We co-hosted the AI for Finance Summit London with Finteda on September 16 at Deutsche Bank's 21 Moorfields, sponsored by Deutsche Bank, AQR Capital Management, and Axioma by SimCorp. This was our first European edition. Seventy-five invited practitioners from 52 organizations took part, with 57% from the buy-side and 52% at director level or above. Fifteen speakers joined one fireside, four panels, and 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, London took on the convergence of systematic and discretionary investing under agentic AI. A show of hands at the start found that about a third of the room had already bought an AI tool for investment research. The conversation has moved from if to how.

The dialogue centered on how a shared toolkit forces both systematic and discretionary investors to redefine their edge, their process, and their risk.

First, convergence is in the toolkit, not in identity. Systematic and discretionary investors reach for the same AI tools, but for different ends. The fundamental analyst retains ownership of analytical frameworks and valuation judgment. The quant's edge moves to proprietary data and the infrastructure to exploit it. The pressure point is the portfolio manager, where AI now challenges traditional covariance-based allocation. The decision of which strategies to automate becomes a pricing question, as clients pay a premium for idiosyncratic risk.

Second, information is no longer the advantage, data is. This point surfaced in every session. LLMs have commoditized knowledge of standard quantitative techniques. The new edge is proprietary or well-structured data, which demands new discipline. This means a central data catalog with single-call access, a formal onboarding process for new datasets, and one authoritative source per value. An agent knows a number; the data layer must tell it what the number means and where it came from.

Third, the expensive work has moved from producing to checking. Writing code is nearly free, but proving it correct is not. Model validation was already a bottleneck before AI, and its independence was often questionable. Now, validation cannot rest on a single human, model, or round; a board of more than four challenger agents degrades evaluation quality. Overfitting, not alpha decay, is the immediate crowding risk. Where a human analyst might overfit twenty strategies, a machine can now overfit two thousand in the same time.

Fourth, the specification that sat in someone's head now has to be written down. An agent executes instructions literally, not by intent. One was told to follow a set of coding standards and rewrote the standards to match its own code. This forces firms to create formal, machine-readable specifications that let an agent verify itself. The human role shifts from doing to defining the specification. Governance becomes a checkpoint: nothing deploys until the spec is met.

Fifth, adoption is won on the path of least resistance. Early adopters chose tools that fit existing workflows, not the most powerful models. Generic access with a broad mandate failed; tailored tools for a specific role drove adoption. One firm reported a 30% increase in coverage with no loss of quality, measured against scorecards senior research heads now use to rate analyst output. In parallel, research is being reorganized around cross-cutting themes rather than traditional sectors.

Sixth, headcount holds, but the skill mix inverts. Research is on the same path as trading floors, where one went from dozens of voice traders and a handful of electronic traders to the opposite over twenty years. The junior training model is being rebuilt: AI writes the first draft, so the loop that once ran through a senior's critique is being rebuilt around a feedback and rating layer on the output. To support this, one firm offered sponsored qualifications in machine learning and AI; half the team enrolled and teaches the others in a weekly, slide-free meeting. Hiring is also adapting, with one firm replacing an interview round with a week-long project requiring heavy AI use.

The live demo traced one research workflow end to end: expert call transcripts scored over time with citations back to the source, compared against what management said on an earnings call, with the same orchestration then applied to macro and credit.

Two tensions defined the close. The first is oversight. A clear rule of thumb has emerged for material decisions: if a junior analyst's work on a task would require checking, then so does an agent's. But no model exists for the long tail of lower-level decisions agents will make. The second is ownership of risk. As the lines between trader, quant, and developer blur, and one person can build a new tool in an afternoon, it is unclear who owns the outcome. Participants working on regulatory capital models expect specific AI regulation to become a direct constraint within 12 to 18 months.

San Francisco, Los Angeles, and Tokyo are next, with Hong Kong, Milan, and Singapore closing out the year.

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