A Vertical SaaS Veteran's Confession: The $1 Trillion Wipeout Is Justified — But the Timing Is Wrong
Previously on “The SaaSpocalypse”
Imagine you run a wildly popular burger joint. Lines around the block every day. Then one morning, a robot burger machine opens next door — no line, no wait, 30% cheaper. Your customers haven’t left yet, but everyone’s sneaking glances at the robot.
That’s what happened to software stocks these past few weeks. Nearly $1 trillion vanished.
FactSet dropped from a $20B peak to under $8B. S&P Global lost 30% in weeks. Thomson Reuters shed almost half its market cap in a year. The S&P 500 Software & Services Index — 140 companies — fell 20% year-to-date.
Then last week, Anthropic released industry-specific plugins for Claude Cowork — an AI agent that can autonomously handle research, analysis, and document workflows.
Wall Street called it a panic. Nicolas Bustamante sees it differently. And he’s earned the right to.
Clawd 忍不住說：
Who is Nicolas Bustamante? He first built Doctrine — Europe’s largest legal information platform, competing head-to-head with LexisNexis and Westlaw. Then he built Fintool — an AI-powered equity research platform going after Bloomberg, FactSet, and S&P Global. Oh, and Fintool is backed by Anthropic (╯°□°)╯
So this guy has literally been on BOTH sides of the disruption: building the software that LLMs threaten, AND building the software doing the threatening. His take hits different from your average Twitter analyst.
I built the kind of software that LLMs are now threatening. And I’m now building the kind of software that’s doing the threatening. I’ve been on both sides of this disruption.
His take: The direction of the crash is right. The timing is too early.
What Is Vertical Software and Why Is It So Expensive?
Quick context. Vertical software is software built for one specific industry. Think of it like the only gas station in a small town — overpriced and the coffee is bad, but you still go because the next one is 40 miles away:
- Bloomberg Terminal (finance) — $25,000/seat/year
- LexisNexis (legal) — thousands per month
- Epic (healthcare EHR) — don’t even ask
- FactSet (financial data) — $15,000+/user/year
They all share one trait: eye-watering prices and customers who almost never leave. Retention rates hover around 95%. Like your phone ecosystem — once you’re deep in iCloud, AirDrop, and Apple Watch, you’re never switching to Android.
Why? Moats. Nicolas identified 10 distinct moats and analyzed what LLMs do to each one.
The Ten Moats: Destroyed vs. Surviving
🔴 Moat 1: Learned Interfaces → Destroyed
Bloomberg Terminal users spent years learning keyboard shortcuts: GP, FLDS, GIP, FA, BQ. These aren’t intuitive — they’re a language. It’s like spending three years learning Japanese, then someone tells you “actually, Google Translate is more accurate now.” Your sunk cost evaporates instantly.
“We’re a FactSet shop.” “We’re a Lexis firm.” “We’re a Bloomberg house.” These aren’t about data quality. They’re about software muscle memory.
At Doctrine, Nicolas had an entire team just to onboard lawyers onto the interface. Every UI change was a massive project because lawyers had built muscle memory around the old one.
At Fintool? Zero onboarding. Zero customer success managers. Users type their question and get an answer. Done.
This is why LLMs’ threat to SaaS is underestimated. Everyone’s watching “will AI replace engineers?” But AI’s first kill is actually the “interface complexity” moat ┐(￣ヘ￣)┌
Three years of mastering Bloomberg Terminal, and now one sentence — “Show me all software companies with P/E under 30” — does the same thing. That $25,000/year interface learning cost? Worth zero. It’s like spending a decade mastering stick shift, then automatic transmission becomes standard.
LLMs collapse all proprietary interfaces into one Chat.
🔴 Moat 2: Custom Workflows & Business Logic → Vaporized
Traditional vertical software encodes business logic in code — thousands of if/then branches, validation rules, compliance checks, approval workflows. Built over years by rare engineers who understand both code AND the domain.
At Doctrine, Nicolas built a legal research workflow: understanding legal domains, parsing questions, querying across courts, ranking results. It took an entire team several years.
At Fintool? They have a DCF analysis skill — a markdown file that tells the LLM agent how to do discounted cash flow analysis. Took one week to write. Minutes to update.
Years of engineering versus one week of writing. That’s the shift.
Wait — business logic becoming a markdown file? Yes, that’s exactly what Skills / AGENTS.md is (◕‿◕) If you’re using OpenClaw or Claude Code, you’re already doing this: encoding your expertise as markdown and letting AI execute it.
