Eight in 10 enterprises report production failures from AI-generated code, CloudBees finds

TL;DR:

• 81% of more than 200 enterprise technology leaders surveyed by CloudBees reported an increase in production issues linked to AI-generated code, even as 92% expressed confidence that their code was production-ready before shipping.
• 61% of organisations’ code is now generated by AI or written with AI assistance, and 64% of engineering teams say AI is “widely or fully integrated” into their workflows, with 70% reporting that test-suite maintenance is now a larger burden than writing code.
• Cost pressures are following: 54% report significant CI/CD infrastructure spending rises in the past 12 months, while only 31% of AI-related spend can be linked to specific business results and just 18% have automated AI spending controls.

The headline gap — 81% reporting more production failures alongside 92% claiming pre-deployment confidence — is the verification gap that AI-augmented development is struggling to close. Failures here are not isolated CI/CD breakages but the full spectrum of post-deployment issues: functional bugs, performance regressions, availability problems, security vulnerabilities and compliance failures that all passed automated review and human gates before reaching production.

A counterweight to the productivity narrative driving UK SME adoption

The Register’s reporting on the CloudBees survey lands the same week the Ada Lovelace Institute called on UK policymakers to challenge AI productivity claims and the UK’s Public Technology coverage warned that “single studies can become powerfully relied on” in driving billions of pounds of decisions. CloudBees and Ada Lovelace are looking at very different sectors — enterprise software and UK public-sector productivity respectively — but the underlying point is the same: the headline AI-productivity numbers are not standing up to detailed scrutiny.

The cost-and-ownership picture in the CloudBees survey is also sharp. 36% of organisations either don’t track AI spending against return or aren’t tracking it at all. Only 12% have dedicated AI governance functions, and where production failures occur, accountability defaults to the CTO or VP of engineering 46% of the time, the engineering lead 32% of the time, and the developer who shipped the PR 7% of the time — a distribution that suggests organisations have not yet developed the AI-specific accountability structures their AI-specific failure modes require. For UK SMEs and FTSE companies that have invested in AI coding tools through 2024 and 2025, the numbers are a calibration point against the productivity promises baked into supplier pitches and internal business cases.

Looking forward

The verification gap is unlikely to close without explicit investment in evaluation infrastructure — and the CloudBees data suggests most organisations are not yet making that investment. Expect tooling vendors to reframe their pitches around governance and verification rather than raw productivity, the FCA and PRA to ask UK financial-services firms how they evidence the safety of AI-assisted code, and the UK government’s incoming AI Bill to wade into questions of accountability for AI-generated outputs that reach production. For UK SMEs, the practical question is whether your AI-coding cost is tracked against the rework cost of AI-induced production incidents — and whether the answer survives the Ada Lovelace test of methodological rigour.