Beyond the API: FactSet on Making Financial Data Ready for Agentic AI

The effectiveness of an AI agent ultimately depends on the quality, context and structure of the information it can access. For investment firms, that puts familiar challenges around data quality, provenance and integration firmly back on the agenda – while potentially changing what they expect from their data providers.

From data delivery to context

That shift has implications for both providers and consumers of financial and alternative data. Supplying a dataset is one thing; making it sufficiently contextualised for an AI system to understand and use within an investment process is another.

As firms bring together market, alternative and proprietary information, much of the complexity lies in normalising those different sources, resolving entities and establishing the semantic relationships between them. Without that groundwork, adding AI on top of the data stack may not live up to its potential benefits.

For Miller, this helps explain why some firms have yet to see the benefits from AI they initially anticipated. The problem may lie less with the models than with the data supply chain feeding them.

It also places a greater burden on data providers. Clients increasingly expect help with the infrastructure, ontology and semantic layers surrounding the data, rather than being left to build those capabilities themselves. That is particularly important where firms lack the engineering resources to manage an increasingly complex combination of external and internal information.

Agents as an extension of the investment process

Once those foundations are in place, Miller sees considerable potential for agents within research and portfolio management – particularly where existing human workflows can be clearly defined and replicated.

“Historically, there’s been a trade-off between breadth and depth within the research workflow,” he notes. “You can only cover so many names deeply because much of that work has to be done sequentially by a human. Agents provide scale to that process that was never previously available. I see the agent as a helper rather than a replacement.”

Earnings analysis provides one example. An analyst covering 50 stocks might spend half an hour before each earnings announcement reassembling information and insights they have already produced, followed by another half-hour afterwards. Across the portfolio, that represents around 50 hours of largely repetitive work.

An agent supplied with the appropriate data, tools and instructions could perform much of that work in parallel, leaving the analyst to concentrate on interpretation and judgement.

Miller likens the approach to creating a “digital twin” of an investment professional’s process. That requires firms to understand the workflow first: what an analyst does, which information they use and where human judgement needs to remain in the loop. Agentic AI then becomes a way of administering parts of that process at scale.

Back to the data foundations

This emphasis on workflow means that FactSet’s work around agentic AI is currently as much about infrastructure as agents themselves.

“A lot of the time I spend on agents isn’t actually on the agents yet,” he says. “It’s going back to the basics. How do we build high-quality data pipelines? How do we reimagine data structures for an AI and agentic world that is consuming this information? And how do we partner with clients to build knowledge services that include market-level intelligence while giving them a harness to build their own intelligence?”

The pace of development complicates those decisions. Firms investing heavily today risk accumulating technology debt as architectures and capabilities change. Waiting for the market to settle brings a different danger: falling behind while competitors experiment and learn.

“Everybody feels like they have to do this, but they’re finding that it’s really hard and really costly,” says Miller. “As a result, two things are happening. Either people are running so fast that they’re creating an infinite amount of tech debt, or they’re paralysed and aren’t doing anything because they feel woefully behind. There’s anxiety in both of those worlds.”

An evolving role for established data platforms

FactSet has approached that problem by drawing on infrastructure originally built to support its own data operations and has brought more of those capabilities into client environments.

The company processes around 26 million portfolios each night, according to Miller, giving it established stores of record across areas including research, trading and risk. Capabilities such as data mesh, concordance and entity resolution are becoming more relevant as clients seek to combine public and private market information alongside proprietary and alternative datasets.

FactSet is also exploring greater use of single-tenant environments and modular infrastructure, designed to give clients more control over their own data while allowing the underlying technology to evolve.

Openness is central to that approach. With new AI vendors appearing rapidly, investment firms face an expanding range of technology choices and the risk of building another generation of silos. Miller believes that is contributing to vendor fatigue and, in some cases, delaying decisions as firms wait to see what comes next.

For data providers, the emergence of agentic AI may therefore expand the job beyond supplying information. As investment workflows become increasingly machine-assisted, clients will need data they can trust, understand and combine with their own intelligence – and infrastructure flexible enough to accommodate whatever comes next.

Gerard Miller, SVP, Agentic Strategy at FactSet, will discuss agentic AI models and financial workflows at the A-Team Group/Eagle Alpha Alternative Data Conference in New York on 10th September.