Your engineering org is shipping more code than ever. That’s the problem.

“It’s crazy to think that Claude Code was released just a little over a year ago,” says Martin Harrysson, Senior Partner at McKinsey & Company. “The improvement of these tools is happening so fast that they’ve moved from AI assistants and autocompletes to full-on agents who can take on real pieces of work end-to-end in a matter of months.” 

With the introduction and widespread adoption of AI coding tools, software development as we knew it will never be the same. Yet, what’s actually happening inside engineering orgs is not what most people would have expected. Drawing on telemetry from 4,000 teams and 20,000 developers over two years, the 2026 AI Engineering Report uncovered the far-reaching effects of AI adoption in software engineering. 

The report, published by Faros in April, found that AI is now the primary author of code, and throughput numbers are up. By every velocity metric engineering leaders have historically reported to their executives, the AI investment is paying off. 

But at the same time, the quality numbers tell a different story. Bugs per developer, PR incidents, time in review, and code churn are also up—and the gap between output and absorption is widening as AI adoption deepens. 

Software engineering maturity is not a shield

Everyone assumed that AI would amplify an organization’s strengths, and that large enterprises with elite DORA metrics and mature engineering practices would benefit the most. But we’re actually seeing the exact opposite. 

“High-caliber engineering orgs often have a harder time adapting than you’d expect,” Martin explains. “These teams usually have a strong engineering identity—and they see themselves as craftspeople—so they’re often more resistant to change compared to those in smaller, newer companies. Just think about it: A new ‘AI-native’ engineering org can make changes to tools and processes at lightning speed, whereas the change management required to get an org with 30,000 engineers to overhaul how they build software is an entirely different challenge.”

At those large companies, there may be individual developers flying high and excelling with AI, but the organization as a whole is not. “The workflow slowdown observed in the data is reminiscent of the 10x individual vs. 10x org problem,” says Martin. “AI works brilliantly for scoped individual tasks, but real companies run on systems, handoffs, and cross-team coordination. That’s where acceleration breaks down.”

The system wasn’t built for this

Engineering organizations are pushing AI-generated volume through a pipeline designed for human-authored volume. Code review processes, QA staffing ratios, team structures, and deployment gates were all calibrated for a world where coding were the bottleneck. They aren’t anymore. The bottleneck has shifted to everything around code generation: review, validation, governance, and deployment.

“The tools have gotten very good, but our ability to apply them—knowing what to do with the output, how to govern it, and how to restructure teams around it—has not kept pace,” Martin explains. “The gap between investment and return is real, and it won’t close by buying more seats or shipping more tokens.”

It’s no wonder organizations are reeling from AI acceleration whiplash.

What actually closes the gap?

In a recent webinar, Martin Harrysson and Vitaly Gordon, CEO at Faros, unpacked the full AI Engineering Report 2026 - Acceleration Whiplash findings and what it means for engineering leaders trying to connect AI spend to actual business outcomes. 

Watch the full webinar →