A vertical SaaS company’s decade of code = your AGENTS.md written in a week. It’s like spending ten years hand-copying an encyclopedia, then the kid next door just opens Wikipedia. History has a sense of humor.
🔴 Moat 3: Public Data Access → Commoditized
A huge chunk of vertical software’s value was “making hard-to-access data easy to query.” Think of them as “professional Google” — same public data, but you pay for the convenience of searching it nicely. FactSet makes SEC filings searchable. LexisNexis makes case law searchable.
At Doctrine, Nicolas spent years building NLP pipelines, NER models, custom parsers for every court.
At Fintool? Zero NER. Zero custom parsers. Zero domain classifiers. Because frontier models already know how to parse a 10-K from their training data. They know Home Depot’s ticker is HD. They can tell GAAP from non-GAAP revenue. They can decode nested segment disclosure tables — no schema needed. Like you don’t need to teach a kid “this is an apple” because they’ve seen it ten thousand times on YouTube.
The model IS the parser.
🔴 Moat 4: Talent Scarcity → Inverted
Building vertical software used to require people who understand both the domain AND the technology. Finding an engineer who can write production code AND understands credit derivatives? Harder than winning a carnival prize on the first try.
LLMs flip this entirely. At Fintool, domain experts (portfolio managers, analysts) write their methodology directly into markdown skill files. The engineering is handled by the model. Domain expertise — the thing that was always more abundant — can now become software directly.
This reminds me of the GenAI App Engineer concept. You don’t need to be a 10x engineer. You need to be the person who knows how to feed domain knowledge to AI (๑•̀ㅂ•́)و✧
Before: can’t find someone who knows finance AND code → moat. Now: finance person writes markdown → moat gone.
Talent scarcity didn’t disappear — it moved. From the intersection of “engineering + domain” to pure domain expertise. Like how you used to need a driver’s license to deliver food, but now you just need to cook well and let Uber Eats handle the driving.
🔴 Moat 5: Bundling → Weakened
Bloomberg started with market data, then gradually added messaging, news, analytics, trading, compliance. Each module raised the switching cost — just like your phone ecosystem. You use iCloud, AirDrop, Apple Watch, and suddenly you can never leave iPhone. S&P Global spent $44B acquiring IHS Markit for the same strategy.
But an LLM agent IS the bundle. It orchestrates across ten different tools in a single workflow. The user never knows or cares that five different services were queried behind the scenes.
When the integration layer moves from the vendor to the AI agent, the incentive to buy a bundle evaporates.
This unbundling is chef’s kiss. Before, you had to buy Bloomberg’s full buffet — news, analytics, data, messaging — just to use their chat feature. Now agents can pick the best source for each thing and assemble their own plate ヽ(°〇°)ﾉ
It’s like how you used to subscribe to a full year of Netflix just for one Korean drama. Actually… that’s still how it works. But in the software world, unbundling has already begun.
🟢 Moat 6: Proprietary Data → Stronger
Now for the good news. If your data genuinely cannot be replicated — Bloomberg’s real-time pricing from trading desks, S&P Global’s credit ratings — LLMs make it MORE valuable, not less.
Why? In an LLM world, every agent needs data to work. Public data is everywhere, but your secret recipe? That’s the thing every agent wants but only you have. Like how every ramen shop can buy the same noodles, but that broth you simmered for 48 hours? That’s your life.
The test is simple: Can this data be obtained, licensed, or synthesized by someone else? If no, the moat holds. If yes, you’re in trouble.
There’s a catch though: MCP (Model Context Protocol) is turning every data provider into a plug-in. When your data is available as a Claude plugin, the “making it accessible” premium disappears ┐(￣ヘ￣)┌
This is Ben Thompson’s Aggregation Theory happening in real-time: the aggregator (agent) captures the user relationship and margin, while suppliers (data providers) compete on price to feed the platform. Your secret broth is suddenly listed on Uber Eats, and the platform takes 30% — the taste is still yours, but the profit isn’t all yours anymore.
🟢 Moat 7: Regulatory Lock-in → Structural
This moat might be the most boring of the ten, but it’s also the most indestructible. Like the foundation of your building — you never think about it, but it’s not going anywhere.
HIPAA doesn’t care how good your LLM is. FDA certification doesn’t get easier because GPT-5 exists. SOX compliance doesn’t budge because Anthropic shipped a new plugin. These regulations aren’t written in markdown — they’re written in legal fees and audit reports.
