06-reference

vertical software selloff

Sun Feb 15 2026 19:00:00 GMT-0500 (Eastern Standard Time) ·article ·source: x.com/@nicbstme ·by Nicolas Bustamante
vertical-saasllmsmoatsdisruptiondata-consultingfintechlegaltechai-agents

10 Years Building Vertical Software: My Perspective on the Selloff

Nicolas Bustamante (@nicbstme) built Doctrine (Europe’s largest legal information platform) and then Fintool (an AI equity research platform competing with Bloomberg, FactSet, and S&P Global). He’s been on both sides of this disruption — building the software LLMs threatened, and then building the software doing the threatening. Published February 2026 during the ~$1T vertical software stock selloff.

The Core Thesis

LLMs are systematically dismantling some of the moats that made vertical software defensible — but not all of them. The selloff is directionally correct but temporally exaggerated. The market isn’t pricing in revenue collapse; it’s pricing in the end of the premium multiple, because the moats that justified 15-20x revenue are dissolving.

The real threat is a pincer movement: hundreds of AI-native startups entering every vertical from below, and horizontal platforms (Microsoft Copilot, Anthropic Claude) going deep into vertical territory from above — something that was previously prohibitively expensive but is now trivially achievable with LLMs + skills + MCP.

The Ten Moats — What Survives, What Doesn’t

Destroyed or Severely Weakened

1. Learned Interfaces — Destroyed. Bloomberg Terminal users spent years memorizing function codes (GP, FLDS, BQ). That muscle memory was a switching cost that justified $25K/seat/year. Chat-based LLM interfaces collapse all proprietary UIs into one. Years of learned behavior become worthless overnight. Fintool has no onboarding, no CSMs, no UI change management — the chat interface absorbs all that scaffolding.

2. Custom Workflows and Business Logic — Vaporized. Vertical software encodes how industries actually work — litigation workflows, DCF model conventions, compliance checks. This business logic took specialized engineers (rare: people who write production code AND understand financial derivatives) years to encode across thousands of lines. Now it’s a markdown skill file. Bustamante’s Fintool DCF valuation skill took one week to write. Updating it takes minutes. It automatically improves as the underlying model improves. Business logic is migrating from code to markdown, written by domain experts without engineers.

3. Public Data Access — Commoditized. FactSet’s and LexisNexis’s core value was making technically-public-but-hard-to-use data searchable. Building that required armies of parsers, NLP pipelines, domain-specific NER models. Frontier models already know how to parse a 10-K — they were trained on them. The parsing layer is now a commodity baked into the model. Additionally, MCP is turning every data provider into a plug-in, eliminating the “making it accessible” premium.

4. Talent Scarcity — Inverted. Vertical software required engineers who understood the domain. This scarcity limited competition to 2-3 serious players per vertical. Now domain experts write methodology directly into markdown skill files, bypassing the engineering bottleneck entirely. The barrier to entry collapses from “hundreds of engineers + years” to “10 engineers + frontier model APIs + a few months.”

5. Bundling — Weakened. Bloomberg grew by bundling messaging, news, analytics, trading, and compliance into one lock-in machine. LLM agents break this because the agent IS the bundle — it orchestrates across multiple specialized tools in a single workflow. Customers no longer need one vendor’s entire suite when an agent can cherry-pick the best provider for each capability.

Intact or Stronger

6. Private and Proprietary Data — Stronger. Data that genuinely can’t be scraped, synthesized, or licensed elsewhere becomes MORE valuable. Bloomberg’s real-time trading desk pricing data. S&P Global’s credit ratings (a regulated opinion backed by decades of default data). The test: can this data be obtained, licensed, or synthesized by someone else? If not, the moat holds and may actually widen — the scarce input every agent needs.

7. Regulatory and Compliance Lock-in — Structural. HIPAA doesn’t care about LLMs. FDA certification doesn’t get easier because GPT-5 exists. Healthcare EHR systems, life sciences platforms, and financial compliance infrastructure may actually slow LLM adoption in exactly the verticals where compliance lock-in is strongest.

