AI is writing almost all startup code. That's creating a new problem.

AI Tools Now Generate Most Code at Many Startups

Anthropic PBC hosted its Code with Claude developer conference in London on May 19, 2026. At Alma, a Menlo Ventures-backed nutrition coaching application, cofounder and chief executive Rami Alhamad stated that nearly every line of code shipped by the company is produced by artificial intelligence systems. Alhamad described the proportion as close to 100 percent.

A survey conducted by Business Insider of more than two dozen startup founders and venture capital investors indicated that generative AI systems have become the main source of code at many early-stage companies. Anthropic’s Claude Code was identified by respondents as the most frequently used tool. Several founders reported that the volume of code produced has increased substantially, while others noted added time required for review and correction.

Dan Lorenc, cofounder and chief executive of Chainguard, an open-source cybersecurity firm, said 100 percent of the company’s code is now generated through Claude Code. He placed the figure at 60 percent one year earlier. Lorenc compared the change to the introduction of power tools in woodworking, noting both the increase in output speed and the requirement for additional controls. He stated that the models, tool interfaces, and supporting systems improved markedly in the preceding four to six months, allowing tasks previously requiring weeks or months to be completed in hours or days, provided engineers continue to supply prompts and direction.

At Wordsmith AI, a platform serving legal teams, chief technology officer and cofounder Volodymyr Giginiak reported that the company’s code is nearly 100 percent AI-generated. He said direct human authorship of code has become minimal and that the relevant distinction is now the degree of autonomy granted to the AI systems. Giginiak estimated that fully autonomous tasks currently represent about 10 percent of engineering work and projected that share could reach 80 to 90 percent within one year. He described engineering roles as shifting toward the design of environments and context in which AI systems operate.

Danny Freed, chief executive and founder of Blueprint, which develops an AI operating system for therapists, stated that nearly all of the company’s code is now written by AI, an increase from 40 percent in August of the prior year. Freed noted that the expense of testing new approaches has declined, yet he said human employees remain essential for decisions about which features to pursue.

Jason Alan Snyder, a futurist and cofounder of SuperTruth and Artists & Robots, observed that rapid code generation can produce systems that are difficult to maintain after initial release, particularly when teams lack capacity for ongoing support. Several founders in the survey described similar patterns of additional review time offsetting some of the gains in generation speed.

Venture capital firms have directed substantial funding toward companies developing AI coding products, including Lovable, Replit, and Cursor. SpaceX announced an agreement to acquire Cursor for $60 billion. Anthropic has submitted filings for a public offering expected later in the year.

Founders interviewed for the survey described the current workflow as one in which engineers retain responsibility for specifying requirements and verifying output. They reported that the proportion of code requiring subsequent modification varies by project and by the complexity of the task. Some companies have introduced internal review processes and testing protocols to address quality issues before deployment.

The shift has prompted changes in hiring and team structure at several firms. Executives said they continue to seek engineers who can define clear specifications and evaluate AI-generated results, rather than those who write every line manually. At the same time, the volume of code under management has grown, increasing the amount of documentation and testing required to keep systems stable.

Industry observers noted that the pattern is most visible at smaller companies where rapid iteration is prioritized. Larger organizations have adopted similar tools more gradually, often limiting autonomous generation to narrower modules. The survey responses indicated that adoption rates differ by technical domain, with web and application development showing higher percentages of AI-generated code than infrastructure or security components.

Alhamad, Giginiak, Lorenc, and Freed each stated that their teams still perform manual oversight on all production code. They described the process as iterative, with engineers issuing repeated prompts and adjustments until the output meets functional and reliability standards. None of the executives reported eliminating engineering positions as a direct result of the change; instead, they said the nature of the work has evolved toward higher-level design and verification tasks.

The Business Insider survey did not include quantitative data on defect rates or long-term maintenance costs across the sampled companies. Respondents offered qualitative assessments that the time saved in initial generation is partially offset by review activities, though the net effect on overall development velocity remains under evaluation at most firms.