AI Pulse · Business Models & Winners

Vercel just shared a number worth sitting with. The company built what it calls an internal “AI SDK software factory,” where each step of building software, writing code, testing it, fixing bugs, is handled by an agent, and a human just merges the final result. Four weeks in, Vercel says this pipeline now authors up to 35% of merged pull requests. It also closed 70% of July’s issues and cut open bugs by 25%. This is self-reported by Vercel, with no independent check on how they measured it, so treat the exact figures as a claim, not a verified fact.

Still, it’s a useful data point because it’s concrete. Not “AI will change engineering someday,” but a specific team saying: over one month, agents handled about a third of the changes that actually shipped, while a human stayed in the loop to approve them. That’s a real shape for what agent-run pipelines can absorb right now, even if the number softens under closer scrutiny.

For a business owner, the lesson isn’t “go automate your codebase.” It’s the pattern underneath it. Break a workflow into discrete steps, let an agent own each step, and keep a human as the final checkpoint before anything goes live. That structure works for customer support tickets, invoice processing, content review, or onboarding, not just software. The hard part is defining the steps clearly enough that an agent can actually do them and a human can verify them quickly.

This is the kind of thing we build into an AI command centre, where multiple agents work in parallel on a workflow and a person just signs off before anything moves forward. If you’re wondering whether a piece of your own operation could run this way, that’s exactly the question worth testing with a small pilot before you plan around numbers like 35%.