We hosted the AI for Finance Summit Hong Kong on November 5 at Eaton Club Central, during Hong Kong FinTech Week. This was the first Hong Kong edition of the series. The afternoon was invitation-only, and more than 100 institutional investors and practitioners were in the room. About two in three attendees work on the buy side, from more than 70 firms. Seventeen speakers led a fireside chat, four panels on buy-side research, sell-side research, data and infrastructure, and wealth management, a data talk from LSEG, and a live demo with AWS, all under the Chatham House Rule. The question has moved on from hallucinations, or whether an agent can help with investment work at all, to how best to use agents and what it takes to trust one that runs for hours or days.

The answers kept returning to the same places: verification, a licensed person who signs, and the first-hand work that no model can yet reproduce. Six findings stood out.

The first is that the bottleneck moved from reading to verifying. Reading through AI summaries is already part of the process, so the human role is to verify systematically: deciding which questions to ask, and building checks that catch what the agent got wrong. Morning routines have shifted from ingesting text to framing the analysis and checking sources. The benchmark to beat is now a consensus that is itself shaped by AI. A mostly correct answer can still hide the one detail that changes the decision.

Second, the human parts of the job are worth more. Management meetings, site visits, and a one-on-one call with an experienced analyst were named again and again. AI saves time on historical numbers and repetitive tasks, but forward projections and judgment calls still need a person. Outcomes often hinge on intent and incentives, which are hard for a model to read. Models learn the middle of the distribution. Events that move markets happen in the tails.

Third, a licensed person signs. However much of the work an agent does, the final check on an investment decision, and the advice given to a client, belong to the person who holds the license. In wealth management this creates a new bottleneck, because every AI-generated result needs a person's review before it reaches a client. One person supported by digital workers may now answer for many more clients than before. Nobody on stage argued otherwise.

Fourth, controlled rollout beats giving AI to everyone. Broad rollouts were described as leaving most people not using the tools and some misusing them, so the favored approach is to mature workflows in a small team first. An agent that runs for days is approved only after a business case, a small proof of concept on a subset of data, checkpoints, and a way to stop a run that drifts off course. In regulated firms, concern about data leaving the firm sets the pace.

Fifth, agents that challenge each other give a sharper answer. Single generic agents were seen as prone to bias, holding on to a conclusion even when the inputs change. A workflow can split the work across specialist agents for analysis, coding, and checking, with deterministic models doing the quantitative work. The preference was to own the workflow, with internal tools for sensitive and client data and external tools for domain depth, and to use several tools on purpose to cross-check.

Sixth, how the next generation learns judgment is still unsolved. Judgment was described as coming from repetition: building model after model and seeing how markets behave through difficult periods. With that groundwork automated, how juniors build the same instinct came up in most sessions. Nobody claimed a full answer, but the direction was clear: judgment now has to be designed rather than left to chance, through deliberate mentorship, writing down past decisions including the mistakes, and hiring for curiosity and data skills.

Our live demo with AWS ran a multi-agent research workflow from source documents to a finished view.

Two tensions defined the close. The first is how research gets paid for. Client agents already read and cite research, and the open question is whether rights management and payment per insight emerge, so that research can reach the long tail of clients who never trade through the house. The second is who carries the risk when an agent acts. The client-facing firm carries primary liability, and shared responsibility with technology providers was discussed. That only works if a firm can prove whether an error came from a hallucination, inaccurate data, or an unclear instruction. That method does not exist yet.

Milan follows on November 14, with Singapore on November 16.

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