In my last post, I talked about how my experience at Google, JPMorgan chase and Silver Lake has shaped how I think about AI and AI transformation.
Today, I’m diving into how to apply those learnings in today’s environment.
Too many firms still use AI to summarize documents or speed up an associate's spreadsheet work. That's the floor, and if your firm hasn't done that, you're already losing.
The interesting work begins when you give an agent a real workflow.
One of our venture debt clients moved from lead-to-LOI in two weeks down to five hours, and lifted conversion by 25%. The deal model itself becomes something new at that velocity.
A general partner can finally answer the question every LP wants asked: how many of my loan book loans are PIK-ing right now? Or which of my deals sponsored by X are underperforming? That used to be a four-week project for an analyst pulling data into a spreadsheet on weekends. With the right agents in place, it's a question you answer in minutes.
When tariffs became a live topic, LPs needed to know what portion of their portfolio was exposed. The firms set up with the right research agents answered in real time. The firms without them hired a consultant and waited a month.
A loan notice has 50 fields you can OCR. Extracting them is commodity work. The work that requires intelligence is the cross-referencing, tying the extracted data back to the credit agreement, the loan book, and the accounting system. Humans do that today. Agents can do all of it.
None of this is theoretical. It's happening right now inside the firms we work with. The common thread isn't the model. It's the fact that someone took the time to teach the agent how the firm makes decisions.
All of us are working with the worst agent we'll ever work with in our lifetime. Today's agents are the starting point. They'll get better, but the more interesting questions is how good you get at managing them?
I think the role of the analyst, the associate, the VP, and the partner is going to evolve from individual contributor to manager of agents. Managing a team of agents looks the same as managing a team of people. You learn each one's strengths, evaluate their work, review their output, and teach them.
This is the part the industry consistently underestimates. Foundation models keep getting better for everyone. Institutional knowledge layers don't, unless someone builds them. The firms that take the apprentice model seriously, the ones that spend real time teaching their agents how their firm underwrites, how it manages risk, how it makes decisions, end up with agents that perform 10x better than their competitors'. Same model. Different teacher.
That's what we're building at Obin. We're not focused on copilots. That doesn’t cut it in finance. At Obin, we’re building production AI workers that run end-to-end inside a firm's controls. They encode the firm's specific logic, produce traceable outputs, and resolve the edge cases that kill enterprise pilots before they reach production.
None of this works without data readiness. Agents are fun to talk about. They don't work without the boring infrastructure underneath.
Every large asset manager I talk to is proud of how much data they have. And every one of them knows that data is buried in archive deal rooms, in IC memos, in back office reports, in credit agreements, and in loan notices. Nobody re-reads those documents. Nobody mines them for institutional knowledge.
If you had 10,000 data engineers who could normalize every document and tie it back to a firm-level taxonomy, would that solve most of your data problem? Yes. Those engineers are the agents. That work is the precondition for everything else.
It's unglamorous, and it's where the real moat in enterprise AI is going to sit.
Listen to the full conversation with Marc Andrew on Modern Capital: The Private Markets Podcast here.
