Why Cheaper Software Creates More Demand For Engineers

The US employment projection for software developers (BLS, 2024) puts growth at fifteen percent from 2024 to 2034 against three percent for all occupations, with roughly 129,000 openings a year — a projection published in full view of generative AI, and squarely at odds with two years of essays announcing the end of the profession. The projection is not a prediction that nothing changes. It is a prediction that the volume of software work rises while its content moves, and the second half is the part worth planning around.

The replacement argument treats software demand as fixed and supply as the only variable. When the cost of producing software drops by an order of magnitude, the question is not how much labor the current workload requires. It is how much new workload becomes economically viable that was previously priced out — and who is accountable for the systems that get built as a result. That question has a name in economics, dating to an 1865 paper on coal, and the answer has been the same across electricity, computing power, and bandwidth.

ML LABS has taken more than ten heavy-workload systems to production over fifteen-plus years, across healthcare, telecom, PropTech, and finance. What gets scarce when production gets cheap is not typing, and one engagement makes the point better than the argument does. A major US TV network came to ML LABS having already been quoted a full software system for a workflow that did not require one. The technology was sound and the build was buildable; nobody had asked whether the workflow needed custom software at all. The scoping session's deliverable was that it should not be built — which is the skill the cheap-software era makes scarcer, not more abundant.

The Mechanism Behind The Expansion

William Stanley Jevons observed that improvements in steam engine efficiency did not reduce coal consumption in Britain — they increased it. Cheaper coal per unit of work made coal-powered work viable where the prior price had made it uneconomic, so total use rose. Economists later formalized this as the rebound effect, and a survey of rebound and backfire across energy efficiency studies (Energy Policy, 2009) finds backfire well documented in productive uses where underlying demand is highly elastic.

Software production is a textbook case of elastic demand. The number of workflows inside any non-trivial organization that would benefit from custom tooling is large, and the reason most of them were never built is cost: engineering time is expensive, scoping is slow, and most candidate projects fail the budget review before anyone writes a line. AI-assisted development changes the unit economics of that backlog. A study on generative AI in customer support (NBER, 2023) found that access to an AI assistant raised resolved issues per hour by fourteen percent on average and by thirty-four percent for less experienced workers. When the production cost of a small system falls far enough, projects that sat permanently below the bar cross it — and they cross it in bulk, because the bar was the only thing holding them back.

Where The New Surface Appears

The projects this unblocks are not the ones the technology industry has been building. They are the ones non-technical enterprises have been deferring. A marketing team at a manufacturer that has wanted custom attribution reporting for years. A finance team at a regional insurer that has wanted reconciliation tooling matching how the books actually work. A clinical research group that has wanted its raw instrument output turned into analysis-ready datasets without a vendor procurement cycle. None of these clears the bar against a traditional engineering hire; the economics change entirely against an AI-assisted operator.

• Internal tooling at small and mid-sized businesses that bought off-the-shelf software and worked around the gaps.
• Line-of-business automation inside non-tech enterprises, where domain experts pair with engineers to digitize processes living in spreadsheets and email.
• Research and analysis code in regulated industries, where bespoke pipelines replace manual data wrangling.
• Customer-facing micro-products at companies whose core business is not software but whose customers expect a digital surface anyway.

No single one is large, and that is the point: the surface is composed of many small systems, each individually viable, each individually needing an owner. The 2025 AI Index (Stanford, 2025) reports that seventy-eight percent of organizations used AI in at least one business function in 2024, up from fifty-five percent the previous year. Adoption of that shape is the leading edge of the demand curve, not the end of it: something has to decide what those systems do, prove they do it, and keep them honest once they are load-bearing.

graph TD
    A["AI lowers cost to build"] --> B["Projects below old bar clear it"]
    B --> C["More systems run in production"]
    C --> D["More review, eval, and ops work"]
    D --> E["Demand rises for skilled operators"]
    E --> F["Operators open further projects"]
    F --> A

    style A fill:#1a1a2e,stroke:#16c79a,color:#fff
    style B fill:#1a1a2e,stroke:#16c79a,color:#fff
    style C fill:#1a1a2e,stroke:#16c79a,color:#fff
    style D fill:#1a1a2e,stroke:#0f3460,color:#fff
    style E fill:#1a1a2e,stroke:#16c79a,color:#fff
    style F fill:#1a1a2e,stroke:#ffd700,color:#fff

