The Economics of Software Changed

For thirty years the rational decision was to adapt the company to the software, because customization was expensive. That constraint is the thing that just moved.

The first thing AI changed was not the enterprise. It changed the economics of creating software.

For the first time, we are moving toward a world where software can be created, modified and customized at a radically lower marginal cost. The early measurements point in the same direction. Three field experiments inside Microsoft, Accenture and a Fortune 100 company, covering 4,867 developers, found that developers with an AI assistant completed about 26% more pull requests each week. On GitHub, more than a million public repositories now import an LLM SDK, 178% more than a year earlier. The gains are not uniform and they are not automatic. But the direction is not really in dispute.

Software is becoming malleable

That matters much more than the fact that AI can write code. Code generation is just the visible part. The deeper change is this: software is becoming malleable.

A traditional enterprise application has to make assumptions about how a business works. AI makes it increasingly practical to build applications around the specific way a business actually works. That reverses the old relationship.

For thirty years: company adapts to software. Now we can increasingly have: software adapts to company.

This is not a small improvement to enterprise IT. It changes what enterprise software can be.

Imagine an enterprise where the software stack is no longer a collection of rigid applications that employees have to navigate. Imagine instead that the company can describe how it wants a process to work. The software can be shaped around that process. The interfaces can change. The workflow can change. The logic can change. The application can change. The intelligence can change. The integrations can change. The system becomes much closer to the company's actual operating model.

That is the first unlock. But it is only half of the story.

Software that does the work

Because AI did something else. It did not just make software easier to build. It made software capable of doing work.

For decades, software was mostly something humans operated. You opened the system. You found the record. You entered the information. You selected the workflow. You approved the action. You moved to the next system.

Now we are building systems that can increasingly understand an objective and execute the steps required to achieve it. Read the information. Look something up. Reason about the situation. Make a recommendation. Call another system. Update a record. Send a message. Escalate an exception. Ask a human when necessary. Continue.

This is measurable too. On software tasks, the length of work an AI agent can complete with even odds has been doubling roughly every seven months since 2019. And in Anthropic's telemetry of a million enterprise API conversations, 77% already follow automation patterns. The task is handed over whole, not assisted.

An operating system for the business

This is the second unlock. AI can build software. AI can operate software. And when those two things come together, something much bigger becomes possible. We no longer have to think about enterprise software as a set of applications. We can start thinking about it as an operating system for the business itself.

But that creates a new problem. And this is where most of the AI conversation is still too shallow. That is the next part.

Sources

• The Effects of Generative AI on High-Skilled Work. Cui et al., Management Science 2026. Three RCTs at Microsoft, Accenture and a Fortune 100 firm, 4,867 developers: +26% completed pull requests, concentrated in junior and short-tenure developers.
• Octoverse 2025. GitHub platform telemetry. 1.1M+ public repositories importing an LLM SDK, +178% year over year. Hard counts, vendor-published.
• The Impact of AI on Developer Productivity. Peng et al., 2023. The famous 55.8% speedup: a single fully-specified greenfield task, GitHub-authored. The upper bound, not the enterprise estimate.
• Meta-Analysis of Generative AI Effects on Programming Productivity. Maier et al., LMU Munich, 2026. Pooled effect positive in the lab, statistically indistinguishable from zero in enterprise settings. The boundary of this essay's claim.
• Measuring AI Ability to Complete Long Tasks. METR, 2025. The 50%-reliability task horizon has doubled roughly every seven months since 2019, on software tasks.
• Anthropic Economic Index, September 2025. API telemetry, ~1M transcripts. 77% of enterprise API conversations follow automation patterns.
• Stack Overflow Developer Survey 2025. Usage rose to 78.5% while trust in accuracy fell to 32.7%. The first year distrust exceeded trust.