
LinqAlpha hosts the AI for Finance Summit Milan at ACM ICAIF '26, sponsored by QRT, Standard Chartered, and AQR Capital Management.

We hosted the AI for Finance Summit Milan on Saturday, November 14, in the Main Hall of Bocconi University. The summit was the 3rd Workshop on LLMs and Generative AI for Finance at ACM ICAIF '26, the largest edition of the series so far and its first full-day programme. More than 300 people were in the room, from investment firms, banks, data and technology providers and universities across Europe, North America and Asia. About three in four registered attendees work in the industry rather than academia, from more than 100 firms. The programme included keynotes from Huyên Pham of École Polytechnique and Rama Cont of the University of Oxford, invited talks from Raffaella Calabrese of the University of Edinburgh and Mathieu Chalvidal of CFM, three panels on fundamental research, quant research and agents in production, 11 oral paper presentations and two poster sessions. The panels were held under the Chatham House Rule.

The answers kept returning to the same idea. AI is making answers cheap, so the edge is moving to asking the right question. A person still signs for the work, and the checks around it matter more than the model. Across the three panels, six findings stood out.
The first is that reading is now industrialised. Overnight summaries, draft theses and coverage of foreign-language markets are routine. Language is no longer a barrier to coverage. Little edge is left in reading filings faster, because the morning read is done before you arrive. The freed time goes into wider data, better system design, and more conversations with management, peers and colleagues.
Second, skills have to be kept on purpose. If AI does the first read and the first draft of code, juniors and seniors alike can lose the judgement to overrule it. Panelists keep some work manual and test fundamentals in interviews. As answers get cheap, the ability to ask a sharp question becomes more valuable.
Third, the stronger pattern is to use AI to build the system, not to give the final answer. Asking a language model which stock to buy was compared to using a hammer as a screwdriver. Instead, firms have AI write the optimiser, the screen, the data pipeline, or a quick replication of a new paper. This keeps the decision itself structured. Some investors also have agents argue for and against an idea before a person reads the case.

Fourth, what gets an agent into production is mostly organisational. Successful agents in production are narrow, owned by the business and tied to one problem. Pilots stall when too much is put into one agent, data sits in silos, or permissions are not harmonised. The business case decides. A recurring example was using agents to serve smaller clients profitably where the traditional model could not.
Fifth, a person still signs. Agents can search, subscribe to data and test ideas freely, even with alternative data like satellite imagery. The line sits at the trade. Placing client money, pushing code to production and sending a report to a client all stay with a human. Legal and reputational responsibility does not move to the model, so nothing reaches production or a client unread.
Sixth, the checks matter more than the model. A full audit trail, a log of every test, and a record of how users react to outputs came up in every panel. Evidence discussed during the day suggests a model's confidence can look the same whether its decision is right or wrong, so confidence alone is not a reliable check. In signal research, the search itself is part of the experiment. This means logging every variant, adjusting for the number of trials, and looking for bias if a result seems too good. Controls scale with the risk of each use case, and logs nobody reads are not controls.

Of the open questions the panels kept returning to, three are worth watching over the next 12 months. The first is how many use cases reach production, as pilots are everywhere but full production is still rare. The second is data architecture. Who benefits will depend less on which model a firm uses and more on whether its data is shared, permissioned and ready for agents. The third is whether rapid adoption is followed by a first wave of visible failures, a backlash and a supervisory response, a sequence many saw as likely.
Learn more at linqalpha.com or contact support@linqalpha.com.