The Vertical Software Reckoning

Notes from reading I've done on how LLMs are reshaping vertical software. The ten moats framework comes from a practitioner who has built both traditional vertical SaaS and AI-native competitors to Bloomberg and LexisNexis.

Context

Nearly $1 trillion wiped from software and services stocks in recent weeks. FactSet: $20B peak to under $8B. S&P Global: down 30%. Thomson Reuters: nearly half its market cap gone in a year. S&P 500 Software & Services Index (140 companies): down 20% YTD.

The argument is that LLMs are dismantling the moats that made vertical software defensible. Not all of them, but enough to redraw what deserves a premium multiple.

Vertical Software Basics

Industry-specific software: Bloomberg (finance), LexisNexis (legal), Epic (healthcare), Procore (construction), Veeva (life sciences). High prices ($15-25K/seat/year), ~95% retention, very few competitors per vertical. Ten distinct moats have historically made these businesses defensible.

The Ten Moats

Moats That Are Breaking

1. Learned Interfaces. Bloomberg users spend years learning GP, FLDS, GIP function codes. That muscle memory was a massive switching cost. LLMs collapse all proprietary interfaces into one: chat. Three sentences of natural language replace years of learned navigation. For many companies, the interface was most of the value. The underlying data was public or licensed.

2. Business Logic. Vertical software encodes industry workflows in code: thousands of if/then branches, validation rules, compliance checks. This required rare engineers who understood both the domain and the technology, and took years to build. LLMs turn this into a markdown file. A domain expert can encode their methodology in plain English in a week. It's readable, auditable, customizable per user, and improves automatically as the model improves.

3. Public Data Access. A huge portion of vertical software's value was making hard-to-access public data queryable: parsing SEC filings, structuring case law, normalizing court records. Companies built thousands of custom parsers and NLP pipelines. Frontier models already know how to parse a 10-K from training data. The model is the parser. The "making it searchable" layer, where pricing power lived, is collapsing.

4. Talent Scarcity. Building vertical software required people who understood both domain and technology, which is extremely rare. LLMs flip this. Domain experts write methodology directly in plain English, the model handles the engineering. The scarce resource (domain expertise) can become software without an engineering bottleneck.

5. Bundling. Bloomberg expanded from market data into messaging, news, analytics, trading, compliance. Each module increased switching costs. LLM agents break this because the agent is the bundle. It orchestrates across multiple providers in a single workflow, picking the best or cheapest for each capability.

Moats That Hold

6. Proprietary Data. If data genuinely can't be replicated (Bloomberg's real-time trading desk pricing, S&P's credit ratings, D&B's 500M+ entity files), LLMs actually make it more valuable. It becomes the scarce input every agent needs. The caveat is that MCP is turning every data provider into a plug-in. If your data isn't truly unique, you become a commodity supplier competing on price. Aggregation theory plays out: the agent captures the user relationship, data vendors compete on price.

7. Regulatory Lock-in. HIPAA doesn't care about LLMs. FDA certification doesn't get easier. Epic's dominance is a regulatory moat: 18-month implementations, compliance certifications, patient safety risks of switching. Regulatory requirements may actually slow LLM adoption in the verticals where this moat is strongest.

8. Network Effects. Bloomberg IB chat is Wall Street's communication layer. If every counterparty uses Bloomberg, you must too. LLMs don't break network effects because the value comes from who else is on the platform, not from the interface.

9. Transaction Embedding. Software in the money flow: payment processing, loan origination, claims processing. An LLM might sit on top as a better interface, but the rails remain essential. Stripe, FIS, Fiserv aren't threatened.

10. System of Record. Not directly threatened today, but agents are quietly building their own. They read SharePoint, Outlook, Slack, accumulate memory across sessions. The agent becomes the one layer that sees everything. Salesforce sees CRM. Outlook sees email. The agent sees all of it, and remembers.

The Scorecard

The five that break are the ones that kept competitors out. The five that hold are the ones that only some incumbents have.

Consequences

Barrier to Entry Collapses

Before LLMs: building a Bloomberg competitor required hundreds of domain-specialist engineers, years of development, massive data licensing, enterprise sales teams. Most verticals had 2-3 serious competitors.

After LLMs: a small team with frontier model APIs and domain expertise can build 80% of the capability in months. Competition goes from 3 to 300. Fifty AI-native startups offering 80% of the capability at 20% of the price craters pricing power.

The Pincer Movement

From below, hundreds of AI-native startups are entering every vertical.

From above, horizontal platforms are going vertical for the first time. Microsoft Copilot in Excel does DCF modeling. Copilot in Word does contract review. Anthropic's stack (agent SDK + MCP + markdown skills) is the entire toolkit needed to go from horizontal to vertical.

Software is becoming headless. The interface disappears. What matters is owning the customer relationship, owning the agent itself.

Timing

Enterprise revenue doesn't disappear overnight. Multi-year contracts, 12-18 month procurement cycles. Current revenue is locked in for 12-24 months.

But you don't need revenue to decline for the stock to crash. You need the multiple to compress. A company at 15x revenue with strong moats reprices to 6x when those moats erode. Revenue stays flat. Stock drops 60%. The market isn't pricing in a revenue collapse. It's pricing in the end of the premium multiple.

Risk Assessment Framework

Three questions for any vertical software company:

1. Is the data proprietary? If not, the accessibility layer is collapsing.
2. Is there regulatory lock-in? If not, switching costs are interface-driven and dissolving.
3. Is the software embedded in the transaction? If not, you're replaceable.

The more "yes" answers, the safer the company. Zero means the moats LLMs are breaking are all you have.

Bottom Line

This isn't about all vertical software dying. It's the market distinguishing between companies that own something genuinely scarce (proprietary data, regulatory certification, transaction rails, network effects) and companies whose value was a learned interface over semi-public data. The first group gets stronger. The second is in existential trouble.