What 20M Pull Requests Reveal About AI's Impact

Jellyfish's Nicholas Arcolano breaks down 20M pull requests from 200K engineers: coding speed has doubled, so why hasn't business output caught up?

Jul 27, 2026 36:39

Nicholas Arcolano

Expert Insights

Nicholas Arcolano brings two years of Copilot adoption data across 200,000 engineers and 700+ companies, and the pattern he's found complicates the easy AI-productivity story. Coding speed has roughly doubled, he says, but that gain barely touches overall business output. The real constraint sits downstream, in sales enablement and roadmap adaptation, with code review capacity as an added cap. Arcolano argues that most companies have only accelerated their existing workflows so far, and the harder cultural shift is still ahead.

Hear him outline:

Why coding speed can roughly double while business output barely moves, and where the real bottleneck sits
What Jellyfish's data on 20 million pull requests across 200,000 engineers shows about real-world AI coding adoption
The three types of companies pulling ahead with AI, and what sets them apart
Why context and infrastructure explain most of the gap between teams
What leaders should be asking about their path to autonomous agents over the next six months

I don't trust the opinion of any leader who isn't working with these tools themselves. When you talk to someone, you can tell immediately whether they're actually living this or just reading about it, and you have to live it.

— Nicholas Arcolano, Head of AI & Research at Jellyfish

Monterail Team Analysis

Here's how to close the gap between faster code and real business impact, based on what Arcolano's data shows:

Look past coding speed. If pull requests are moving faster but revenue and margins aren't, the real constraint is downstream, in sales enablement and marketing, and in how fast your roadmap can absorb the extra capacity.
A human developer can find their way through a messy codebase on instinct, something an agent can't do; someone still has to fill in those gaps before agents can work independently.
Expect to build your own agent stack for now. There's no shared standard yet for how teams structure agent tooling, so plan for iteration rather than a plug-and-play setup.
Budget for token spend as a real line item, and check it delivers value. Rising inference costs are a normal early signal of AI adoption, but they need to be tied back to actual output, not treated as a given.
Get your own leadership using the tools. Arcolano holds this as a hard rule for himself, and it's a fair bar for any leader: your read on AI adoption is only as credible as your own hands-on use of it.
Revisit code review capacity before it becomes the bottleneck. As developers ship more, review and accountability need to scale right alongside quality checks, not lag behind.
Watch for product, design, and engineering roles blurring at the edges. The teams pulling ahead already let people move fluidly across the build process, keeping those functions connected.