In 2023 a company called Harvey started selling AI into law firms, and they did it in a way that sounded almost boring. A private deployment trained on each firm's own documents, running where the firm wanted it, used by lawyers the way they already worked. No migration. No rip and replace. Nobody had to transform anything.
Three years later Harvey is worth $11 billion, has raised over a billion dollars, and most of the AmLaw 100 pays them. Over 100,000 lawyers across 1,300 organizations. Sequoia co-led three rounds in a row, which they basically never do.
I think about Harvey a lot, because the conditions that made legal work are sitting right there in investment management and nobody has built the equivalent yet. So I'm building it.
The thesis, in one paragraph
Vertical AI wins where the work is document heavy, judgment intensive, and stuck in a workflow nobody has touched in a generation. Legal checked every box. Investment management checks every box. The firm that gets private, trained, structured AI infrastructure running first pulls ahead of every firm that waits, and the gap widens with every deal that runs through it. The infrastructure is the moat, and the data sitting on top of it makes the moat deeper every quarter.
That's the whole idea. The rest of this post is what it means in practice, and why most of what currently calls itself AI for investment teams doesn't qualify.
What everyone else is selling
There are plenty of products marketing themselves as AI for investment teams right now. Look closely and most of them are doing one of three things.
Retrieval over PDFs. A chat window pointed at a folder of documents. Genuinely useful for "find me the rent roll from the Atlanta deal." Useless for "what's our actual underwriting standard on suburban office in tertiary markets, and how has it moved since 2023." The first question is search. The second needs structure, and a folder doesn't have any.
Workflow templates. Pre-built screens and drag-and-drop pipelines. They demo beautifully. Then the firm tries to encode its own underwriting standards instead of the vendor's defaults and the whole thing fights back. Your criteria are your edge. A tool that forces you into someone else's template flattens the one thing that makes you different.
Horizontal AI in a vertical costume. A general model wrapped in a finance-flavored UI. Easy to ship, easy to copy, easy to leave. Nothing about it gets better the longer your firm uses it, which tells you everything about how durable it is.
Antonine is none of those.
What Antonine actually is
Antonine is four products on a fleet of agents, built for the lean deal team: private equity, private credit, family offices, and real estate investors who are expected to run the screening and diligence volume of firms three times their size. DealFilter screens inbound teasers against what your firm actually buys, before an NDA is ever signed. Diligence takes over after the NDA, ingesting the full CIM and deal package and drafting a first-pass IC memo in your firm's own format with citations back to source. The Vault is the firm's institutional memory, where every deal, memo, and decision becomes permanent and searchable. Pipeline ties it together with a single view of where every deal stands, from inbox to decision.
The Vault is the part that matters competitively. Most document AI is retrieval. The Vault is structure. Every deal, memo, decision, and figure is a typed object that knows its relationship to every other object, so each new deal lands already cross-referenced against your history. “This looks like the business-services deal we passed on in 2023, and here's why” surfaces on its own. It beats off-the-shelf retrieval on accuracy and on cost, and both gaps come from the same place: the structure does the work the model would otherwise have to guess at.
All of it runs on Palantir Foundry, the same infrastructure Palantir deploys for Fortune 100 companies. We built Antonine through Palantir's founder program, one of 23 startups selected to build on the platform, working directly with their engineering team. That matters for the Harvey comparison more than it sounds. Harvey's wedge was private deployments that serious firms could actually trust. Ours is the same, on infrastructure that defense and intelligence agencies already trust.
You can copy a chat UI over a weekend. You can copy a workflow template over a weekend. You cannot copy three years of a firm's decisions encoded into a graph. That moat deepens on its own.
The tailwind nobody is pricing in
Anthropic put $200 million into a joint venture with Blackstone, Hellman & Friedman, and Permira to push Claude into their portfolio companies, with room for that vehicle to grow toward a billion. OpenAI struck a similar deal with TPG and Bain.
Read that twice. The two leading AI labs looked at the entire enterprise landscape and independently decided that PE firms are the distribution channel for enterprise AI. Not the hyperscalers. Not Salesforce. The funds. The bet is that whoever owns the portfolio pushes AI down into the operating companies, white glove, implementation first.
Jensen Huang now talks about token consumption per employee as an operating metric. PE firms are reportedly targeting eighty cents of AI tokens for every dollar of salary inside their portfolio companies. Work that used to take a consulting team six months gets done by agents in days, at a fraction of the cost.
None of this is theoretical. It's happening now, and the firms that build internal AI infrastructure first will post returns the rest of the market can't reverse engineer.
The portco multiplier
Here's the part most people miss about the economics.
Once a fund's internal infrastructure is live, it doesn't stop at the fund. The same systems roll out across the portfolio. One engagement at the fund becomes ten, twenty, forty at the operating companies, each configured around its own data, each feeding what it learns back into the fund's intelligence layer.
That's not a pricing trick. It's a distribution model, and it's the same dynamic Anthropic and OpenAI are betting billions on at the macro level. Antonine captures it at the firm level.
The moat, restated
Software people have argued about moats forever. AI settles a lot of the argument, mostly by killing the old answers. It's not IP anymore, because anything visible gets cloned in a quarter. It's not switching costs, because AI cut migration pain by an order of magnitude. It's not even scale, because the entire point of this technology is that ten people can now do the work of fifty.
The only moat left is a system that compounds your firm's knowledge faster than your competitors compound theirs.
That's what I'm building Antonine to be. Not a chat interface. Not a template. A layer under the firm that gets sharper every time a deal runs through it.
Where this leaves us
Investment management hasn't had its workflow reinvented in a generation. Everyone is aware of AI. Most firms have ChatGPT or Claude licenses floating around and an analyst or two using them off the books. Awareness is not infrastructure, and the distance between the two is where the entire opportunity lives.
Harvey closed that distance for legal. I'm closing it for investment management. We're bringing on a small group of founding users right now, with implementation run directly by me and the Palantir engineering team, and the platform tuned to how each team actually runs deals instead of forced into a generic mold. The model is proven, the vertical is open, and the next couple of years decide who owns the category.
Zach