8. Network Effects — Sticky. Communication networks within an industry (Bloomberg IB chat, Veeva’s pharma network, Procore’s construction stakeholder network) become more valuable as more participants use them. LLMs don’t break these — they may strengthen them.

9. Transaction Embedding — Durable. Software embedded directly in the money flow (payment processing, loan origination, claims processing) is infrastructure, not interface. LLMs may sit on top as a better UI, but the rails themselves remain essential.

10. System of Record Status — Threatened Long-Term. AI agents are quietly building their own system of record — they read Slack, Outlook, SharePoint, and accumulate richer contextual memory than any single SoR. This is directional, not immediate. Epic isn’t going away this year.

The Framework: Three Diagnostic Questions

For any vertical software company:

  1. Is the data proprietary? If yes, the moat holds. If no, the accessibility layer is collapsing.
  2. Is there regulatory lock-in? If yes, switching costs are structural. If no, they’re interface-driven and dissolving.
  3. Is the software embedded in the transaction? If yes, LLMs sit on top. If no, you’re replaceable.

Zero “yes” answers: high risk. One: medium risk. Two or three: probably fine.

Risk Tiers

High risk: Search-layer plays — financial data terminals built on licensed exchange data, patent search tools, anything where the core value prop was “better search for your industry’s data.” These traded at 15-20x revenue on interface lock-in and limited competition. Both are evaporating.

Medium risk: Mixed portfolios with some defensible and some exposed business lines. The selloffs of 20-30% reflect market uncertainty about which segments dominate valuation.

Lower risk: Regulatory fortresses — healthcare EHR, life sciences platforms, financial compliance infrastructure. May actually benefit as customers consolidate around trusted regulated-workflow vendors while switching away from information retrieval vendors.

The Selloff Diagnosis

FactSet dropped from $20B to under $8B. S&P Global lost 30% in weeks. Thomson Reuters shed nearly half its market cap. The 140-company S&P 500 Software & Services Index fell 20% YTD. Bustamante argues the market is right on direction, wrong on timing: enterprise contracts are multi-year minimums (Bloomberg Terminal: 2-year; procurement cycles: 12-18 months). Revenue is largely locked in for the next 24 months. But the premium multiple justified by moats that are now dissolving is appropriately repricing — the stock can drop 60% while revenue stays flat if the market is removing a moat premium.

His contrarian observation: some mid-sized companies with genuine proprietary data moats that adopted AI early will capture disproportionate value versus both incumbents (complacent) and new entrants (no proprietary data).

Actionable / Consulting Application

This analysis maps directly to both the phData consulting role and Ray Data Co’s positioning.

For phData client conversations:

The moat framework is a selling tool. Enterprise clients with vertical software spend are trying to figure out where they’re exposed. The three diagnostic questions (proprietary data? regulatory lock-in? embedded in transaction?) translate directly into a consulting assessment: “Here’s your vertical software portfolio — here’s what you should protect, what you should challenge, and where AI-native alternatives deserve a pilot.” This is the kind of strategic framing that differentiates Principal Consultant advisory work from tool implementation.

The skill-as-markdown-file insight is particularly powerful for the Snowflake/Cortex AI context: organizations don’t need specialized engineers to encode business logic anymore. Domain experts can write methodology directly. The consulting play is helping enterprises build their first skill libraries — capturing tribal knowledge that currently lives in spreadsheets, heads, and PDFs — before AI-native competitors do it for them.

For Products for Agents and the Data Marketplace:

Bustamante’s proprietary data thesis is the intellectual foundation for the data marketplace bet. If “making data searchable” is commoditized but “owning data that can’t be replicated” gains value, then curated, exclusive, or hard-to-assemble datasets become the premium layer. The data marketplace model — licensing unique datasets for agent consumption — is exactly the bet that this analysis validates.

For Ray Data Co positioning more broadly:

The consulting career question is whether to be the interface layer (fragile, commoditizing) or the proprietary-data/domain-expertise layer (durable, appreciating). The phData role is explicitly on the right side of this — designing AI agent systems on Snowflake Intelligence for enterprise clients, which is workflow design and judgment work, not interface-layer work. This is a position the market will pay more for as the interface layer gets eaten.

Connections