A Boring 100 Percent
The year should not end on a rant, so here is my catharsis: a 100 percent productivity gain with AI is possible outside greenfield projects and PoCs. I know because we did it.
The metric
The metric is crude, which makes it hard to game.
We do not count lines of code, pull requests or tickets closed. We do not report DORA metrics. We look at two numbers only: how many product services our team owns, and how many people are on the team.
| Change in 2026 | |
|---|---|
| Services owned | +50% |
| People on the team | -30% |
That is a productivity increase of more than 100 percent.
Owning a service means you run it, fix it at night and ship its features, and nobody asks how many PRs that took. The only way to improve the number is to take on more real work with the same or fewer people.
No greenfield
Ours is a decade-old enterprise application that serves hundreds of thousands of customers. It carries all the warts you would expect after ten years and many teams: dozens of repositories and services, each with its own history and conventions.
That matters. A greenfield project is an easy place for an LLM to shine. Our setup is the opposite, and it is the reality most of us work in. If a gain shows up there, it is not a trick of circumstances.
What we skipped
For most of the year it did not feel like success. It felt slow.
Colleagues in other teams wrote entire MCP servers and agent orchestrators, wired LSP into their agents, tuned harnesses and tried dozens of new skills the week they appeared. We did almost none of that, and at times it was a little embarrassing to sit in the meeting where everybody showed their latest setup in pursuit of AI-native enlightenment.
What we did was boring. We worked through our backlog, took over more services and shipped our feature increments. We used every new LLM as it arrived, but we never asked our Product Owner to slow down while we figured out AI.
I do not claim the others are wrong. Some of it might pay off, and maybe we will copy some of it next year. It is just not what we think creates the business value we are being paid for.
Our setup
Our setup is short: mostly GitHub Copilot with Opus, plus a tiny handful of skills we wrote for recurring tasks in our own codebase. That is all.
Beyond that, everybody finds their own way of working with the agent. Time spent learning how to work with it pays off better than time spent building tooling, hooks and skills around it. All of that becomes obsolete with the next model generation anyway.
The rule for anything new is just as short: if it works without fuss, it goes into our standard workflow. If it needs fuss, we drop it.
Everything else stayed as it was. Every change goes through a pull request and a human review. We still understand our codebase and can explain every change we merge. The agent writes a lot of our code now, helps us understand and evaluate concepts, and reviews PRs against guidelines the team agreed on. It runs inside the quality habits we already had and makes them faster.
Sceptics
We had AI sceptics on the team, and they did not embrace agents with enthusiasm. The sceptics eventually saw the backlog shrink, saw a new service land in our ownership without drama, and started using agents for the same reasons as everybody else.
Results convince better than demos.
The loop
Our trajectory still points upwards, and that is the part I find most interesting.
We take on more work. To cope, we pull the new tasks into our standard workflow. A standard workflow lets us simplify: fewer special cases, fewer one-off scripts, more of the codebase looking the same. Simpler code is easier for the agent and for us, which frees capacity for more work. Then the loop starts again.
Each round is small, and none of them would make a conference talk. After a year they add up to a factor of two. I do not know where it ends.
So the story is a modest one. It took no swarm of agents and no transformation programme. It took tools that work, proven habits, and a stubborn focus on quality and on what the customer gets.
What is your 2026 success story?