
LinqAlpha hosts the AI for Finance Summit Singapore, co-hosted with CFA Society Singapore and Databricks, and sponsored by AQR Capital Management and AWS.

We hosted our second AI for Finance Summit in Singapore on Monday, November 16, at the Databricks office, during Singapore FinTech Festival week. It was co-hosted with CFA Society Singapore and Databricks, and sponsored by AQR and AWS. The afternoon was invitation-only, every registration was screened, and more than 120 people were in the room. About seven in ten attendees work on the buy side. Twenty-one speakers led two fireside chats, three panels on the investment desk, sell-side research, and data and infrastructure, a quant research session on systematic macro, a session on the context layer, and a live demo with OpenAI, all under the Chatham House Rule, with no press and no recording. A year ago, the question in this city was what AI could do for investment research. This time it was what is actually running inside real investment processes today, and what it takes to let AI run on its own.

The answers kept returning to the same places. AI now runs throughout the workflow but rarely on its own, the hard work has moved to checking output and controlling cost, and judgement and accountability stay with people. Six findings stood out.
The first is that AI is everywhere, but it does not act alone. AI now touches most of the workflow, from research and earnings notes to client onboarding and reconciliations, but very little of it runs without a person checking. Testing showed an agent's behaviour drifting further the longer it ran alone, which is why AI-enabled tools are in production and autonomous agents are not. Work that can be undone is automated first. Anything that cannot be undone, like a market order or an investor report, still goes through a human.
Second, your process is the edge, and the model is a commodity. Everyone has access to the same general models, so everyone tends to get roughly the same answers. The teams with the most trusted output wrote down how their best people work and which sources to trust, then had AI follow those steps with a check at each stage. That written process is the firm's own IP.
Third, checking now takes more time than producing. On a results day AI produces the numbers and the argument fast, but it can make a convincing case for a wrong number, so validation now takes longer than drafting. Reviewers approving output all day get tired and stop reading closely. That is why some argued for more reviewers, not fewer.

Fourth, agents need context, not just data. Decades of structured data can be read by machines, but agents also need to know what it means, where it came from and when. Was the market open? Was the feed delayed? Did a batch fail? Without that context, a precise answer can be completely wrong. Agents also need a map of how a firm's data fits together, so that signals spread across datasets can be combined.
Fifth, cost and access have to be controlled centrally from day one. Left alone, people tend to pick the strongest model for everything, and the budget can be gone within months. Firms that kept spending flat routed each task to the right model and set rate limits before launch. The useful measure is cost per outcome, not per token, because the bill also covers data, compute and memory. Each agent should be treated like a new hire, with its own identity and access only to the data its task needs.
Sixth, accountability and judgement stay with people. Someone has to own every recommendation, and that someone is never the model. A model cannot be fired or held to account, and a person can. Clients still pay for access on the ground and for a senior analyst to test their views against. As AI takes over much of the spreadsheet work, juniors have to be trained on what that work used to teach: questioning assumptions, working out what matters and explaining the reasoning.

Our live demo with OpenAI showed an agent working through how an election outcome could affect data centre build-out, linking companies, filings and events, then building a scenario dashboard on request.
Of the open questions the afternoon left on the table, three are worth watching over the next 12 months. The first is whether agents can trace second-order effects. Summarising the market's view of a company is routine, but tracing how news flows through to customers, suppliers and competitors is much harder, and progress is early. The second is who will read everything AI writes. Hundreds of reports a day help no one if nobody reads them. The third is whether AI can be trusted to check AI. Human review cannot keep pace with AI output, and how to score an agent's forecasts against what actually happened is still open.
Learn more at linqalpha.com or contact support@linqalpha.com.