The software industry is having its worst month since October 2008. The iShares Expanded Tech-Software Sector ETF (IGV) has dropped 22% from its highs, officially entering bear market territory. ServiceNow plunged 10% despite beating earnings. SAP fell 16%. Salesforce, Adobe, and Intuit all posted double-digit declines.
The narrative driving this selloff: AI will replace software subscriptions. Why pay for SaaS when you can just tell an AI to build it?
But here's what the market is getting wrong: it's pricing the demo, not the deployment.
The Debt Market Is Even More Spooked
The equity selloff is dramatic, but the leveraged loan market tells an even starker story. According to Nomura research, software debt in collateralized loan obligations has posted the lowest returns of any sector in 2026.
The damage is specific and measurable:
- LegalShield, which provides document review and legal services, saw its $1.15 billion first-lien loan sell off to 94 cents on the dollar
- Consilio, another legal document review company, watched its $2 billion loan slide to 84 cents from par
- Verint, a call center software provider, couldn't find buyers for its $2.7 billion buyout loan; Santander is stuck holding a portion they couldn't syndicate
- Getty Images priced bonds at 10.5% yield, far wider than companies with comparable credit ratings
The thesis seems logical on the surface: if AI can review legal documents, why do you need LegalShield? If AI bots can handle customer service, why do you need Verint's call center software?
But this logic contains a fundamental flaw.
The Gap Between Demo and Deployment
I use AI coding tools every day. Claude Code is genuinely useful; it accelerates my workflow and handles boilerplate tasks efficiently. I'm not an AI skeptic.
But there's an enormous gap between "AI can write a function" and "AI can build and maintain enterprise software." The market is pricing as if that gap doesn't exist.
The data tells a different story.
CodeRabbit's State of AI vs Human Code Generation Report, which analyzed 470 real-world pull requests, found that AI-generated code produces significantly more defects across every major quality category:
- 1.75x more logic and correctness errors
- 1.64x more maintainability problems
- 1.57x more security findings
- 2.74x more XSS vulnerabilities
This aligns with what I wrote in The Security Debt Crisis from AI-Generated Code: AI models are getting better at writing functional code, but they're not getting better at writing secure code. Security performance has remained flat across models of varying sizes and sophistication.
Microsoft patched 1,139 CVEs in 2025, the second-largest year for CVEs by volume. Microsoft says 30% of code in certain repos was written by AI. Analysts predict this number will rise in 2026 as AI-generated bugs become more prevalent.
"Vibe Coding" Isn't Enterprise-Ready
The term "vibe coding," coined by Andrej Karpathy, describes using natural language to generate software without reviewing the code. It's an impressive demo. It's also not enterprise-ready.
According to GitLab's 2025 DevSecOps survey:
- 45% of enterprises use AI coding assistants for isolated tasks
- 22% have integrated AI into formal development workflows
- Only 7% rely on vibe coding for mission-critical systems
That last number is critical. The market is pricing mass displacement of enterprise software, but 93% of enterprises won't even trust AI-generated code for their most important systems.
Gartner estimates 40% of new enterprise production software will use vibe-coding techniques by 2028. But "new" doesn't mean "replacement of existing." And 40% adoption in two years, for a fraction of software, is very different from the "AI replaces all software" narrative driving this selloff.
Enterprise Integration Is the Moat
Here's what the "AI replaces software" thesis fundamentally misses: enterprise software isn't just code. It's decades of integrations, compliance requirements, institutional workflows, and accumulated institutional knowledge.
New research on enterprise AI adoption reveals the reality:
- 42% of enterprises need access to 8 or more data sources just to deploy AI agents
- 86% require upgrades to their existing tech stack for AI agent deployment
- Only 5% of custom AI projects reach production (MIT study)
You can't vibe-code your way to HIPAA compliance. You can't prompt-engineer your way to SOC 2 certification. As I discussed in AI Governance in Enterprise Data Management, governance frameworks that took decades to build don't get replaced by a better language model.
