"Software patents are dead." For nearly a decade after Alice v. CLS Bank, this was conventional wisdom in Silicon Valley. AI-powered software companies—especially those building on foundation models—were told not to bother. But Director John Squire's USPTO is rewriting that narrative, and vertical AI companies should pay attention.

Why Vertical AI Companies Dismissed Patents

The post-Alice landscape was brutal for software patents. Examiners routinely characterized any software innovation as an "abstract idea" implemented on "generic computer components." Vertical AI companies building on APIs from OpenAI or Anthropic faced an even steeper hill: if you're "just" orchestrating someone else's model, what's patentable?

This thinking was always flawed—significant innovation happens in how AI is applied to domain-specific problems—but the USPTO's hostile examination environment made pursuing patents seem futile. That environment has fundamentally changed.

The New Guidance Changes the Calculus

The August 2025 USPTO memo contains several provisions favorable to vertical AI companies:

• "Involves" vs. "Recites": Claims that merely "involve" an abstract idea but don't explicitly recite one are patent-eligible without further analysis. Your AI orchestration claims may not even trigger Section 101 scrutiny.
• Technical improvements count: Even when using "generic" components (like API calls to foundation models), technical improvements to the overall system can establish patent eligibility.
• Close calls favor applicants: Examiners should only reject when ineligibility is "more likely than not." This shifts the burden significantly.

What Vertical AI Companies Can Patent

Even if you're building "just a wrapper," your implementation likely contains patentable innovations:

• Multi-model orchestration: Novel architectures for routing, chaining, and combining outputs from multiple AI models
• Domain-specific validation: Technical systems for ensuring AI outputs meet industry-specific requirements (regulatory compliance, safety constraints, accuracy standards)
• RAG implementations: Novel approaches to retrieval-augmented generation, especially domain-specific chunking, embedding, and retrieval strategies
• Error correction pipelines: Systems that detect, correct, or gracefully handle AI failures in mission-critical applications
• Human-AI workflow integration: Technical implementations of how AI assists human decision-making in specific domains

The "No Moat" Counterargument

VCs and critics often claim AI wrapper companies have "no moat." Patents directly address this concern. A competitor can call the same OpenAI APIs, but they cannot replicate your patented orchestration architecture, your novel validation pipeline, or your domain-specific error correction system without licensing or infringement risk.

In fundraising and M&A contexts, patents transform "we built a better product" into "we own IP that competitors cannot legally replicate." That's the moat critics claim doesn't exist.

Building Your Portfolio Now

The favorable USPTO environment may not last. Vertical AI companies should:

• Identify your genuine innovations: What technical problems did you solve that the foundation model alone couldn't?
• Focus on system claims: Patent the architecture and workflow, not just individual prompt techniques
• Document the technical improvements: Speed, accuracy, reliability, cost efficiency—quantify how your implementation improves on baseline approaches
• File provisionals immediately: Secure priority dates under the current guidance while refining your claims

Conclusion

The "software patents are dead" narrative was always oversimplified, but it was understandable given the hostile examination environment. That environment has changed. Vertical AI companies now have a clear path to building defensible IP portfolios—and those who act quickly will establish positions that slower competitors cannot easily challenge.