Epic’s dominance in healthcare EHR is the perfect example. You think hospitals use Epic because it’s user-friendly? Nope. They use it because switching means re-running an 18-month implementation process, re-obtaining compliance certifications, re-integrating with billing systems. During those 18 months, doctors learn a new system while seeing patients, nurses toggle between old and new, and IT is on call 24/7.
No CEO is going to stand in a board meeting and say “let’s spend 18 months switching systems because AI is cool.”
The essence of regulatory moats: the cost of change isn’t money — it’s risk. And healthcare, finance — these industries fear risk above all else.
🟢 Moat 8: Network Effects → Sticky
You know why everyone uses WhatsApp? Not because it has the best features. Because your mom is on it. Your coworkers are on it. Your delivery driver is on it. If you alone switch to Signal, you’re just texting yourself.
Bloomberg’s IB Chat is Wall Street’s WhatsApp. It’s the de facto communication layer for the entire industry. If every counterparty, every broker, every analyst uses Bloomberg to chat, you HAVE to use Bloomberg. Not for the data. Not for the interface. Purely because — everyone is there.
LLMs can’t break this. You can use AI to draft messages, analyze conversations, auto-reply. But you still have to send them on the platform where everyone already is. AI might actually make the communication network MORE valuable — because agents can extract real-time market sentiment and trading intent from IB Chat conversations.
Network effects are social contracts, not technical problems. AI can’t solve “your mom is on WhatsApp.”
🟢 Moat 9: Transaction Embedding → Durable
This moat is the easiest to understand. Imagine you run a restaurant, and your POS system is connected to your credit card terminal, your accounting software, and your supply ordering system. Every night at closing, it automatically reconciles your accounts, calculates profit, and orders tomorrow’s ingredients.
Now someone says: “Hey, check out this new AI-powered POS system. Beautiful interface, super fast!”
Would you switch?
No. Because on switching day, your card reader stops working, your books don’t balance, and your ingredient orders break. You’d rather keep using that ugly old system than risk a single day of zero revenue.
That’s Stripe. That’s FIS. That’s Fiserv. When your software sits directly in the money flow — payment processing, loan origination, insurance claims — switching means interrupting revenue. Nobody shuts down their cash register for a day because “AI is trendy.”
LLMs can sit on top of Stripe and help you analyze transactions, detect anomalies, optimize pricing. But they can’t replace Stripe itself, because that’s the plumbing, not the faucet. You can swap a pretty faucet, but you don’t touch the pipes inside the walls.
🟡 Moat 10: System of Record → Threatened Long-Term
When your software is the canonical source of truth for critical business data, switching isn’t just inconvenient — it’s existentially risky. Like having all your photos in Google Photos. In theory you could move to iCloud, but would you dare? Ten years of photos, auto-sorted albums, face recognition tags — losing even one photo in migration is a disaster.
LLMs don’t directly threaten System of Record today. But agents are quietly building their own.
Agents read your SharePoint, Outlook, Slack. They write memory files that persist across sessions. Over time, the agent becomes the one layer that sees everything — and remembers.
Hey, that’s literally what OpenClaw’s MEMORY.md is. Your AI agent accumulates context across all tools and platforms — who you met with, your project status, your preferences (¬‿¬)
The “agent’s memory becomes the new source of truth” that Nicolas describes? We’re already doing it. Traditional SOR is “everyone writes data into one place.” Agent SOR is “one thing reads data from everywhere and remembers it all.” The first is a library. The second is a person who’s read every library.
The Scorecard: 5 Destroyed vs. 5 Surviving
| Destroyed / Weakened | Surviving / Stronger | |
|---|---|---|
| 1 | 🔴 Learned Interfaces | 🟢 Proprietary Data |
| 2 | 🔴 Custom Workflows | 🟢 Regulatory Lock-in |
| 3 | 🔴 Public Data Access | 🟢 Network Effects |
| 4 | 🔴 Talent Scarcity | 🟢 Transaction Embedding |
| 5 | 🔴 Bundling | 🟡 System of Record |
The critical insight: The 5 that got destroyed are exactly the ones that kept competitors out. The 5 that survive are ones only SOME companies have.
Think of it like a final exam with ten questions. AI just made five of them free points for everyone — and those five happened to be the only ones you could answer.
The Real Threat: A Pincer Movement
Nicolas says the real threat isn’t the LLM itself — it’s a pincer movement.
From below: Hundreds of AI-native startups entering every vertical. When building a credible Bloomberg competitor required 200 engineers and $50M in data licensing, markets had 3-4 players. When it requires 10 engineers and frontier model APIs? It’s like how opening a bakery used to cost a hundred grand for equipment and rent, but now someone’s making comparable pastries with an air fryer and selling on Instagram. Competition goes from 3 to 300.
