Six months of AI, measured in git instead of vibes

It’s almost 1am and I’m reading four years of my own git history over SSH from my phone, because a conversation about revenue turned into a question I couldn’t answer: what did AI actually do to my team?

Everyone I talk to has a story about AI making them faster. The stories are all vibes. Mine was too, until tonight, when I pulled the actual numbers.

Context: I run a 20-plus-year-old sports SaaS with a tiny team. A support dev, a designer, a reviewer, and me. The AI coding agent arrived about six months ago. Here is what the record says it did.

the numbers

No survey, no self-report. One loop over git:

# commits per author per year, merges excluded
for y in 2023 2024 2025 2026; do
  echo "== $y =="
  git log --no-merges --since="$y-01-01" --until="$y-12-31" \
    --format='%an' | sort | uniq -c | sort -rn
done

What it printed, by author:

My own 1,105 in six months annualizes to about 2,200, roughly 2.3 times my decade-long rate. The support dev sat around 300 a year for three straight years and is now pacing about 670, the same 2.3x. The designer had never committed a line of code in his life and has shipped 106 commits in six months. He is a developer now.

The reviewer is the exception. His commits are flat to down, but his review comments tripled year over year and concentrated: about five per issue he touches now, versus two a couple years back. He barely uses the agent himself. More on him shortly.

Tickets resolved as a team: we peaked near 1,470 a year in 2019, cratered to about 510 in 2022 during our biggest product bet, and are now pacing about 1,270. Back near peak, at roughly half the historic contractor cost.

the part the numbers underplay

The multipliers are the headline. The structural change underneath them is what actually reshaped the company: work stopped queuing on people.

For twenty years this team had two serialization points. Anything heavy came back through my review. Anything visual waited on the designer. Those two queues were the actual shape of the company, and every process we had was built around them.

Six months in, both queues are just gone.

The support dev’s backlog used to be “things waiting for Ben to review.” When he got comfortable with the agent, one of the first things he did was clear that backlog himself. When his tasks need design thinking, he doesn’t wait for the designer anymore either. He prototypes with the agent and the designer reviews a working thing instead of a request.

The designer’s change is bigger. Our old flow was hours of scope meetings, then I code alone, then we iterate. Now he has ideas I don’t fully agree with or even understand, and instead of spending meetings convincing me, he explains them to the agent and builds them. One of those, a feature I would have called a multi-month project, went from created to live in about eight weeks with exactly one commit from me out of thirty. Another one, a concept I was lukewarm on, he’s nearly finished building without needing me to be sold on it first.

And the styling work that used to eat my evenings, the endless pixel tweaks that were the worst use of my hours and the best use of his eye, I just hand to him now. Whole category of my time, returned.

the teammate who didn’t adopt

The reviewer is the interesting case. He uses the agent the least of anyone, and he became more valuable, not less.

When everyone else’s volume doubles, the scarce resource stops being production and becomes judgment. His deep review threads are the quality gate that lets the rest of us run this fast. If I had measured adoption by “who uses the AI most,” I would have scored our most important adaptation as a failure.

every gain in a different currency

My sense of it was just “we’re faster.” The data said something more specific. Every person’s gain landed in a different currency. Volume for me and the support dev. Capability for the designer. Depth for the reviewer. Each of those needs a different thing from me now, and I couldn’t see any of it until I queried it.

Two caveats, because the numbers tempt overclaiming. Commit counts measure activity. They say nothing about size or quality, and AI-era commits trend smaller and more frequent. Some of the activity under my name is agent sessions committing as me. The direction and the multiples survive both. The precision doesn’t.

Six months ago I would have told you AI made me faster. That undersold it. It reorganized my company, and I only found out by reading the logs.