social contract that has held the post-war American economy together. Some of them are utopian — abundance for all, the end of drudgery, the citizen empowered with capacities no court or guild could ever have ratified. The two camps argue about the destination. Neither side has really spent substantial time on the route... because it was literally the stuff of science fiction less than five years ago. We've reached the point of no-return. Now comes the ride from point A to point B that the sides have envisioned. Now we have to think about it. Fortunately, the technology that is set to cause the disruption and power the transition, is a tool that literally amplifies our thinking power... and is as ubiquitous, on release, as the smart phone in your pocket. So. How are we going to do it.
A post-AI economy in which Mode-2 deployment has become the dominant institutional pattern — the friction layer of the firm absorbed by the tool, the substrate-qualified work of presence and judgment and recognition "expanded into the room" the friction used to occupy — is a real and plausible end-state. But the country does not get there in one step from where it now stands. It gets there by the inside of every firm performing some specific work that has not yet been named. The boss of a regional manufacturer that runs everything on Excel does not have a coherent picture of what his firm looks like after the pivot, or a method for taking the first step toward it. Neither does his accounts-receivable clerk, who has been quietly afraid of the news for a year. Both are right to feel the gap. Neither has been handed a useful answer.
We are not claiming AI is the first technology where users could teach themselves. VisiCalc taught bookkeepers; the personal computer taught accountants. The claim is narrower: this is the first general-purpose technology where the conversational interface — built into the tool's core capability, not bolted on — collapses the teach-the-user step into the do-the-work step at a quality that scales.
The convergence — model capability + interface + ambient compute — happened in 2022–2023. Earlier instances of self-taught diffusion were possible because the work was discrete and the manual was concrete; the convergence makes self-onboard possible where the work is continuous and conversational.
Earlier compliance-AI deployments make this concrete. HSBC's anti-money-laundering pipeline (2012–2020), built in response to a $1.9 billion U.S. fine for prior cartel-money-laundering failures, deployed transaction-monitoring, customer-due-diligence, and sanctions-screening AI at scale. The result was Mode 1: thousands of compliance analyst seats consolidated, analysts reclassified or let go as the system absorbed the pattern-detection work. The deployment looked identical to what this paper proposes — same architecture, same institutional shape, same set of seats being refactored — and produced the opposite outcome. The reason is the convergence above. HSBC's AML systems were classifiers, not partners. They could not explain their reasoning, accept corrections, or learn an analyst's style. The transformative claim of this paper is specific to conversational and agentic AI — systems where the diffusion mechanism IS the work mechanism. The Mode 2 attractor only operates when the AI can be the user's onramp. Identical-looking deployments before the convergence are consistent with the paper's claim: pre-conversational compliance-AI produced Mode 1 because pre-conversational compliance-AI was Mode 1 by design.
The remaining question is the practical one. If Mode 2 is the architecturally correct (most likely to lead to a positive state outcome at the end of the transition) deployment, and the substrate law is the structural reason it is correct, what makes Mode 2 actually happen inside a firm that has none of the apparatus — no consultancy engagement, no dedicated AI team, no executive who has used a frontier tool for more than a sentence rewrite, no budget — that previous industrial transitions required to land their target technology inside an institution?
The historical analogy that makes the question vivid is the previous industrial revolution the United States is widely understood to have missed. The diffusion of internet-of-things sensors, integrated enterprise software, just-in-time inventory tracking, and the broader package of cyber-physical integration that Klaus Schwab named the Fourth Industrial Revolution[1] landed unevenly across the developed world over the 2010s. China — by virtue of a centralized authority capable of telling a manufacturer to install the sensors and a labor force conditioned to accept the instruction — landed it more thoroughly inside its industrial base than the United States did, along with more strategic gains, such as control of rare earth element supplies and battery manufacturing capacity. The United States, lacking that centralized lever, did not land it. A great many American small and mid-sized firms still run their operations, in 2026, on the same Excel workbooks they were using in 2009. The Fourth Industrial Revolution required a hand on the throttle that the American system was constitutionally unable, or unwilling, to provide.
