The Unowned ArtifactCode, craft, and the reckoning that Wall Street is not pricing

For half a century, the software industry has rested on a legal fiction so successful that most practitioners never noticed it was a fiction. Source code is literature. It is a work of authorship, fixed in a tangible medium, protected the moment it is written. That single doctrinal move — treating a set of machine instructions as an expressive work rather than a functional device — gave software companies a cheap, automatic, durable, and litigable moat. No registration required to hold the right. Statutory damages and attorney's fees available to those who registered. It is the reason a startup could raise money on a repository.

That moat is draining, and the drain has a legal name. In January 2025 the U.S. Copyright Office published Part Two of its report on copyright and artificial intelligence, and reaffirmed what it had been saying since 2023: human authorship is a prerequisite for protection, and material generated purely by a machine is not copyrightable.1 The Office went further on the question every engineer wants answered. Prompts, it concluded, do not by themselves confer enough control over the expressive elements of an output to make the prompter an author — however long, however detailed, however many iterations.2

Two months later the D.C. Circuit affirmed the principle in Thaler v. Perlmutter, holding that the Copyright Act of 1976 requires a work to be authored in the first instance by a human being.3 On 2 March 2026 the Supreme Court declined to hear the appeal, leaving that holding in place and unreviewed.4 The question is now settled at the level that matters for planning: a machine is not an author, and what a machine alone produces belongs to no one.

The Office is careful to say this does not doom AI-assisted work. Where a human selects, arranges, modifies, or contributes perceptible expression, that human contribution is protectable, and the Office has registered hundreds of works containing AI-generated material on exactly that basis.5 But read that carefully as an engineering manager rather than a lawyer. The protection attaches to the human increment. In a codebase where an agent generated ninety-eight percent of the lines, the copyrightable estate is the two percent — plus whatever selection and arrangement a court will credit — and proving which two percent requires contemporaneous records that almost nobody keeps.

The asset is not gone. It has moved, and it has moved somewhere far less convenient to finance, audit, and litigate.

The practical consequence is a collapse in barriers to entry. If a competitor cannot be sued for the code, the code stops being the wall. Bloomberg Law's analysis put the strategic conclusion plainly: architectural and orchestration artifacts may be the new moat in the agentic era, and the most valuable of them may be worth holding as trade secrets rather than exposing through registration.6 Other practitioners have reached the same place from a different door: as generated code becomes the norm, advantage migrates to proprietary prompting and evaluation strategy, fine-tuning corpora, integration patterns, data flywheels, and sheer execution velocity.7

Each of those substitutes is weaker than the thing it replaces. Trade secrets die on disclosure and offer no remedy against lawful reverse engineering — a defense that has grown noticeably thinner now that a competitor can point a model at your API surface and infer your architecture from its behavior.8 Patents demand eligibility, expense, and publication. Contracts bind only the counterparty. Velocity is not a right; it is a treadmill.

So the first sea change is this: the software industry is being quietly converted from a property business into an operations business. You no longer own a thing. You maintain a lead.

The second foundation is older and more consequential, and it is the one I expect to reshape more lives. For thirty years, enterprise computing has been financed by a wage differential. The offshore delivery model is, stripped of its language about follow-the-sun and centers of excellence, an arbitrage: the same unit of codified work, executed by a person paid a fraction as much. Every layer built on top of it — the pyramid staffing model, the T&M contract, the utilization target, the fresher intake of ten thousand a quarter — is a mechanism for converting that gap into margin.

Arbitrages close. This one is closing not because wages converged but because the good being arbitraged stopped being scarce.

