Devs are coding faster. Coding reviews are eating the gains

What happened

New research from Bain & Company finds that while AI coding assistants are measurably speeding up software development, much of that time saved is being absorbed by the additional work needed to review, test and validate AI-generated code before it reaches production, according to CIO Dive.

Bain's analysis points to a widening gap between how quickly code can now be written and how quickly it can be safely shipped. Engineering teams are generating code faster with AI assistance, but reviewers — human or automated — are spending more time scrutinising that output, checking for errors, security issues and architectural fit, which erodes the net productivity gain organisations expected from these tools.

Why it matters

The finding complicates a widely held assumption in enterprise technology: that AI coding assistants translate directly into faster software delivery and lower development costs. Bain's research suggests the real constraint in modern software delivery is shifting from code generation to code verification — meaning organisations that invest heavily in AI writing tools without equally investing in review capacity, testing automation and quality gates may see little net improvement in delivery speed.

For technology and transformation leaders, this reframes where to focus AI investment. The bottleneck is no longer just "can we write code faster" but "can we trust and ship it faster" — a distinction that should shape how CIOs sequence AI adoption across the software development lifecycle, not just at the authoring stage.

The Renascence take

This is a familiar pattern dressed up in new technology: automating the easy part of a workflow while leaving the harder, judgment-heavy part untouched simply moves the bottleneck — it doesn't remove it.

Most organisations measure AI success at the point of generation — lines of code written, tickets closed, drafts produced — because that's the visible, easy-to-demo moment. But experience and quality are determined downstream, in review, validation and handoff, which is exactly where this research shows the friction has relocated. The lesson isn't that AI coding tools don't work; it's that speed gains anywhere in a workflow are only real if the surrounding process is redesigned to absorb them. A customer-obsessed operator should treat AI-assisted code the way they'd treat any new high-volume input channel: invest in the review and quality-assurance layer first, instrument where time is actually being spent end-to-end, and resist declaring victory on productivity metrics that stop at the point of output rather than the point of delivery.