The integration complexity alone creates a moat. Legacy infrastructure wasn't designed for real-time model inference. It lacks the compute capacity, modularity, and scalability that AI demands. When 86% of CIOs report unprepared network capacity for AI workloads, technical constraints become fundamental barriers to the mass displacement the market is pricing.
The Innovator's Dilemma, Misapplied
CNBC recently cited Clayton Christensen's Innovator's Dilemma to explain the software selloff. The framing is seductive: AI is the disruptive technology that will eat software from below.
But the Innovator's Dilemma theory has a specific pattern. Disruptive technologies start by being "good enough" for the low end of the market, then progressively improve until they threaten incumbents at the high end.
AI code generation doesn't fit this pattern. It's not even "good enough" for the low end of production software. It produces more bugs, more security vulnerabilities, and requires senior developers to review and refine everything. The Cortex Engineering Benchmark Report found that PRs per author increased 20% year-over-year, but incidents per pull request increased 23.5% and change failure rates rose 30%.
That's not disruption from below. That's a productivity tool that creates more work downstream.
Contrarian Voices Are Emerging
Not everyone on Wall Street is capitulating. D.A. Davidson's Gil Luria argues that "the death of software has been greatly exaggerated," pointing to three years of AI coexistence without mass software extinction.
Mizuho analysts are calling the selloffs "silly," noting that AI won't easily replace complex enterprise workflows that took decades to build.
LPL Financial's Adam Turnquist summarized it well: "The market's pricing a worst-case scenario that software is dead because AI is disrupting the space."
Some analysts identify companies like Snowflake and MongoDB as less vulnerable because they're "data connectors," infrastructure that makes data usable across systems rather than end-user applications. This distinction matters: the market is treating all software as equally threatened when the actual exposure varies dramatically by category.
The Yellow Pages Analogy Is Wrong
Citigroup's Michael Anderson asked the question on every credit analyst's mind: "Who is the Yellow Pages of today?" The implication is that some software company will be made obsolete by AI the way search engines made phone books obsolete.
But the Yellow Pages analogy doesn't hold. Phone books were simple directories with no integration complexity, no compliance requirements, no institutional workflows built on top of them. Enterprise software is fundamentally different.
Consider what it actually takes to replace a company like Verint. You're not just replacing call center software. You're replacing integrations with CRM systems, telephony infrastructure, compliance recording requirements, quality assurance workflows, agent training programs, and reporting systems that feed into executive dashboards. All of that took decades to build and refine.
An AI bot that can answer customer questions is impressive. It's not a replacement for enterprise call center operations.
What I See From Both Sides
I sit at an unusual intersection on this topic. I use AI coding tools daily and see their genuine productivity benefits. I also have an investment banking background and understand how debt markets price narrative risk.
From the developer side: AI assistants are tools, not replacements. They accelerate certain tasks and create new problems in others. The net productivity gain is real but modest, and it comes with a security debt burden that most organizations haven't yet reckoned with.
From the credit market side: leveraged loan traders are notoriously reactive to narrative. When they hear "AI replaces software," they sell first and analyze later. The LegalShield, Consilio, and Verint selloffs show this pattern: headline risk driving positioning, not fundamental analysis of how AI actually works in production environments.
The pattern reminds me of the dot-com crash. Microsoft lost 58% of its value and took 17 years to fully recover its peak valuation. But the companies that survived the crash became trillion-dollar giants. The market oversold the internet's impact on traditional business; it may be doing the same with AI's impact on software.
The Bottom Line
Software stocks are in bear market territory. Leveraged loans for software companies are selling off. The market has decided that AI will make software subscriptions obsolete.
But the data doesn't support this thesis:
- AI-generated code produces 1.7x more bugs than human code
- Only 7% of enterprises trust vibe coding for mission-critical systems
- 86% of enterprises need major tech stack upgrades just to deploy AI agents
- 62% of AI code optimizations contain bugs
- Complex enterprise workflows took decades to build and can't be replaced with a prompt
Wall Street is pricing a future where AI produces flawless production code instantly. That future doesn't exist today, and the path to get there is far longer than the market implies.
The demo is impressive. The deployment is a different story.
