We co-hosted the AI for Finance Summit Los Angeles with AWS and Databricks on October 12, at the AWS office in Santa Monica during a16z LA Tech Week, sponsored by AQR Capital Management. About 50 practitioners took seats chosen from about 480 requests, 89% buy-side and 70% director level or above. Nine speakers joined two firesides, three panels, and a live demo, all under Chatham House Rule. Where San Francisco asked three days earlier what must be true before an agent runs on its own, Los Angeles put credit on stage next to equity for the first time.

The through-line was what an investor can trust a system of agents to do, and where a person still has to make the call. Six findings emerged from the dialogue.

The first is that credit and equity reach for the same tools, but for different ends. Credit, where the payoff is mostly downside, uses them for early warning: a change in tone or in the numbers that flags deterioration while there is still time to act. Equity uses them to move faster and widen the search, with text as a new input and model forms far beyond a linear one.

Second, the bar for a model-made signal has gone up, not down. It must still make economic sense, hold up over a long history, and add something new. It now must also prove it is not an echo of what the model saw in training, and several teams watch it live, sometimes for years, before giving it weight. A signal nobody can explain does not raise assets.

Third, reading is solved. Judgment and relationships are not. An agent can get through a full credit agreement and find every term a borrower's owner wrote in. Getting those terms taken out is still a negotiation, and the answer is often no. Company communication is now written with these filters in mind, so knowing who is speaking, and why, matters more, not less.

Fourth, adoption follows demonstration, not debate. Earlier technology cycles on the investment desk turned the same way: senior investors came around when they saw the tool do the work, not when they were argued with. Small teams are now building their own reading tools in a matter of months and letting the output make the case.

Fifth, crowding is the risk nobody argued against. If every model reads the same filings and reaches the same view, positions crowd, and then something moves the other way. As inputs and models converge, the difference has to come from the firm's own history and the judgment of the people reading it.

Sixth, the constraint has moved from the model to the firm: cost, governance, and the pace of change. One user on one report can outspend a whole team in a matter of days. A yield can be calculated thirty ways in thirty systems, so an agent has to be told which one it is reading. Adoption is uneven even within a single team, and never as fast as promised.

The live demo with AWS traced one research workflow end to end: a firm's own notes and meetings organized into a second brain, agents that join the calls an analyst cannot attend, and the runtime, access controls, and guardrails needed to run them inside an institution.

Two tensions defined the close. The first is the record. Client-facing use is held back until every input and output of a model can be stored, retrieved, and shown to be unaltered, and the industry may need to ask regulators for a ruling. The second is where the next edge lives. With public-market text already heavily mined, several speakers pointed to private credit and private markets. Through all of it, the call on a position, a covenant, or a portfolio stayed with a named person.

Hong Kong, Milan, and Singapore are next.

Learn more at linqalpha.com or contact support@linqalpha.com.