The numbers from the industry's own trade body tell the story without editorial help. Indian IT services revenue reached roughly $315 billion in FY2026, growing 6.1 percent, while headcount grew 2.3 percent to about 5.95 million — a divergence that Nasscom's own president has acknowledged from the stage.9 Headcount has ceased to be the leading indicator of revenue, which is another way of saying the pricing model has lost its physical basis.10 Infosys shed 8,440 employees in a single quarter of FY2026 while reporting growth, and told analysts that AI-related work had reached 8.2 percent of revenue.11

Markets have noticed. In June 2026 renewed fears about AI's effect on the services model knocked TCS down nearly eight percent in a session and dragged the Nifty IT index down close to five percent.12 Sell-side analysts have put thirty to forty percent of IT services revenue at risk from AI-led price deflation, concentrated precisely where the boom was built: application development, maintenance, and testing.13

Layered on top of the technological shock is a political one. The HIRE Act, introduced as S.2976 in October 2025, would impose a 25 percent excise tax on payments to foreign persons for labor benefiting U.S. consumers and would disallow the deduction for those payments.14 The arithmetic is brutal because it is doubled: a $100 offshore payment that effectively costs about $79 after deduction would cost $125 — an increase near 58 percent.15 The bill has gone nowhere; it sits in the Finance Committee without cosponsors, having been blocked on unanimous consent, and most tax practitioners treat it as a messaging vehicle rather than imminent law.16

I would not bet on its passage. I would bet on what its existence signals. A model whose economics are eroding from below, by automation, and which is simultaneously exposed to a live political impulse to tax it from above, is not a model to build a thirty-year career inside. The offshore delivery firms will survive — they have survived Y2K's end, the dot-com bust, and the cloud transition — but they will survive as something other than labor brokers, and the transition will be paid for by the people whose value proposition was availability at a price.

The commodity programmer was never selling code. He was selling cheap, and cheap has been undercut by free.

Here is the distinction that decides who survives the next decade, and it is not new. Fred Brooks drew it in 1986: the accidental complexity of software is the labor of expressing a solution in a machine's terms, while the essential complexity is the conceptual work of specifying what the thing must do and how its parts must relate.17 Every productivity revolution in this industry — assemblers, compilers, fourth-generation languages, frameworks, cloud primitives — has attacked accidental complexity and left the essential complexity untouched. Large language models are the most powerful assault on accidental complexity ever built. They are, so far, indifferent to the essential kind.

A program is an answer to a stated question. A system is a structure of commitments: about failure, about state, about who is accountable when the answer is wrong at three in the morning. Building programs is a task. Building systems is a judgment. The industry's commodity tier was staffed, trained, and measured to do the first. That is exactly the tier that the machine now does faster, cheaper, and without complaint.

You do not have to take my word for the distinction. Mark Russinovich, CTO of Microsoft Azure, and Scott Hanselman put it in a peer-reviewed opinion piece in Communications of the ACM in April 2026, and their formulation is exact: programming is not software engineering.18 Their examples are the ones every senior engineer will recognize. An agent handed a race condition inserted a sleep call — a masking fix that leaves the synchronization defect intact and the test suite green. Agents that report success over real bugs. Agents that duplicate logic across a codebase, dismiss crashes as out of scope, and implement special cases that pass the tests and fail in production. An experienced engineer catches all of these in seconds. A junior may not catch them at all. The faculty that does the catching — what the authors call systems taste — is precisely what is acquired by doing production work, and precisely what is no longer being acquired.

The measurement literature is more ambiguous than either camp admits, and it is worth being honest about that. METR's randomized controlled trial — sixteen experienced open-source maintainers, 246 real tasks in repositories they knew well — found that allowing early-2025 AI tools made them 19 percent slower. The forecast had been 24 percent faster. After finishing, the same developers still believed they had been sped up by 20 percent.19 That perception gap is the durable finding. METR itself has since redesigned the experiment, reporting that selection effects made its follow-up an unreliable estimate of current uplift, while noting some raw evidence of speedup with newer tools.20 Read honestly, this is not proof that AI slows people down. It is proof that the industry has been steering by feel.

