February 2026
On January 30, 2026, Anthropic released 11 open-source plugins for Claude Cowork spanning sales, finance, legal, data, and marketing. With these plugins, Claude can act as a sales expert, financial analyst, legal advisor, or data analyst. Work that once required a full software stack and process can now be handed to AI. Wall Street quickly declared “SaaS doom,” and software stocks sold off sharply.
Nicolas Bustamante—a vertical software founder (Doctrine, Europe’s largest legal information platform; Fintool, an AI-driven equity research platform competing with Bloomberg, FactSet, and S&P Global)—published a widely discussed piece on February 17: “Ten Years Building Vertical Software: My Take on the Selloff.” Below is an adapted translation of the core argument.
The Selloff in Numbers
In recent weeks, nearly $1 trillion in market cap evaporated from software and services. FactSet fell from a $20B peak below $8B. S&P Global lost about 30% in weeks. Thomson Reuters was nearly halved over the year. The S&P 500 Software & Services index (140 names) is down 20% year to date. Last week Anthropic released industry-specific plugins for Claude Cowork—an AI agent for knowledge workers that autonomously handles research, analysis, and document workflows. Wall Street called it panic.
Bustamante’s thesis: LLMs are systematically eroding the moats that made vertical software defensible—but not all of them. The result is a re-pricing of what makes vertical software valuable and what multiple it deserves.
Ten Moats of Vertical Software (and What LLMs Do to Each)
1. Learned interface — Destroyed. Bloomberg users spend years learning shortcuts and codes (GP, FLDS, GIP…). That muscle memory is a moat. When the interface becomes natural-language chat, years of training become worthless and $25K-per-seat switching cost melts away. At Fintool there is no onboarding—users ask in plain English. The chat interface absorbs the whole support structure.
2. Custom workflows and business logic — Evaporated. Vertical software encodes how an industry works into code: thousands of if/then branches, compliance checks, approval flows. At Doctrine, building the legal research workflow took a team of engineers and legal experts years. At Fintool, a DCF valuation “skill” is a Markdown file—written in a week, updated in minutes. Domain experts can encode methodology without writing code. Years of engineering vs. a week of writing.
3. Public data access — Commoditized. Much of vertical software’s value was making hard-to-use data queryable (SEC filings, case law). That required parsers, pipelines, and maintenance. LLMs can parse 10-Ks and case law out of the box. “Making data searchable” is becoming a baseline capability; the premium on that layer is collapsing.
4. Talent scarcity — Reversed. Building vertical software required people who knew both the domain and engineering. That scarcity was a barrier. LLMs make the engineering part cheap; domain experts can now turn expertise into software via Markdown skills. The bottleneck flips.
5. Bundling — Weakened. Vendors locked in clients by bundling adjacent functions (data + news + analytics + trading). AI agents are the new bundle: one workflow can call ten specialized tools. When the integration layer moves from the vendor to the agent, the incentive to buy the bundle shrinks.
6. Private and proprietary data — Stronger. Data that truly cannot be replicated (e.g. Bloomberg’s real-time pricing from trading desks, S&P’s ratings) becomes more valuable, not less. LLMs need scarce inputs. If your data can be obtained, licensed, or synthesized elsewhere, you are at risk. If not, the moat holds.
7. Regulatory and compliance lock-in — Structural. HIPAA, FDA, SOX, audit trails, implementation cycles—none of that changes because of a new Claude plugin. Epic’s position in healthcare EHR is partly a regulatory moat. LLMs may slow adoption where compliance lock-in is strongest.
8. Network effects — Sticky. Bloomberg chat is the comms layer for the street; if everyone is on it, you have to be. LLMs don’t destroy that. Networks may become more valuable as the information flowing through them feeds context and signal.
9. Transaction embedding — Durable. Software embedded in the flow of money (payments, lending, claims) has switching cost that is about revenue interruption. LLMs can sit on top; they don’t replace the rails. Stripe, FIS, Fiserv are not threatened the same way.
10. System of record — Under long-term threat. When your software is the authoritative source for critical data, leaving is existential. LLMs don’t remove that today—but agents are building their own systems of record. Agents read SharePoint, Outlook, Slack; they accumulate rich, persistent memory. Over time, the agent’s context can become a more complete picture than any single SOR. The direction of travel is that agent memory undermines traditional record-system moats.
View on the Selloff
Net effect: barriers to entry collapsed. Five moats are destroyed or weakened; five hold. The five destroyed are the ones that kept competitors out. So the industry doesn’t go from 3 players to 4—it goes from 3 to 300. A small team with frontier APIs, domain expertise, and good data pipelines can build 80% of a vertical product in months. Fintool was built by a team of six and serves funds that used to rely only on Bloomberg and FactSet.
But timing matters. Enterprise contracts are multi-year. Revenue doesn’t vanish overnight; the cliff is a slope over 12–24 months. What the market is pricing is not immediate revenue collapse but multiple compression: when pricing power and retention are seen as eroding, a 15x revenue multiple can become 6x. Same revenue, much lower stock.
The Real Threat: A Pincer Move
From below: hundreds of AI-native startups entering every vertical. From above: horizontal platforms going vertical via AI. Microsoft Copilot does DCF and financial parsing inside Excel, contract review and case research inside Word. Anthropic is doing the same—a general agent toolkit, pluggable data (MCP), and domain skills (Markdown). No domain engineers, no years of build. Software is going “headless”; the interface is the agent. The real threat is horizontal giants like Microsoft moving seriously into verticals because it has never been easier—and in an AI-first world they must own the core workflow.
Risk Framework
High risk: Pure “search layer”—making public or licensable data queryable. If that’s the core value, the moat is gone.
Medium risk: Mix of defensible and exposed lines (e.g. proprietary ratings + repackaged public data).
Lower risk: Regulatory fortresses—HIPAA, FDA, compliance, deep integration into mission-critical workflows.
Three questions: (1) Is the data proprietary? (2) Is there regulatory lock-in? (3) Is the software embedded in transactions? Zero “yes” = high risk; one = medium; two or three = likely okay.
Source: Nicolas Bustamante, “十年构建垂直软件：我对抛售的看法” (Feb 17, 2026). Doctrine, Fintool. Adapted and translated for Claudery.