The crucial fact about the AI industrial revolution is that it does not require the same hand. The tool that does the work in a Mode 2 deployment is the same tool that explains the work, costs the upgrade, walks the implementation in, trains the operator, debugs the integration, and answers the question about why the integration broke three weeks later. The tool is its own onramp. We argue what makes this technology category distinctive is that the mechanism of diffusion is identical to the mechanism of doing the work. Previous revolutions all required a separate diffusion apparatus — the consulting engagement, the trade journal, the trade school, the credentialed implementer — to bring the technology into the institution. The personal computer tried to be its own diffusion apparatus (the manual, the help system, Clippy) and the diffusion mostly relied on human teachers anyway. The frontier model is the instance, at scale, where the diffusion-equals-work property actually functions in practice — well enough that an accounts-receivable clerk can install the upgrade by asking the upgrade what to do. The empirical record on diffusion speed is consistent with this: Bick, Blandin and Deming's 2024 NBER work finds generative AI diffused faster than the PC and the internet to American workers in the same time-since-launch window.[2][3]
This is the load-bearing claim of the paper, and it is worth being concrete about what it looks like in practice. Three institutions, three different refactor surfaces, three different worker archetypes. None of the three has a McKinsey engagement. None of the three has an AI department. All three have a person inside who has decided to walk Mode 2 in by hand.
A regional manufacturer of, say, custom architectural metalwork. The firm has thirty-five employees, the founder still walks the floor, and the sales pipeline lives in a shared Excel workbook that has been organically extended over twelve years and is now a quiet horror to anyone who has to touch it. The accounts-receivable workflow involves cross-referencing the shop floor's job-completion records against a separate accounting workbook, manually reconciling job numbers that have drifted across three numbering conventions, and producing invoices that are routinely late because the reconciliation takes a junior bookkeeper an entire afternoon. Mode 1 would be to fire the bookkeeper and put a chatbot on it. Mode 2 is to sit the bookkeeper down with a frontier model and ask her to describe the workflow in her own words for ninety minutes.
The model reads the workbooks she names, asks her clarifying questions about the numbering drift, and writes — over the course of an afternoon, in dialog with her — a small script that ingests the shop-floor records and the accounting workbook, reconciles the numbering, and outputs a stub invoice for her review. The bookkeeper's role is unchanged in title but transformed in content. She no longer spends afternoons doing the reconciliation. She spends afternoons calling the three customers whose invoices are unusually large, building the relationship that is the actual point of a small B2B firm's accounts-receivable function. The financial-press story on the quarterly call, were the firm public, is that the founder retained his bookkeeper, watched days-sales-outstanding drop by twenty percent, and learned that he had been understaffing the relationship-building part of the function by an order of magnitude. He did not get a consultant. He did not buy software. He asked his bookkeeper to ask the model what to do.
A mid-sized county in the rural South, population roughly eighty thousand, operates its land records, vital records, and voter registration through a clerk's office of fourteen full-time staff, most of whom have been there for decades. The office's largest pain point is records retrieval: a title attorney, an heir, or a journalist arrives in person or by phone with a request that requires somebody to walk to the basement, retrieve a microfilmed roll from 1973, read it on the one functioning reader, and either photocopy or transcribe the entry in question. The work is unevenly distributed across the staff. Two clerks bear most of it. One of them is six months from retirement. The Mode-1 framing would have a vendor digitize the rolls and replace the two clerks with a search index. The Mode-2 framing — and this is the version the office's elected clerk actually walked in — is to digitize the rolls in dialog with the two clerks themselves, asking the model to describe the file structure, identify the inconsistencies in how the 1973 entries were indexed versus the 1991 entries, and produce a unified search layer that the clerks themselves operate and continue to extend as new records come in.
The clerks' roles, post-deployment, include the substrate-qualified work the office had never had time to do under the old workflow: actually helping the family that has come in to settle a contested estate, walking them through what the entries mean, advising them on what to do next, making the public service the office had always been supposed to be. The clerks did not lose their jobs. The office did not hire a vendor. The elected clerk paid a contract programmer for two weeks and produced an outcome that the previous decade's digitization vendors had quoted at eighteen months and four hundred thousand dollars. The county commission noticed. The next county over has begun asking how it was done.
