Driving AI inside an investment firm

When people outside the firm ask me what it means to “drive AI” inside an investment manager, I usually answer with what it isn’t.

It isn’t picking the LLM. It isn’t the vendor demo. It isn’t the eight-figure annual contract for a copilot that the rest of the firm treats as a curiosity. Those things happen, but they’re the easy part — the part the trade press writes about because it’s the part that has logos.

The actual work is harder, slower, and far less visible.

The committee

Most large investment firms now have an AI committee or working group. Mine sits at the intersection of the technology org, the investment org, and risk and compliance. The committee meets, debates, and decides things like: which use cases get capital, where the firm’s data goes (and doesn’t), what the procurement bar is for vendor models, what the firm’s posture is on training data and PII, what we’re willing to do with public LLMs versus what stays inside the perimeter, what gets logged and reviewed.

The committee work matters because every one of those decisions becomes a constraint on the next year of building. A firm that decides “no public LLMs touching investment data” has a very different platform roadmap than one that decides “approved providers, with reviewed logging.” Both are defensible positions. Neither is reversible without a quarter of cleanup.

The job of the people on these committees — the job I think most people misunderstand — is not to advocate for AI. It’s to translate between the part of the firm that wants to ship and the part that wants to be careful. Both sides are right. The translation is the work.

The platform

After the committee gets the constraints right, you need a platform that makes it possible to actually use AI inside those constraints. For us, that’s a Snowflake-and-dbt foundation, an in-house Streamlit visualization framework, a model registry, evaluations that run on real investment workflows (not benchmark datasets), and the boring infrastructure that makes any of it auditable.

This is where most firms get stuck. The vendor demo runs on a dataset the vendor controls, against an evaluation the vendor designed, with a workflow the vendor invented. The day after the demo, you have to make it work on your data, with your governance, against your portfolio managers’ actual questions. The gap between demo-day and production-day is roughly twelve months, and most of those twelve months are infrastructure and team-building, not model work.

The gap

The gap between vendor pitch and investment reality is the part nobody talks about, so I’ll talk about it. A portfolio manager doesn’t want a chatbot that summarizes earnings calls — they want to know how a rate move will affect their book under three different macro scenarios, in the next ninety seconds, with the underlying data they can audit. The first thing is a feature. The second thing is a system. AI helps, but the system is mostly governance, data engineering, and the trust the user has built up with the people who built it for them. A model that hallucinates on a portfolio manager twice is a model that doesn’t get a third try.

So when I talk about “driving AI,” I’m really talking about three things: setting the constraints early, building the platform that makes the constraints livable, and earning the trust that lets the platform actually get used. The model choice is the smallest of the three.

It’s also the only part the press writes about. That’s fine. The work is in the other two.