MUMBAI: AI may be writing more code than ever, but it still cannot debug a bad idea. As artificial intelligence becomes an increasingly common co-pilot in software development, Boris Cherny, creator of Anthropic’s coding assistant Claude Code, has cautioned that fully automating software engineering remains far more complicated than many companies anticipated.
Speaking at a fireside chat hosted by Scale AI, Cherny argued that while AI-generated code is dramatically improving developer productivity, organisations are discovering that software development’s biggest challenge is no longer writing code, it is knowing what to build in the first place.
His remarks come as technology companies face growing scrutiny over whether soaring AI investments are delivering meaningful business returns.
“ROI is absolutely the right framing, you spend something on it and you get something back,” Cherny said, joining a wider industry conversation around the economics of AI adoption. The debate has intensified in recent months, with executives including Uber COO Andrew Macdonald questioning whether rising AI expenditure is translating into enough real-world outcomes for customers.
According to Cherny, measuring progress through the volume of AI-generated code is becoming increasingly irrelevant. Many developers already rely heavily on AI tools, making the percentage of machine-written code a poor indicator of success.
“Once you get it to this point… the bottleneck is going to be good ideas,” he said.
The observation reflects a broader shift taking place across the software industry. As AI takes over repetitive coding tasks, human value is moving further upstream toward product thinking, innovation and defining problems worth solving.
In other words, the challenge is no longer building faster, it is building smarter.
Cherny also highlighted the financial realities underpinning large-scale AI deployment. Every interaction with an AI model consumes computational resources, commonly measured in tokens, which carry tangible costs for providers.
“Every token we use is a token we do not give to a customer, so there’s an opportunity cost,” he said, underscoring the importance of balancing experimentation with efficiency.
Yet he warned against becoming overly cautious too early. Restricting AI usage in pursuit of immediate cost savings, he argued, could prevent organisations from uncovering more valuable long-term applications of the technology.
The discussion also touched on the next phase of AI-assisted software development.
Cherny said the industry is increasingly moving towards what he described as “loop engineering,” a model in which AI agents generate, refine and optimise prompts autonomously, reducing the need for constant human intervention.
Under this approach, developers interact with higher-level systems that coordinate tasks across multiple AI agents and sub-agents. While this promises greater automation and efficiency, it also introduces new complexity and potentially higher operating costs as several AI systems work simultaneously behind the scenes.
The comments reflect a growing recalibration across the technology sector.
After the initial excitement around generative AI’s ability to write code, businesses are now confronting tougher questions around cost, scalability, governance and measurable returns. Productivity gains remain significant, but the path to fully autonomous software development appears less straightforward than early predictions suggested.
For now, AI may be accelerating how software is built, but Cherny’s message is clear: code can be automated, creativity cannot. As the industry races towards an AI-powered future, the real competitive advantage may lie not in generating more lines of code, but in generating better ideas.