The Best Engineering Teams Are Moving Dramatically Faster With AI. Here's How to Measure Yours.

The best engineering teams are moving at a pace that would have sounded made up two years ago.

At OpenAI, three engineers driving Codex merged about 1,500 pull requests in five months, without writing code by hand. Lauren Tan, the engineer behind the open-source agent toolkit pstack, says she ships about 2,000 PRs a month. Our own team writes code about 50x faster than it did a year ago. In one 30-day stretch, our team merged 690 pull requests, a median of 16 minutes apart. (Our CEO wrote up exactly how we do it.)

That gap is the opportunity. The hard part is knowing where your own team sits on it.

The short answer? Look at the work that merged, not at usage. Usage is who has a seat, how many prompts they send and how many tokens they burn. It tells you AI is being used, not what it shipped. Compare AI-assisted and human-only pull requests team by team, check whether that work got reviewed, and use what you learn to decide where to lean in, what to fix and which tools to keep.

Three questions to ask about AI productivity for engineers

1. How much more are we getting done?

Start with work that merged, not work that started. Prompts, accepted suggestions and opened PRs are all activity. A merged pull request is about as close as engineering gets to a unit of shipped work.

Then compare PRs written with AI help against PRs written without it, on the same repos over the same window, and see how long each takes from first commit to merge.

Do it by team. An org-wide average hides the most useful part, which is the spread. One team may already be working like the examples above while the team next to it hasn't changed at all. If I'm a manager, my own team is the first place I'd look. Those are the people I know, and the ones I'd ask what's working.

One caveat: this is a comparison, not a controlled experiment. Engineers choose what to hand to AI, so AI-assisted PRs might be smaller or in easier parts of the codebase. Treat a gap as a signal to dig into, and compare within a team where you can. Inline completions and pasted code also leave no trace in a PR, so any AI share you can see is a minimum.

2. Where's our next speedup?

Every time you speed up one stage, the next one becomes your biggest opportunity. When engineers and agents open far more PRs, review, testing and CI are where the next gains are.

The teams furthest ahead say this out loud. Revolut gave its roughly 2,000 engineers unlimited tokens on a self-hosted model. By its own comparison, lead time (from picking up a task to merge) improved about 70%, more than coding time at about 60%. Its AI operations lead put it simply: the bottleneck for engineering is not necessarily coding.

So for every speed number, track what keeps it going:

If AI work merges faster partly because fewer people look at it, that's your cue to build better safeguards. You don't put brakes on a car to stop; you put them on so you can go faster. The teams moving fastest are the ones that can see when something goes wrong.

3. What do we do next with AI?

Measuring only matters if it changes a decision. The first two answers usually point to four:

• Lean in where it's working. A team where AI-assisted work merges faster and holds up has figured something out. Find out what (the tool, the workflow, the type of work) and spread it.
• Fix the new drag. If review is now the constraint, change how review works: smaller PRs, clear owners, or an independent AI first pass with a human sign-off.
• Look hard at teams that aren't moving. No speedup can mean the wrong tool, missing context, or work AI isn't good at yet. Each needs a different fix.
• Renew with evidence. Before the next renewal, check which tools are used in which repos and how their work holds up.

Before your next AI tool renewal, check:

• Which tools are actually used, and in which repos?
• For each tool, does AI-assisted work merge faster than other work?
• Is that work getting reviewed, and does it hold up after merge?
• What does each tool cost, across seats and usage?
• Which teams aren't seeing a speedup, and why?

Stop flying blind

AI tools and agents are now a real part of the engineering workforce. You wouldn't run a team without knowing what it delivers and where it gets stuck, and the AI doing a growing share of your work is no different.

The teams at the top didn't get there by accident. They measured, found what worked, fixed what didn't, and kept going. A year from now, your team could be the example someone else writes about. The first step is seeing where you are today.

That's why we built Proxon Productivity, which we're launching today. It helps engineering leaders answer a question that matters more every month: how much faster is AI making our engineers, and what return are we getting on what we spend? Connect GitHub or GitLab, with nothing to install on anyone's laptop, and within minutes you see AI-assisted and human work side by side, with six months of history, by team and tool, including how much of it got human review. With Proxon's Cost Intelligence, you see it next to what you spend, so you can tell what's working before the bills grow.

We've made this shift ourselves, and we'll help you make it too. A Proxon expert can work on site with your engineers: take a baseline, train the team on a modern, AI-first way of building software, then hand it off while Proxon keeps measuring. Email hi@proxon.ai to talk to one of our experts.

Want to see where your team stands? Book a demo and we'll walk you through it on your own repos.

FAQ