AI for Subtraction, Not Addition

By Martin Bergmann | AI Project Lab | September 2026

Every shared drive already has a folder like this: a file, a “final” version of that file, and a “final, actually final” version after that, and nobody agrees which one is current. That habit did not start with AI. What has changed is how quickly the next version shows up and how little it costs to make a new one rather than to fix the one that already exists.

The cost AI didn’t remove.

AI made producing things nearly free: code, reports, documentation, and process. It did not make owning and maintaining those things free. Each new report, script, or workflow looks reasonable in isolation, a small time-saver, an easy ask. What accumulates is not obvious until someone tries to add it up.

Steve Yegge has been building an AI agent management system he calls Wheelhouse to run the coding agents working on his long-running video game project. By his own account, Wheelhouse has grown to roughly 600,000 lines of code and 450 artifacts governing what the agents are allowed to do. Maintaining it now consumes 20 to 25 percent of his total effort, with no sign of shrinking. His own description: the factory is growing faster than the game.

The pattern shows up at scale, too. A 2025 Faros AI study tracking over 10,000 developers across 1,255 teams found that teams with high AI adoption completed 21% more tasks and merged 98% more pull requests than teams with low adoption. But review time on those pull requests rose by 91%, the pull requests themselves grew by 154%, and bug counts rose by 9%. The output did not disappear. It moved downstream, into a growing queue of larger, slower-to-review, buggier changes that somebody still has to deal with.

It shows up in writing, too. A September 2025 survey of 1,150 US desk workers by BetterUp Labs and Stanford’s Social Media Lab found 40% had received AI-generated “workslop” in the past month, content that looks polished and complete but is actually unhelpful or wrong. Workers estimated that on average, 15.4% of the work they receive fits that description, and BetterUp modeled the resulting cost at roughly $9 million a year in lost time for a 10,000-person company.

None of this means AI made things worse. It means AI removed the friction that used to slow down addition, and did nothing to the friction that governs subtraction. Those were never the same cost, and treating them as if they were is the actual mistake.

Counting the wrong thing.

Most AI adoption dashboards track consumption: prompts sent, tokens used, active users, and tools deployed. Consumption is easy to count and easy to grow. It rewards building more, not needing less. Subtraction, retiring a report nobody reads, killing a script three other scripts already do, collapsing two dashboards into one, is harder to see and almost never shows up as a metric anyone tracks, even though it is the more honest signal that something is actually working.

Where AI actually helps.

Point AI at the reports, dashboards, and recurring documents your team maintains and ask it two questions: which of these has nobody opened in the last month, and which ones say close to the same thing as something else on the list. That inventory is the addition AI quietly helped create.

Then ask the harder question: for each one it flags, what would actually be lost by deleting it. If the honest answer is nothing, that is the subtraction nobody has been counting, and the tool that made the mess is also the fastest way to find your way out of it.

AI did not just make it easier to build. It made it easier to stop. Most teams have only been using it for the first one.

---

Martin Bergmann is the Director of an IT PMO and founder of the AI Project Lab. He writes about AI, project management, and the future of intelligent work.