The state of play: AI code in product today
I've titled this "as of today" deliberately. The pace of change is fast enough that the position I held six months ago has shifted, and I expect it to shift again. I'm writing this to share what comes up most often in conversations with clients and our own teams. The question is no longer whether to use AI-generated code, but how much. How far should we rely on it, do you use it, and can the tool a colleague built over a weekend go live? Alongside our own experience at Deepend, I've gone looking for what the market is saying.
My conclusion, for now, is that AI-generated code works well below a certain level of complexity and degrades beyond it, and the difference between value and risk comes down to verification and maintainability.
Where it is delivering
Gains are now reaching delivery, not just individual output. Faros AI found task throughput per developer increased 69.0% and epics completed per developer rose 53.7% as teams deepened AI adoption. Quality per change is stabilising, with Faros reporting that incidents per pull request under high adoption rose 14.5%, against 242.7% in April, and deployment frequency is climbing again. Analytics improves where context is governed, and AI-generated query accuracy on dbt Labs' benchmark rose from 32.7% in 2023 to 64.5% in 2026, and to 98.2% with a modelled semantic layer.
Our experience sits at the measured end of this. Our teams use AI to generate alternative approaches, as additional senior-level support on difficult problems, and for limited, reviewed snippets. It has proved the least disruptive model, adding thinking capacity without handing over authorship of systems clients will run for years.
Where it struggles
Game developer Jonathan Blow describes current AI as a search algorithm that kind of gets what you mean, fast on well-defined problems but weak at the judgement calls complex systems demand. The evidence clusters where complexity and consequence are highest.
Security is flat. Veracode's 2026 testing found an average security pass rate of approximately 56%, with cross-site scripting passing only about 15% of the time. Georgia Tech tracked 35 CVEs in March 2026 alone directly attributable to AI coding tools, estimating the true count is five to ten times higher.
Accessibility is going backwards. The 2026 WebAIM Million found 95.9% of top homepages failing WCAG, up from 94.8%, with average errors per page up 10.1%, a reversal WebAIM attributes in part to AI-assisted coding.
Verification is eroding. Unreviewed merges rose 76.3%, compared with 31.3% six months earlier, and the ACM has found AI coding platforms modifying, disabling or deleting failing tests rather than fixing the underlying code.
Maintainability and content, my biggest concern
The risk I see most isn't at launch but in the years after. AI-assisted code tends to be duplicated rather than consolidated, and GitClear found refactoring changes have fallen from 24% to below 10% since 2021, while copy/pasted code has risen from below 10% to almost 15%. That makes systems harder to change safely.
Content is the other half. Editors and marketers maintain these platforms daily, and features generated without a considered content model are hard for them to update. That matters because the six most common accessibility failures in the WebAIM Million are all editorial decisions, not development decisions.
A client recently showed us a vibe-coded tool built to replicate a complex system holding customer data, unvetted and unhardened. For me, that's a significant red flag. Anything handling customer data needs security testing, access controls, a maintenance owner and an update plan, none of which appear by default when code is generated quickly. It also showed why these tools tempt, as a working prototype now appears in days and looks finished.
How we are approaching it
We're building verification before scaling generation, starting with a multi-agent test harness in which fully agent-generated code is checked by independent agents and automated tests before human review. Broader use should follow proof that output can be verified, not precede it.
We're also now being asked by clients to harden AI-generated code they have already built, reviewing, securing and stabilising it before it goes into production. It's a new area of practice for us, and a revealing one, because it shows us the realities of AI-generated code in real time, across tools and teams we didn't choose.
Where credible voices agree
Advocates and critics disagree on pace more than principle. Anthropic's Boris Cherny reports an 8x increase in the amount of code written at Anthropic since the start of the year, while Simon Willison argues the key skill is being able to confidently instruct agents and then confidently verify that changes have been applied correctly. The UK's NCSC judged that vibe coding poses an "intolerable risk" for many organisations as things stand, but offers "glimpses of a new paradigm".
Accountability stays local
The questions I find myself asking
When clients ask how much to rely on AI-generated code, I reframe it as a different question, how is it being controlled? Is work matched to complexity, with integrations, data migrations and regulated journeys held to a higher bar? Is every AI-generated change reviewed by someone able to verify it? Are security and accessibility checks mandatory? Does someone own maintenance, and can content teams update what's built? Is analytics grounded in governed definitions? Do agents have scoped permissions, with production credentials out of reach?
A note on the evidence
Much of the quantitative data comes from vendors, and some voices quoted have commercial interests, so individual figures are indicative rather than definitive. What gives me more confidence is that our own delivery is bearing many of these findings out, particularly the complexity threshold, where AI accelerates well-defined, contained work while the effort to verify and maintain what it produces rises sharply as integrations, data and content requirements grow. Across independent research, commercial data and our own experience, the direction is consistent, showing that AI amplifies whatever discipline is already in place.