There's a question I ask every investment firm I talk to, and the answer tells me more than their entire deck of AI initiatives.
When an analyst at your firm uses AI, does it know your firm?
If the answer is no, you're at level one. If the answer is sometimes, inside one or two tools, you're at level two. If the answer is yes, every time, on every deal, with everything your team has ever decided behind it, you're at level three. Almost nobody is at level three. The firms that get there first are going to be very hard to catch.
This is the framework I use to think about that ladder, and what it actually costs to climb it.
01 The plugin
Most firms are here. An analyst has a ChatGPT or Claude subscription. They paste a few pages of a CIM into the window, get a summary back, drop it in the memo, close the tab.
I want to be fair to level one: this is real productivity, and I worked this way too. But nothing accumulates. The next analyst on the next deal opens a fresh tab and starts from zero. The model has never heard of your underwriting standards. It doesn't know which deals you passed on or why, which sponsors you have history with, which sectors you actually have a thesis in versus the ones you just watch.
Every analyst at every firm gets the same boost from the plugin. A productivity gain everyone has is an edge for no one.
And the real cost of staying here is sneaky. It's not that you're slow. It's that the firms above you are getting faster at being themselves while you're getting faster at being generic.
02 The workflow
Some firms have moved up a level. They built or bought tooling that wraps AI around a specific process. A screening interface. A diligence checklist with AI baked in. A memo template that auto-fills from source documents.
This is genuinely better. The process gets faster, the output gets more consistent, and new analysts ramp quicker because the firm's checklist lives inside the tool.
But level two has a ceiling, and it's the same ceiling level one had: the AI still doesn't know your firm. It knows the workflow. The deals that run through it leave no trace. Close the tool and the institutional memory closes with it.
I watch firms mistake level two for level three all the time. The team is using the tools, the demos look great, leadership concludes the AI program is working. They're right about the productivity and wrong about the compounding. Two firms running the same level-two tooling produce roughly the same output. The tooling can't be your differentiator if your competitor can buy it too.
The tell of a level-two program: the AI gets faster, but the firm never gets smarter.
03 The engine
Level three is what almost no firm has, and it's what the next decade's category leaders get built on.
At level three the AI stops being a tool and becomes infrastructure. Every deal that moves through the firm leaves a trace in a structured knowledge graph. Every memo, every IC decision, every pass and the reason for it, every sponsor relationship becomes a typed object connected to everything else.
The questions change completely. You're no longer asking the model to summarize a CIM. You're asking it: given the last forty deals we looked at in this sector, the ones we passed on and why, the ones we closed and how they performed, what should we be worried about in this one? And the answer comes back grounded in your firm's actual history instead of internet best practice.
The compounding is the whole point. A level-one tool is identical on day one and day three hundred. A level-three engine is sharper on day three hundred because three hundred days of decisions are encoded in the thing it queries. That's the difference between a team with institutional memory and a team without it, except the memory no longer depends on who happens to still work there.
Why most firms are stuck at one or two
Two reasons, in my experience.
First, level three is invisible from the outside. Level two and level three look identical in a demo: chat window, document upload, summary out. The difference is structural and shows up in months three through twelve, not week one. Most evaluations don't run that long, so most buyers can't tell the difference at the exact moment they're deciding.
Second, level three is hard to build. A typed knowledge graph is not a chat interface with a vector store glued to the back. You have to model the firm's actual work. What is a deal. What is a sponsor. What is an underwriting standard. How do they relate. That ontology work is unglamorous, and most teams underestimate it. The most common failure mode in this category right now is a level-two system wearing a level-three costume.
What changes when a firm gets there
Three things, in the order they show up.
The analyst's job changes first. Not in the AI-replaces-analysts way. In the way where the analyst stops spending sixty percent of the day on retrieval and reformatting and spends it on judgment instead, which is the job they were hired for in the first place. The team gets leverage without growing.
Then the underwriting gets better, and you can measure it. Because every passed deal is encoded with its reason, the firm starts catching repeats it used to miss. The same sponsor back with a new vehicle. The same flawed assumption dressed up in a new sector. The same mispricing the firm has seen and walked from twice before. Pattern recognition is what senior partners get paid for, and the engine puts it in front of every analyst from day one.
And the third one, the one that actually matters at the partnership level: the firm's edge becomes durable. Your criteria, your thesis, your decision history live in a system that sharpens every quarter. When a senior person leaves, the memory stays. When one joins, they onboard into a system that already knows what the firm believes.
Where Antonine fits
I built Antonine for level three. The Vault is the graph. The Assistant is how you talk to it. DealFilter and Diligence are the workflows that feed it, and the agents underneath handle everything that doesn't need a human in the loop.
You can run it at level two if that's where your firm is today, and you'll get the productivity wins. The difference is what happens underneath: every deal you process builds the level-three engine whether you're thinking about it or not. So the real decision isn't which tool to buy this quarter. It's whether the place you start lets you compound.
The firms that reach level three first compound on the ones that don't. That's the whole game.
Zach