A mid-Atlantic regional bank with roughly four billion dollars in assets — call it forty branches, four hundred employees, a compliance department of twelve. The bank's pain point is its Bank Secrecy Act and Anti-Money-Laundering compliance review workflow. Every transaction over a threshold triggers a manual review. The reviewers are reading the same eight templates of customer activity for the thousandth time, flagging the unusual ones for escalation. The flagging accuracy is, in the unhappy way these things are quantified by the regulator, about as good as it has ever been — which is to say, mediocre — and the manual workload is, in the equally unhappy way these things are felt by the staff, soul-destroying.
The Mode-1 framing replaces the twelve compliance reviewers with a model that scores every transaction and surfaces the top half-percent for human review. The framing fails — and it fails specifically because of the substrate law. The compliance reviewer's substrate-qualified contribution is not the templated scan. It is the judgment-with-stake that says: this looks like the template, but my body has spent seven years in this branch and I know this customer's family, and the transaction that statistically fits the templated pattern is, in this case, a daughter helping her father transfer assets before a hip replacement. The model, lacking the substrate, would have escalated the family transaction. The reviewer, possessing the substrate, would not.
The Mode-2 framing reverses the workflow: the model does the templated scan, surfaces every potentially-flagged transaction with a paragraph of context, and asks the reviewer to ratify or override. The reviewer, freed of the rote scan, now reviews eight hundred transactions a day instead of eighty, and applies actual judgment to the ones that need it. The compliance department's headcount holds. The flagging accuracy, on a generation of regulator audits, improves measurably. The financial-press story on the quarterly call is that the bank's compliance reviewers became the best in their peer group — not by being replaced by an AI but by being given an AI to dispatch the templated work the reviewers had never been the right substrate for in the first place.
In none of the three cases was the upgrade landed by a consultancy. In none of the three cases did the firm have an AI department or an executive with a frontier-model résumé. In each case the upgrade was landed by a person inside the firm asking the tool what to do, in plain English, with the tool willing and able to explain what it was doing and why. This is the self-onboard property that previous industrial revolutions tried for but did not achieve well enough to matter. Steam engines taught nothing; the diffusion required guild apprenticeship. Computers tried — Clippy is the memorable mascot for the failure mode — and the diffusion still mostly required human teachers. The frontier model is the instance, at the convergence of model capability + conversational interface + ambient compute, where the self-onboard property functions at a level the worker can use. It is what makes the American AI transition, in principle, performable at the speed of decision rather than at the speed of consultancy. It is the mechanism by which Mode 2 can become the dominant deployment pattern without a centralized authority forcing it from above. The tool teaches its own use. The tool is the onramp.
Every industrial revolution has produced disruption. None has been absorbed without dislocation. The question is never whether to allow the new technology — that question is, in any meaningfully competitive economy, already answered before it is asked — but rather, whether the benefits of the new technology flow back to the ordinary people whose lives the disruption affects, or get captured at the top of the curve by whoever owned the apparatus that produced the technology. The First Industrial Revolution produced both the modern factory and the modern factory worker, and the slow construction over two generations of the institutions — the labor union, the public school, the social-insurance state — that turned the worker's productivity into the worker's standard of living. The Second produced both the modern corporation and the modern middle class. The Third, the digital revolution, produced both the modern platform and the modern productivity wedge between owner and worker, and the institutions that would have closed the wedge are still under construction. The Fourth, by the conventional count, mostly skipped the United States. The Fifth — the AI revolution — can be deployed deliberately, with the question of whether and how benefits flow back to ordinary people answered at the moment the architecture is named rather than retrospectively over two generations. That is the choice the next 3-10 years is making. And the outcome is totally opaque.