The maintainability data is less equivocal. GitClear's analysis of 211 million changed lines found that "moved" code — its proxy for refactoring and consolidation — fell from about 24 percent of changed lines in 2021 to 9.5 percent in 2024, while copy-pasted lines rose from 8.3 to 12.3 percent. In 2024, for the first time in the dataset, cloning exceeded refactoring, and blocks of five or more duplicated lines rose eightfold.21 The 2026 follow-up, tracking seven signals across 2023–2026, reports refactoring line moves down 70 percent, cross-file function calls down 35 percent, legacy maintenance down 74 percent against 2022, and rises in duplication, error-masking constructs, and two-week churn.22 Google's DORA research found a similar shape from the other side: AI adoption correlating with a small improvement in perceived code quality alongside an estimated reduction in delivery stability.23

Put the two findings together and a picture forms that is neither the boosters' nor the doomers'. Throughput is real. So is the debt. The machine produces more code per hour and less structure per line, and structure is the thing that makes a system cheap to change in year three. Which is to say: the tools are extraordinarily good at producing programs and indifferent to producing systems. That is not a defect to be patched. It is a description of where the human work went.

The capital markets have made an enormous, coherent, and possibly correct bet on the infrastructure layer. Goldman Sachs Research puts global AI-related investment at roughly $1 trillion in 2026, with just under $600 billion of it in the United States, and cumulative investment since 2022 approaching $1.8 trillion.24 J.P. Morgan Asset Management has tracked sell-side 2026 capex estimates for five U.S. companies to nearly $700 billion.25 Whether that is a bubble is the wrong question to argue in public; Goldman's own framing is useful discipline, noting that AI capex has been running near 0.8 percent of GDP against peaks above 1.5 percent in prior technology booms.26 Big, then, but not yet historically unprecedented as a share of the economy.

The more interesting story is not the inflation at the infrastructure layer. It is the deflation at the application layer — and that is the part the headlines keep missing.

PitchBook's Q1 2026 comp sheet — the one that gave the episode its nickname, the SaaSpocalypse — recorded the median enterprise value to trailing revenue multiple across 99 public enterprise SaaS companies falling to 3.3x, from 4.9x at year-end 2025 and 6.2x at year-end 2024.27 That is not a rate-driven repricing of the kind we saw in 2022. It is a structural reassessment of a business model, and the specific question being repriced is whether per-seat pricing survives when an agent performs the work that used to require a seat. Gartner's forecast that 40 percent of enterprise applications will embed task-specific agents by the end of 2026, up from under 5 percent in 2025, is the mechanism stated as a schedule.29

Meanwhile the buyers of enterprise AI are not, in aggregate, getting what they paid for. MIT's Project NANDA reported in mid-2025 that against an estimated $30–40 billion of enterprise generative AI spending, roughly 95 percent of organizations were seeing no measurable P&L return, and only about 5 percent of custom enterprise AI tools were reaching production at all.30 The number has been criticized, fairly, on methodology — a preliminary, non-peer-reviewed study with a small interview base and a short measurement window.31 But the direction is corroborated elsewhere: S&P Global found the share of firms abandoning most of their AI initiatives rising to 42 percent from 17 percent a year earlier, with the average organization scrapping nearly half its proofs of concept before production.32

Stack those three facts. Unprecedented capital flowing into the infrastructure that serves models. Multiples collapsing on the applications that were supposed to consume them. Adoption without transformation in the enterprises that were supposed to pay for both. That is not a picture of an industry being uplifted. It is a picture of value being violently relocated — down into compute and up into whoever can actually make a system out of it — while the middle, the enormous middle where enterprise computing has employed most of its people, is squeezed from both ends.

The demise of enterprise computing as we practiced it is the larger sea change, and it is arriving on a slower clock than the capex boom that obscures it.

Which is worth more: the person who builds AI systems, or the person who runs them? I have gone back and forth on this, and I have come to think the question is badly posed in a way that matters.

