AI coding tools are spreading through engineering teams faster than trust in them is growing. That gap is becoming a real business problem for startups selling the next wave of coding agents.
The newest warning for AI-first engineering is not coming from people who refuse to use the tools. It is coming from developers who have used them, reviewed their output, fixed their mistakes and watched managers turn a useful assistant into a workplace mandate.
As Stack Overflow's 2025 Developer Survey makes clear, the market is moving in two directions at once. The survey found that 84% of developers either use or plan to use AI tools, up from 76% in 2024. At the same time, 46% said they distrust the accuracy of AI output, and positive sentiment toward AI tools fell to 60% from above 70% in both 2023 and 2024.
The executive story is much cleaner. Google recently said AI now generates about three quarters of its new code, Microsoft CEO Satya Nadella said last year that 20% to 30% of code in the company's repositories was written by software, and Microsoft CTO Kevin Scott has predicted that 95% of code could be AI-generated by 2030. Anthropic leaders have also talked about very high internal use of Claude for code generation. The message to the market is simple: software is about to get cheaper, faster and less dependent on large engineering teams.
That may be true in some workflows. It is not the whole story.
The strange part is that developers are not walking away from AI. They are using it while trusting it less. For founders, that is exactly the kind of signal that should slow down a go-to-market strategy built only around speed. Adoption is not the same as confidence, and usage is not the same as return on investment.
The labor issue sits inside the workflow. AI can produce a large amount of code quickly, but code still has to be understood, tested, reviewed, secured and maintained. If one developer generates 1,000 lines that another developer has to untangle, the company may have gained output while losing throughput. The dashboard looks busy. The team feels slower.
There is research pointing in that direction. METR's 2025 randomized study of experienced open-source developers found that, on familiar mature repositories, developers using early-2025 AI tools took 19% longer to finish tasks. The study was narrow, and it does not prove AI slows every team down. But it does challenge the easy assumption that newer coding tools automatically create more productive engineering organizations.
Hidden tech debt is a buyer problem
The deeper risk is not that AI writes bad code once. Bad code has always existed. The risk is that companies create a process where nobody feels fully responsible for the code because everybody assumes the tool, the prompt, the model or the reviewer will catch the problem later.
That is how hidden technical debt forms. Not in one dramatic failure, but in hundreds of small decisions where generated code is accepted because it works for the narrow case in front of the team. The cost arrives later, when the system needs to scale, security teams ask for proof, or a new engineer has to maintain code nobody really remembers designing.
This is where AI coding startups face a tougher buyer conversation. Selling agents as efficiency tools is easy when the buyer is chasing headcount savings. It is harder when the engineering organization asks for evidence that the tool reduces total cycle time, defect rates and maintenance burden. Buyers will want evals based on their own codebases, not generic benchmarks. They will want governance that says when AI can write code, when humans must design first, what gets logged, and who is liable when generated code creates a security or compliance problem.
The winners in this market may not be the companies promising the most autonomous agent. They may be the ones that make review easier, preserve context, expose uncertainty and fit into the way serious engineering teams already protect production systems. That sounds less glamorous than AI writing everything. It is also closer to how software actually gets shipped.
For entrepreneurs, the takeaway is practical. AI-first engineering can be an advantage when it removes low-value work and helps good engineers move faster. It becomes a liability when management treats code volume as productivity and turns review into unpaid cleanup. The next phase of the market will be less about whether agents can generate code, and more about whether startups can prove that generated code is worth owning six months later.
Also read: Banks are trying to narrow the CLARITY Act stablecoin deal • LinkedIn is cutting staff while its business is still growing • Tower Semiconductor turns AI demand into capacity commitments
