The Light We See in Vertical AI
What we look for in vertical AI startups in financial and professional services, when the consensus worries the labs are going to eat it all.
For the past 6–9 months, the consensus has been that vertical AI applications are dead — or are about to be. As the frontier labs develop more capable models, as they go after verticals like coding, financial services and healthcare, and as they build out large teams of forward deployment engineers and Applied AI staffs, and expand their extensive connector ecosystem, many worry that Anthropic and OpenAI are going to eat the lunch of most — if not all — vertical AI application startups.
I agree with part of that. Many thin wrappers will die.
But for the most part, I disagree. And I disagree with huge respect for Anthropic and OpenAI — where many of my friends work, and I believe both will be enormously successful companies.
Because there were so many debates and uncertainties with this question, we made zero investment for the first four months of this year — which felt very uncomfortable at times. But believing in the discipline for investment deployment, we chose to take the time to clarify our thinking rather than rushing towards writing another check.
That meant having private 1:1 chats with insiders working at the frontier labs, observing closely what our portfolio companies were seeing from their customers, and talking with many founders building vertical AI applications. From that, I’ve formed some thoughts below.
These conversations have sharpened our thinking on where we think the best opportunities are in vertical AI in financial and professional services, what we look for in the startups that we back and partner with, and directly resulted in our latest two investments — Narrative (AI for compliance) and Bilrost (AI for commercial real estate underwriting).
I don’t have all the answers, and some of what I put out here could be proven wrong — especially as the AI landscape changes so rapidly, month to month.
But I’ve always believed it’s important to think independently and have a point of view (POV), especially at times that feel uncertain and in a world that’s becoming increasingly consensus-driven. So my intention here is to share our current POV and welcome any thoughts, perspectives, and debate from others, as we keep refining our thinking.
So where are the opportunities we see in vertical AI for financial and professional services? Where is the light?
Below are the six things we look for, including specific examples from startups we are lucky to partner with.
1/ The moat is the full embedded workflow, not automating a single task.
2/ The proprietary layer is the orchestration, not the model.
3/ The action layer is commoditizing; the judgment layer is not.
4/ The best opportunities are the ones the frontier labs won’t prioritize.
5/ The internal data and unspoken rules compound inside the vertical product.
6/ Mid-market can be a more attractive customer base.
#1. The moat is the full embedded workflow, not automating a single task.
The companies I’m most excited about aren’t selling a chat interface on top of a model. They’re rebuilding entire workflows from the ground up.
Bilrost builds AI for commercial real estate underwriting. While early customers spent the past year trying to incorporate ChatGPT or Claude into their workflow themselves, they’ve all abandoned the workaround once they saw Bilrost had already captured the full underwriting workflow: dozens of document types per loan, cross-validation of leases against rent rolls against bank statements, the firm’s own risk and origination guidelines applied to each decision, and direct integration with email, Drive, APIs, and borrower portals.
Because Bilrost is embedded across the entire workflow — intake, classification, extraction, cross-validation, decision-ready output — it has already produced an outcome that’s rare this early: its first customer has 3x’ed its underwriting capacity. In a world where most AI tools still deliver modest efficiency gains, deep workflow embedding has proven its ability to drive material revenue uplift.
Solving the last mile of the day-to-day workflow is what drives adoption and creates stickiness. While the frontier lab optimizes for the average case across many domains, a vertical company optimizes for the long tail of one.
#2. The proprietary layer is the orchestration, not the model.
Accrual (an Elsa Capital portfolio company) builds AI tools for accountants, and this past tax season showed what a vertical company focused entirely on the most mundane, mechanical parts of an accountant’s job can do. It processed more than 1.4 million pages across over 10,000 clients at many of the largest accounting firms in the country. 99.7% of worksheet edits in the platform were drafted by AI; across hundreds of thousands of edits, humans stepped in fewer than 1,000 times. A year ago, every one of those pages would have needed a human to read, extract, and incorporate into a return — and the firms using Accrual closed out meaningfully ahead of where they were last year.
So why won’t Anthropic or OpenAI just do this? Accrual’s founder and CEO, Cosmin Nicolaescu, explained: “the labs will absolutely build tools where the model itself solves the problem, but accounting workflows involve dozens of integrations, compliance rules, edge cases, and process logic accumulated over years — hard to replace with a single model interface.” Ask yourself: are you comfortable letting Claude or ChatGPT fully handle your taxes?
What makes the largest accounting firms trust Accrual is its proprietary layer — the expert instructions and tax-specific orchestration sitting on top of the foundation models. The model is an input, swapped for whatever’s best this month. The defensible asset is that instruction layer, refined against real returns season after season. The model improving is good for them: it upgrades a component they don’t have to build, while the part that’s theirs keeps compounding.