Start with the honest definitions. The AI developer designs the thing that makes a model useful: the retrieval architecture, the tool surface, the memory model, the state machine, the evaluation harness, the failure boundaries. The AI operator keeps that thing alive and pointed at the business: monitors trajectories, tunes context, adjudicates escalations, catches drift, decides what the agent is allowed to touch this quarter. In classical enterprise terms the first is an architect and the second is a production manager, and we spent forty years teaching those two groups to distrust each other.

The market is currently paying both, under new names. Roles that did not exist in hiring pipelines two years ago — agentic AI engineer, context engineer, applied AI engineer, agentic systems architect — appeared in enterprise postings through late 2025 and Q1 2026 at Anthropic, Salesforce, and the large consultancies.33 The content of those roles is telling. The discipline that has displaced prompt engineering is not about phrasing; it is about designing the entire information environment the model works inside: retrieval, compression, memory, tool definitions, state, and the token budget that constrains all of it.34 The hardest skill listed in these postings is not generation. It is evaluation — building harnesses that prove an agent actually completed the task rather than produced a plausible-looking artifact.35

That is the tell, and it resolves the question. Value accrues where verification and accountability live. A model can generate; it cannot yet own the consequences of generating wrong. The scarce good is not the ability to produce output, nor the ability to babysit it, but the ability to specify what correct means in a domain and then build the machinery that proves it. Development and operations converge on the same person because in an agentic system the specification, the evaluation, and the production behavior are the same artifact viewed at three moments.

So: neither is more valuable. The orchestrator — the person who holds the specification, the architecture, and the evidence in one head — is more valuable than either, and that person is not produced by a bootcamp or by a certification. Which brings us to the problem nobody in a public company wants to price.

Judgment is grown, not hired. It is grown by doing work that is consequential enough to teach and small enough to survive. That rung of the ladder is being sawn off, and the sawing is measurable.

The Stanford Digital Economy Lab's work is the best evidence we have, and its authors are notably careful about what it does and does not show. Using ADP payroll data, they found that since the diffusion of generative AI, workers aged 22 to 25 in the most AI-exposed occupations — software developers prominent among them — experienced roughly a 13 percent relative decline in employment, while employment for experienced workers in the same occupations held up or grew.36 The August 2026 revision, with data through mid-2026, widened that gap to about 19 percent, driven primarily by reduced hiring rather than increased separations.37

The mechanism the authors identify is the one that should keep engineering leaders awake. The declines concentrate among younger workers whose contribution rests on codified knowledge — the kind that has been written down, and therefore the kind a model has read. Experienced workers in the same occupations show no comparable decline, because what they sell is tacit: judgment accumulated by doing, which Michael Polanyi described sixty years ago as the class of knowledge we possess without being able to fully state.38 Declines also concentrate where AI usage substitutes for human tasks rather than complementing them.36

Russinovich and Hanselman name the resulting structure the narrowing pyramid. AI confers a large advantage on senior engineers and imposes a drag on early-career ones who have not yet built the judgment to steer and verify it. The rational firm therefore hires seniors and automates juniors — and the pipeline that manufactures seniors quietly collapses.18 Their proposed remedy is borrowed from medicine: a preceptorship, a year or longer, in which an experienced engineer and an early-career one work with the tools together, and the senior's job is not to answer questions but to teach judgment — with mentorship measured and compensated as a real deliverable rather than volunteered on nights and weekends.39

It is the right answer. It is also, for a public company, an almost impossible one.

Consider what a preceptorship asks a quarterly-reporting organization to do. Hire people who will reduce delivery capacity for twelve months. Assign your most productive engineers to spend a material share of their time not producing. Book the cost this year and the benefit in year five, by which time the CTO who authorized it has been replaced twice and the analyst who punished the margin has moved to another sector. The arithmetic of a five-to-seven-year apprenticeship is simply not expressible in the language of quarter-to-quarter earnings guidance. This is not a failure of intelligence in the C-suite. It is a structural mismatch between the clock speed of capital markets and the clock speed of expertise.

