Why SaaS Fails AI: The Proto-Market Playbook is Premature by Derek Xiao of Menlo Ventures

The crowded landscape of generative AI applications—from code generation to legal support—looks like a chaotic mess of overlapping products with no clear moats or margins. But according to venture capital analysis, the problem isn't poor execution; it’s that the traditional software playbook is fundamentally broken.

AI categories are not mature markets at all. They are "proto-markets," evolutionary precursors where demand exists but the rules of engagement, technology stability, and economic structure have yet to settle. This distinction, argues Derek Xiao, a principal at Menlo Ventures, means the classic SaaS strategy—moats first, then margins, then market leadership—is "actively wrong."

The core issue is technological instability. In SaaS, the definition of "done" was stable. In AI, the foundation constantly shifts. When Claude Sonnet 4.5 was released, it forced companies like Cognition to entirely rebuild their agents because the underlying model behaviors broke previous assumptions about tool use and context awareness. Products like Manus have reportedly reconstructed their agent framework four times.

This instability collapses the boundaries that defined traditional software. Horizontal versus vertical, sales versus IT—these distinctions blur when a single generalized AI capability can span every department. Salesforce’s Agentforce and ServiceNow’s NowAssist are now competing on customer support, recruiting, and finance, far outside their original domains.

The New Moat: Learning Velocity

In proto-markets, the heavy, upfront engineering efforts that once built durable SaaS moats—proprietary foundation models or sophisticated search architectures—are failing. What wins instead is "product plasticity and learning velocity."

The goal is to build products that are "hackable by design," starting with the simplest utility and watching how users break or adapt it. This approach was central to the development of Claude Code, which began as a simple chat loop wired to the filesystem.

When Anthropic observed power users creating plans before coding, they productized that pattern into Plan Mode. When users built libraries of executables, that evolved into Agent Skills. Every feature addition formalized a behavior users were already hacking together.

This evolutionary approach is critical because it creates a path-dependent advantage. The learning—captured in decision traces and feedback loops—is encoded directly into the product’s structure. By the time these proto-markets harden and selection pressures sharpen, the winners will be the companies whose products have accumulated the deepest, most structural understanding of what users actually need.

AI’s first durable moat is not proprietary data or unique models. It is the accumulated understanding derived from treating every user interaction as raw material, compounding learning into an advantage before the market even knows what it wants to be.