AI coding got faster. Why didn’t engineering?

AI is helping individual engineers move faster. The catch is everything around the code is getting in the way, and the bigger the company, the worse that seems to get. DX’s new State of AI Impact in Engineering report says AI investment has jumped 28 times for most companies, especially those with more than 99 engineers, while velocity is flat or even down.

Justin Reock, DX’s deputy CTO, sees the core problem in the mismatch between cost and output. Spend keeps climbing, but the expected leap in shipping speed never shows up. The report also says the innovation ratio, meaning the share of engineering effort going to new features instead of maintenance, toil, and overhead, has stayed flat. So the money is going into AI tools, but not into noticeably more business value.

The developer experience data is even stranger. DX says AI is making code easier to read and easier to edit, yet change confidence has fallen below zero. Engineers feel more able to work in the codebase, but less willing to trust what gets pushed to production. That matters because DX says every point of improvement in its DXI benchmark returns ten hours a year to each engineer, and this is the first time it has seen that score drop across the industry by two points.

The size of the organization seems to be the big divider. Martin Davidson of a2bic.ai argues that small teams avoid the communication taxes that crush larger ones, and can get a lot out of AI agents because there’s less coordination overhead to begin with. His point is blunt: code writing is now cheap, but many companies are still using structures built for a world where it was expensive.

That pattern shows up in pull requests too. DX says the median PR grew from 42 lines of code in July 2025 to 72 lines a year later, while incremental delivery took the biggest hit in the latest quarter. LinearB’s separate AI engineering productivity gap report makes the same basic point from another angle: its top 10% of organizations average under 100 lines per pull request, while the bottom group sits above 228. And then there’s the measurement gap. LeadDev says only 31% of teams are even tracking AI’s impact, which makes all the victory laps feel a bit premature.