Each turn of that loop changes what an engineer is paid for. Less of the work is keystroke production and more of it is judgment about what the system should do, evidence that it is doing it, and recovery when it is not. Prompt design becomes a real specialty, because the same model produces wildly different output depending on how the task is framed and what context is loaded. Evaluation design becomes load-bearing, because a change a model accepts as correct is not thereby correct, and without a regression suite the team is flying blind. Integration work gets no cheaper at all — a model does not know which internal service owns which data. And on-call becomes a different job, because an AI system fails in ways that look like normal latency right up until it has taken an action nobody wanted. A framework for productivity gains from general-purpose technologies (NBER, 2018) describes the J-curve underneath this: visible output lags the intangible investments that make a technology durable, and the intangible investment here is the operating capability, not the agents.

The work that scales is not writing code. It is deciding what to build, proving it works, and keeping it honest in production.

When The Expansion Stalls

The thesis has a real boundary. An organization already digitized end to end, where the marginal internal task is already cheap, has no deferred backlog for the rebound to reach; cheaper production raises the floor on existing work without opening a new layer of projects. The same holds where every new system carries a regulatory lift heavy enough to swamp the production savings, because the binding constraint there is validation and audit rather than build time.

The other boundary is organizational, and it is the one that bites hardest. A company that cannot articulate which problems it wants software to solve does not gain capacity from a cheaper way to write software — it generates more half-built systems that nobody owns, which is the failure the build-versus-buy decision is supposed to catch before the money moves. The rebound depends on someone knowing which projects were previously deferred, being able to sequence them, and having the authority to say that a candidate on the list should not be built at all.

First Steps

1. Audit the functions that historically failed budget review. Walk finance, operations, marketing, and customer success, and list every internal tool that was scoped and shelved. That list is the addressable surface, and it already exists.
2. Separate acceleration from viability. For each candidate, ask whether AI-assisted development makes an existing initiative faster or makes a previously uneconomic one possible. The second category is where the rebound concentrates value, and it is the category nobody has a budget line for.
3. Make one role accountable for AI-built tools in production. Name the operator who owns evaluation, on-call, and integration health across the new systems. Without that role, the cheaper-to-build advantage is spent on cheaper-to-fail outcomes.

The Bar This Era Sets

The bar is not generalist coding fluency, and it is not model expertise either. It is the operating discipline that carries an AI-assisted build from a working demo to a system behaving correctly in production, and it is short enough to hold your own organization against right now. Someone can name which deferred projects are now viable and in what order. Every system that ships has one accountable owner. Every system has an evaluation that would catch it being wrong. And someone in the room has the standing to say that a proposed build should not happen — the decision that returns the entire budget.

"Omar delivered in two weeks what our team estimated would take six months. The scoping session alone saved us from a $200K mistake." — AI Program Manager, a major US TV network

An organization that clears that bar treats the expanded surface as a sequencing problem, which is what it is. Sequencing the backlog is a discrete decision that can be bought on its own: an AI scoping session turns one candidate into a written recommendation — scope, delivery path, budget range, and a go or no-go — for $750 credited against the work if the work goes ahead, and as the network above found, the recommendation is sometimes that the cheapest possible build is still the wrong one. For a first candidate rather than a queue of them, how to find the one worth doing is the companion piece.

References

1. Sorrell, S. Jevons' Paradox revisited: The evidence for backfire from improved energy efficiency. Energy Policy, 2009.
2. Brynjolfsson, E., Li, D., and Raymond, L. Generative AI at Work. NBER Working Paper Series, 2023.
3. Brynjolfsson, E., Rock, D., and Syverson, C. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. NBER Working Paper Series, 2018.
4. U.S. Bureau of Labor Statistics. Software Developers, Quality Assurance Analysts, and Testers. Occupational Outlook Handbook, 2024.
5. Stanford HAI. The 2025 AI Index Report. Stanford Institute for Human-Centered AI, 2025.