#3. The action layer is commoditizing. The judgment layer is not.
Reading a document, extracting a field, drafting a response — these outputs get better and cheaper every quarter as the models improve. What doesn’t commoditize is the unwritten judgment that drives decisions inside a regulated workflow. Most consequential decisions in a bank aren’t rules-driven; they live in the heads of a few experienced people. A frontier model can tell you what the regulation says. It can’t tell you how this firm weighs a marginal complaint against a top-tier partner relationship, or which exceptions the senior underwriter has been quietly making for fifteen years.
Narrative, which builds AI-powered compliance and risk management for financial services firms, solves this gap. One layer does the general actions — read, extract, summarize — and commoditizes as the model providers improve. The other encodes what isn’t written down: the institutional judgment that makes the output actually valuable.
Because Narrative encodes that judgment, it frees its customers from the “judgment constraint” that has historically bottlenecked how fast financial services firms can grow. That’s why, even as a seed-stage company, it has signed seven-figure contracts with bank customers.
Along a similar line, I recently heard Winston Weinberg, founder and CEO of Harvey — the legal AI startup now with over $200M ARR — explain their moat. Customers pay for two things: the work product and the judgment. “The first will get commoditized and the latter will not,” he said.
The successful vertical AI startups bundle the product and the judgment, and that bundle is the value.
#4. The best opportunities are ones the frontier labs won’t prioritize.
The durable companies build in markets too niche, too regulated, or too operationally complex for the labs to care about. Not niche as in small — commercial lending and compliance are massive. Niche as in the workflow demands deep domain knowledge, proprietary context, and years of relationship-building. These markets have been underserved so long that when a founder finally shows up and cares, the reaction is visceral. One founder told us a billionaire customer flew to San Francisco just to meet them after seeing the demo. That pull doesn’t happen in well-served markets.
In addition, what we’ve heard from friends who’ve worked at the frontier labs is that winning a vertical takes more than a well-designed product. To sell into these verticals, the labs would most likely need to stand up a deep, expert sales team aimed squarely at that buyer. That organizational lift — not the product — is why so many verticals stay open to startups whose entire GTM motion is dedicated to one kind of customer.
#5. The internal data and unspoken rules compound inside the vertical product.
The frontier labs train on external, public data. But ~90% of all data lives inside a company, and on top of that there’s an enormous body of unspoken rules, institutional judgment, and firm-specific preferences that were never written down. None of that is on the open internet for a model to learn. It only accrues to a vertical product that embeds deeply enough in the workflow to see it, capture it, and act on it.
So every workflow execution makes the system understand that customer better — its quirks, its exceptions, the things no one wrote down. This isn’t model weights getting smarter for everyone; it’s a decisioning engine getting dialed in on one institution, and it can’t be shortcut. Narrative sharpens on a firm’s risk profile with every run. Bilrost tunes nightly against each customer’s historical loans. Accrual built a loop that reconciles every draft prepared in the platform against the final filed return — reviewed and corrected by human accountants — and ships fixes before the next season. The work doesn’t reset each year; every returning client hands the system a richer starting point.
A model release doesn’t replicate that, because the most valuable asset — a customer’s own data and the unspoken rules around it, compounding over time — isn’t in the model. It’s in the vertical product that owns the end-to-end workflow.
#6. Mid-market can be a more attractive customer base.
Plenty of great vertical companies win in the enterprise. But for many verticals, mid-market can be the more attractive base: underserved, faster to close, and far less contested. The Fortune 100 is precisely where Anthropic, OpenAI, and the best-funded startups all concentrate their attention; the mid-market sits well outside that spotlight. And the economics aren’t a step down. Solve a real pain point and the ACV can run $500K to seven figures even for a pre-seed or seed-stage startup — signed much faster than a Fortune 100 procurement cycle, where the buyer also has a frontier lab’s engineers on speed dial.
There are countless unsexy, unloved verticals and mid-size customers whose names I’d never heard before this work — starved of attention from AI startups, and glad to pay seven figures to one that cares enough to solve their problem.
Final thoughts
Application software was never defensible on its own. The defensibility was never going to come from the technology itself. It comes from the orchestration, the judgment layer, the compounding data inside each account, the customers the labs ignore, and solving the last mile and edge cases of the day-to-day workflow.
AI will automate pieces of the work. The real, more exciting opportunity is redesigning how the work gets done — rebuilding the entire workflow from the ground up. The frontier labs supply the first half. The founders who know the work, and dedicate their lives to a single vertical, own the second.
That’s where we believe the light is in vertical AI applications.
If you are building in the space, or know a great founder who is, we’d love to hear from you. You can reach out here (if you’re a founder pitching) or email us directly.