On an equal budget

The meeting almost always unfolds the same way. The demo impresses, the teams get excited, and then someone — usually the CFO's office, sometimes procurement — asks the question that seems self-evident: “so, for the same scope, how much can we expect to save? Twenty percent? Thirty?” The question is legitimate, the reasoning is logical, and I spend part of my job explaining why it is, strategically, wrong. This essay is the long version of that explanation.

A question inherited from another world

In 2003, Nicholas Carr published a now-famous article in the Harvard Business Review, “IT Doesn't Matter”, defending a simple idea: once a technology becomes available to everyone, it ceases to be a competitive advantage and becomes a commodity — and the only rational strategy is then to minimize its cost^[1]. That is exactly the reasoning finance departments apply to AI today, and they deserve credit for its consistency. But Carr was writing about infrastructure technology — electricity, rail, the server — whose use is undifferentiated. Software, however, is no longer just another piece of infrastructure: as Marc Andreessen diagnosed back in 2011, it has eaten the world and become the product itself, the distribution channel, the customer experience, the competitive edge^[2]. Applying commodity cost logic to the software that differentiates you is mistaking the generator for the light.

And there is something more awkward still: even on Carr's own turf, the argument turns against itself. A cost advantage only lasts if it is exclusive. Agentic AI is available to all your competitors, at the same price, at the same time, often from the same vendor. If the entire market pockets the same productivity gain, nobody has gained anything — except those who converted it into something other than savings. The saving is real on paper; it is paid for in market share.

Fixed scope does not exist

The “same scope, cheaper” reasoning rests on an assumption so familiar we no longer see it: that software has a finite, known scope, which merely needs to be delivered for less. Forty years of software engineering demonstrate the opposite, and the Defense Innovation Board at the Pentagon — hardly suspect of agile romanticism — made it the very title of its 2019 report: Software Is Never Done^[3]. A living product has an effectively infinite backlog: technical debt, usability, integrations, security, data, documentation. The constraint has never been the stock of ideas; it is the capacity to build. Agentic AI loosens precisely that constraint. Cutting the budget at the exact moment it loosens is tightening it on purpose.

Besides, we have known for over a century and a half what happens when the unit cost of a resource collapses: we do not consume less of it, we consume vastly more. William Stanley Jevons observed it as early as 1865 with coal — every efficiency gain in steam engines increased total consumption instead of reducing it^[4] — and his paradox has become tech leaders' favorite argument for predicting the explosion, not the contraction, of software demand in the AI era^[5]. The consequence is concrete for any budget committee: while you deliberate over your 20% cut, more software is being produced everywhere, and your customers recalibrate their expectations against what the market shows them — not against what your budget allows. The bar rises, whether you invest or not.

Running to stay in place

In Through the Looking-Glass, the Red Queen explains to Alice a rule of her strange country: there, it takes all the running you can do to stay in the same place^[6]. Evolutionary biologists turned it into a law — in Leigh Van Valen's formulation, a species that stops adapting within an adapting ecosystem is eliminated at a constant rate^[7] — and I know of no better description of industrial software in 2026. When windows of opportunity close within months, coming in second often costs more than the entire development budget. The first use of the agentic dividend is therefore defensive as much as offensive: at minimum, do not fall off the market's pace; at best, set it. The saving you pocket, meanwhile, is speed handed back to your competitors.

Where the dividend goes

At constant budget, the agentic gain can be reinvested in three directions, and they reinforce one another. Speed, first — compressed cycles, parallel exploration of options, prototyping measured in days. Quality, next, and it is perhaps the most underrated reinvestment: it has always been the adjustment variable of projects under constraint, the one sacrificed by cutting tests and deferring debt. Yet the DORA research, the largest empirical study ever conducted on software team performance, established a counter-intuitive result: speed and stability are not traded against each other, they go together — the best organizations ship both faster and more reliably^[8]. Quality is a form of speed: less rework, fewer incidents, fewer regressions. This quarter's quality investment shows up in next quarter's time-to-market.

The third reinvestment carries the highest leverage, because it bears on the scarcest resource. That resource is not the ability to write code: it is judgment — knowledge of the business, the stakes, the customers. And yet we have our best experts spend most of their time on implementation, where their judgment is least differentiating. As agentics takes over a growing share of the building, expert time can move back up toward requirements, trade-offs, architecture, product strategy. Machines build, experts decide. Provided, of course, that this freed-up time is not immediately reabsorbed by production or handed back to the budget — it must be ring-fenced, like any investment.

Measuring without fooling yourself

Which leaves the most serious objection: “prove it”. It deserves better than a marketing number, because the history of information technology is littered with productivity promises that were slow to materialize. “You can see the computer age everywhere but in the productivity statistics,” quipped Robert Solow in 1987^[9]; Erik Brynjolfsson and his co-authors have since shown that this paradox follows a J-curve — the gains only arrive after the complementary investments, in organization, processes, skills, and are even preceded by a dip^[10]. Translated for our subject: handing out licenses and hoping is the surest way to pay for the technology without ever touching the gain. The agentic dividend cannot be decreed; it is industrialized — versioned, replayable workflows, systematic verification gates, activity measurement, tooling capitalization so that every project accelerates the next. That is what I call an agentic factory.

And it is measured honestly, which means being wary of both extremes. The most-cited controlled studies measure spectacular gains on standardized tasks — 55% less time on a benchmark exercise with GitHub Copilot^[11] — while a randomized trial by METR on seasoned developers working in their own repositories measured the opposite: 19% slower with AI, while believing themselves 20% faster^[12]. The lesson is not that AI does not work; it is that perception is a broken thermometer, that context is everything, and that the dividend must be steered with value indicators — time-to-market, test coverage, debt paid down, value delivered per euro — stated with their explicit limits, not with theoretical savings or declarative enthusiasm.

Cut here, reinvest there

Should we conclude that cost reduction never has its place? No — it has one, in the right spot. Every application portfolio contains software that does not differentiate the company: commoditized run, support systems, off-the-shelf applications. There, Carr is right, and AI is a perfectly legitimate efficiency lever. The strategic mistake is not seeking efficiency; it is applying across-the-board cost logic to the software that carries your differentiation. The right answer to a finance department fits in one sentence: cut where software does not set you apart, reinvest where it does. Sorting one from the other is, in fact, a more strategic conversation than any debate about tooling.

Within a few years, almost every organization will be using agentic AI in delivery — the technology will set no one apart, and that is precisely where this whole essay begins. What will set them apart is the question asked upstream. Those who ask “how much will AI save us?” will optimize the past. Those who ask “what will we do with the capacity it frees up?” will build tomorrow's lead — more value, better quality, experts in the right place, at the same cost. On an equal budget, that is where everything is decided.