We hosted our first AI for Finance Summit San Francisco on October 9 at the AWS office on Market Street, during a16z SF Tech Week. Close to 1,000 people requested a seat for about 60 in the room: 81% buy-side, half of them fundamental investors, 62% director level or above. Thirteen speakers took part across a fireside and three panels, alongside a live demo, all under the Chatham House Rule. Where Boston centered on governance, New York on alternative data, and Seoul on the research-to-production gap, San Francisco asked what has to be true inside an investment firm before it moves from a copilot a human checks at every step to an agent that runs many steps on its own. The labs building the models sit about a mile from the room. Adoption inside investment firms has not kept pace.

Six findings defined the path from here to the long-horizon agent.

The first is that the edge is depth, not speed. Long-horizon investors are not in a race to cover more names or trade faster. They use AI to go deeper on what they already own, and to let agents think for longer. The race is for conviction, not velocity.

Second, the bottleneck has moved from building to verifying. A small team can now produce twenty variations of a strategy in the time one used to take, and the work has shifted to proving that a number, a citation, or a backtest can be trusted. The constraint is no longer the ability to generate an answer, but the ability to believe it. Generation is cheap. Trust has a cost.

Third, a model must be treated as a measurement, not an estimator. A good measure is reproducible. If the answer changes when the prompt, market, or model changes, it is an unstable estimator with unknown bias. The defense is old discipline in a new form: state the hypothesis first. P-hacking is now available to everyone.

Fourth, AI works at the task level, with human judgment layered on top. End-to-end replacement of an analyst's workflow did not survive contact with a real process, but discrete tasks did. The most valued use case was not efficiency, but a devil's advocate to challenge confirmation bias. Productivity is not effectiveness.

Fifth, AI-ready data is an operational contract, not a marketing stamp. Buyers, platforms, and data providers are converging on the same requirements: every point must be traceable to its source, actuals must be separated from estimates, and agents need delivery they can join with internal data. Without evaluation, it is only a demo.

Sixth, context is the new moat. The models were described as good enough for now. The differentiator is the firm's own data, the years of internal meeting notes, and the business meaning only the firm can put in front of a model. Commoditized history is losing value. The next proprietary data point is gaining it.

A live demo with AWS ran a long-horizon agent on a single research theme, checking its answers for traceability along the way.

The afternoon closed on tensions nobody resolved. One view was that firms combining their own expertise with AI have a few years of alpha, and no one could say what comes after. While the constraint has moved to checking the output, the roles, tools, and budgets for verification are not yet settled. And per-seat pricing, built for a human logging in, is giving way to usage and outcomes that neither buyers nor vendors can yet predict.

San Francisco returns in 2027.

Los Angeles follows on Monday, October 12, with Tokyo on October 22. Hong Kong, Milan, and Singapore are next.

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