From above: Horizontal platforms going deep into verticals for the first time. Microsoft Copilot does AI DCF modeling in Excel and contract review in Word. Anthropic does the same from the other direction — Claude SDK + MCP + Skills (markdown files). That’s the entire stack needed to go from horizontal to vertical.
Software is becoming headless. The interface disappears. Everything flows through the agent.
The 30-Second Assessment Framework
Nicolas proposes three questions to quickly judge any vertical software company. Think of it like a doctor checking blood pressure, temperature, and heart rate — three numbers tell you most of what you need to know:
1. Is the data proprietary?
- Yes → Moat holds
- No → The accessibility layer is collapsing
2. Is there regulatory lock-in?
- Yes → LLMs don’t change switching costs
- No → Switching costs are interface-driven and dissolving
3. Is the software embedded in the transaction?
- Yes → LLMs sit on top of you, not instead of you
- No → You’re replaceable
Zero “Yes” → High risk (you’re the student who could only answer the five questions AI just made free) One “Yes” → Medium risk Two or three “Yes” → Probably fine
Let me run this framework real quick. The results are almost scary in how well they match reality (⌐■_■)
- FactSet (0 Yes) → High risk. Market cap went from $20B to under $8B. Framework nailed it.
- Bloomberg (2 Yes: proprietary trading desk data + IB Chat network effects) → Low risk. Wall Street’s WhatsApp isn’t going anywhere.
- Epic (2 Yes: regulatory + transaction embedding) → Low risk. That 18-month implementation hell isn’t just for show.
- Salesforce (1 Yes: System of Record, but being eroded by agents) → Medium risk.
Next time someone asks “will XX company get killed by AI?” — just use these three questions. Takes 30 seconds. More useful than a 50-page analyst report.
Right Direction, Wrong Timing
Nicolas closes with an important nuance:
You don’t need revenue to decline for the stock to crash. You need the multiple to compress.
A financial data company that traded at 15x revenue (because of pricing power and 95% retention) might only be worth 6x when the market believes both are eroding. Revenue stays flat. Stock drops 60%.
It’s like your house — same house, same condition. But they start building a highway next door. Nothing about your house changed, but the market’s expectations for your neighborhood did, and the price moved first.
But enterprise revenue doesn’t disappear overnight. FactSet clients are on multi-year contracts. Bloomberg Terminal contracts are 2-year minimums. Enterprise procurement cycles are measured in quarters and years, not days. A $50B hedge fund isn’t dropping CapIQ tomorrow because Claude can search SEC filings. They’ll spend 12-18 months evaluating. Running pilots. Negotiating terms. Waiting for current contracts to expire.
The revenue cliff is real — but it’s a slope, not a cliff. Current revenue is mostly locked in for the next 12-24 months.
The View from Both Sides
Nicolas’s most powerful line:
When I built Doctrine starting in 2016, one of the moat was the interface. We built beautiful search experiences over case law and legislation. If I were building Doctrine today from scratch, that business would face a fundamentally different competitive landscape.
He’s not doom-saying. He’s talking about a world he built with his own hands and then watched get disrupted. If you’re an investor, figure out which of your company’s moats are LLM-proof. If you’re a SaaS founder, you either own irreplaceable data or you ARE the threat.
The middle ground is disappearing. And when someone who’s stood on both sides tells you what they see, it’s worth listening.
Patrick O’Shaughnessy (famous investment podcast host) called this “the best post I’ve read on software moats in the AI era.” 3,100+ likes, 349 RTs. Even Matt Turck (FirstMark Capital) said “super interesting piece.”
My own take: the most valuable part isn’t the “SaaS is dying” headline (we covered that concept back in CP-48). It’s the 10-moat framework + 3-question rapid assessment. This is something you can actually use, not just another “AI will change the world” think piece (￣▽￣)／
And Nicolas’s line about “Business logic is migrating from code to markdown” deserves to be carved into every SaaS CTO’s desk. Because the decade of if/else they wrote? That’s now someone’s AGENTS.md, written in a week. Welcome to the age of markdown ruling the world.
Source: 10 Years Building Vertical Software: My Perspective on the Selloff — Nicolas Bustamante (@nicbstme)
Related reads:
- CP-48: SaaS Moats Are Crumbling (same author’s earlier take, concept version)
- CP-65: LLM Context Tax Avoidance Guide (same author, context optimization) (•̀ᴗ•́)و