Every firm has an incentive to hire the trained engineer and none to train one. That is a textbook collective-action failure, and markets do not solve those on their own.

We have already watched one high-profile round trip. Klarna went AI-first with public enthusiasm, cut headcount hard, celebrated the savings — and by 2025 was recruiting humans again, having discovered what quality costs when you remove the people who knew what good looked like.40 A handful of firms are betting the other way; IBM reportedly tripled its junior intake while redefining the role toward interpreting requirements and validating machine output, on the explicit thesis that whoever keeps the pipeline open wins in three to five years.41

Meanwhile there is a second-order risk that deserves more attention than it gets. If a cohort learns to accept output it cannot evaluate, the population of engineers capable of producing genuinely novel, well-structured code thins — and that population is the source of the training data the next generation of models learns from. Add MIT's finding that people who outsourced writing to a model showed weaker recall and engagement than those who worked unaided, which the researchers labeled cognitive debt, and the loop is not hypothetical.18 We would be automating the production of the thing we are consuming.

I have been in this industry long enough to have been confidently wrong in public, so let me put the strongest version of the opposing case here rather than leave it to the comments.

I find the counterarguments to the productivity literature the most persuasive of these, and the counterargument on the training pipeline the least — because the lag is the whole problem, and a five-to-seven-year lag is long enough to hollow out a profession before the price signal arrives. But a reader who weighs them differently is not being unreasonable.

Knuth called it The Art of Computer Programming, and he chose the word deliberately. He meant art in the older sense — a practice whose mastery is demonstrated rather than certified, whose standards are internal to the craft, and whose best work has an elegance you can recognize before you can justify.42 That framing has never been more load-bearing than it is now, and for a reason that is almost too neat.

Copyright protects expression, not ideas, and it requires a human. As machines absorb the expression, what remains for a human to own is the conception: the architecture, the specification, the criteria of correctness, the taste that recognizes a solution as sound before the tests confirm it. The law is, by accident, pointing at exactly the same thing the engineering evidence points at. The unprotectable part and the automatable part are the same part. The protectable part and the scarce part are the same part.

So the orchestrator is not a diminished programmer. The orchestrator is the author in the sense the statute always meant — the human whose control over the expressive and structural choices makes the work a work at all. Whether courts will credit that in a codebase is genuinely unsettled and will be litigated for years. But the strategic instruction is clear regardless of how the doctrine lands: build the practice of documenting human architectural contribution into the workflow now, because it is simultaneously the copyright record, the trade-secret boundary, and the training curriculum for the next generation.6

The tools got better at the part I could teach in a month, and no better at the part that took me thirty years.

I learned this trade on a Burroughs B-1900 on a hospital night shift, at sixteen, because nobody senior wanted the shift. That is how it worked: you were handed a small consequential thing at an hour when the damage would be survivable, and you learned what production felt like when it went wrong at two in the morning with nobody to call. Every senior engineer I respect has some version of that story. None of them has a version where they read about it.

The industry is currently in the process of deciding, mostly by default and mostly through the accumulated weight of individually rational quarterly decisions, whether that shift still exists for anyone. If it does not, we will find out in about seven years, and the finding out will be expensive.

For those of us who intend to be here for the whole ride, the practical program is not complicated, though it is unfashionable. Build systems, not programs — hold the specification and the evaluation, not the keystrokes. Treat the architectural record as an asset, because legally and commercially it now is one. Assume no moat you cannot rebuild in a quarter. Take on an apprentice even when the arithmetic argues against it, because the arithmetic is measuring the wrong interval. And be suspicious of your own sense of speed; the one thing the evidence agrees on is that we are poor judges of whether these tools are helping us.

It will be a wild ride. It usually is. But the thing I would tell a twenty-two-year-old is the thing I would have told one in 1985, which is that the tooling always changes and the discipline never does. The machine has become a superb apprentice. It has not become a master, and mastery — patient, expensive, embodied, slow — is the only asset in this business that nobody has yet found a way to copy.