In our model, Mode 2 is the deployment pattern that preserves the wage base, where the other possibility is to... decimate the wage base? The wage base is what funds the consumer demand that turns enterprise productivity into national prosperity, and powers the global economy. Mode 1, taken at scale, eats the wage base. Mode 2, taken at scale, preserves it — and, because Mode 2 redirects the freed time toward the substrate-qualified parts of the role, increases the productivity of the wage base on the same headcount. The resulting equilibrium is the one in which the technology pays for itself in margin without the country having to pay for the technology in disruption. The tax base survives. The consumer market survives. The political coalition that funds the rest of the institutional infrastructure — the schools, the pensions, the deferred maintenance on the highways and the grid — survives, with AI improving both product quality/outcome and operational efficiency. Concurrent precedent exists: China's industrial policy is explicitly directing companies to embrace AI without firing workers,[4] Chinese courts have ruled against AI-driven termination without retraining (Hangzhou Zhou case; Beijing Liu case),[5] and California's Executive Order N-6-26 (May 2026) directs state agencies to modernize the Cal-WARN Act and the displaced-worker safety net for AI-driven workforce reduction.[6] The trilogy's substrate-grounded wage-base argument is complementary to this empirical and legal precedent. Rules and regulations are not the American way, freedom and individuality are. In the end, in America, you are better off convincing someone that doing something hard is in their interest, than trying to force compliance. That means the safe and successful transition into the next technological age becomes a communications problem, not a technology or policy-setting problem.
The closing observation belongs to the work that frames the rest of the series. The substrate-asymmetry paper, in its current form, ends with the observation that the country does not really have a choice — that things cannot stay as they are, that the technology has already arrived and is already reshaping the institutions whether the architecture is named or not, and that the honest posture is engagement-with-eyes-open rather than disengagement-as-precaution. The same closing belongs here.
Disengagement is not safer than engagement. It is a different bet, with the same downside and a worse expected payoff. The architecture this paper has described — Mode-2 deployment, substrate-respecting workflow, the self-onboard property of the tool itself, the distribution of benefit through the preserved wage base — is not guaranteed by inaction. The default outcome of inaction is the wholesale-replacement mode that wins by being the easier sell on a quarterly call, and the wholesale-replacement mode produces the disruption everybody is justifiably afraid of. The choice between the two modes is being made, distributed across millions of conversations between executives and clerks and consultants and founders, this year and next. The paper's argument is that the right choice is available and inexpensive, that the technology itself is the mechanism that carries the choice into the institution, and that the missing piece is not capability but commitment.
A question this paper has not directly addressed: why deliberate the white-collar transition when the blue-collar workforce got the shaft a generation ago? Part of the answer is that the blue-collar trades — building a house, fixing a furnace, framing a wall, treating a patient at the bedside — turn out to be exactly the substrate-asymmetric work the law predicts cannot be displaced. The trades are coming back, not going away. The white-collar workforce is where the disruption is now concentrated, and where the substrate asymmetry runs unevenly across altitudes — starkest at the seat, softer at the executive office where the AI-prepared plan carries weight with the board and the CEO's day is also substantially refactored, but the seat stays. Some jobs in this transition will be obliterated. Some will be created. All are likely to be disrupted. But the facts suggest we can be optimistic about what's on the other side... if we do it right.
This paper has not solved the political-economy problem that turns Mode 2 from an available choice into the dominant pattern — what the firm's incentive structure, the executive's incentive structure, the regulator's incentive structure all do to the choice once the architecture is on the table. That problem is of a newly common sort. We'll know when we get there. And it will be soon. This one has demonstrated that the architecture is available at the speed of individual decision, that the tool itself is the transition mechanism, and that the missing piece is commitment, not capability. In other words, we have a real shot. We have the standing and the fortitude as a nation to carry this through responsibly. Not just California; not just any single administration; Big Business has the choice in front of it now.
China Wants Its Companies to Embrace AI — Without Firing Workers.The Wall Street Journal. Documents Chinese industrial policy explicitly directing AI augmentation rather than substitution at the firm level.
Inclusion of Clinicians in the Development and Evaluation of Clinical Artificial Intelligence Tools: A Systematic Literature Review.Frontiers in Psychology, April 7, 2022. Empirical finding that 82% (19 of 24) of clinical-AI design studies consult clinicians only at later stages of the design cycle. Cited as empirical anchor for the substitution-default critique. Verified URL: frontiersin.org/articles/10.3389/fpsyg.2022.848035/.