The financial services industry has historically rewarded firms that quickly act on the best information. But something fundamental is shifting. The differentiator is no longer only who has the best data, it’s also who has built the right AI architecture to put that data to work.
Given the competitive stakes are high and the technology landscape is rapidly evolving, in this article we offer perspective and a strategic framework for firms across the financial sector to consider.
In the past 12 months, we’ve seen the launch of Claude Opus 4.7, Gemini 3, GPT-5, and a wave of open-source alternatives. The frontier labs are moving faster than most enterprise technology roadmaps can track.
Meanwhile, the volume of financial data is compounding at a rate that would have seemed implausible a decade ago. At FactSet, the flow of data delivered through standard feeds has grown 30 times over the past 10 years. The number of entities covered has increased 97% in just the last four years, now reaching approximately 20 million entities and 90 million securities. That is 130% in the last four years in terms of growth.
The dilemma: more accessible data does not automatically create more intelligent decisions. It creates more complexity, unless the right architecture sits beneath it.
This is what Jevons Paradox tells us about the current moment. Historically, when a resource becomes more efficiently usable, total demand increases rather than declines. That happened with coal and steam power, with electricity, with fuel-efficient engines, and most recently with cloud computing.
We are now at the inflection point for financial data. As it becomes more accessible through new channels—MCP servers, agentic workflows, embedded APIs—demand will compound rather than plateau. Firms that have built the right data and technology foundations will be positioned on top of the wave.
Also 12 months ago, the dominant AI use case in financial services was retrieval-augmented generation (RAG): sophisticated search over structured data, delivered through a conversational interface. That was genuinely useful. It was also relatively limited in its ability to drive fundamental workflow change.
Today, task-level agents are live and in production across a growing number of firms. Multi-agent systems—specialized agents that are collaborating to complete complex, multi-step processes—are moving from demonstration to deployment.
The next stage is always-on personalized agents that understand individual user preferences, analytical habits, and portfolio contexts to act proactively on behalf of users. It is a near-term reality that multiple organizations are actively building toward.
When that stage arrives, the best-positioned firms will have built clean data foundations, established agent governance frameworks, and began the harder work of workflow redesign.
Before evaluating any specific AI solution, financial institutions need to be honest about their foundational readiness. Three questions cut to the heart of this:
1. Is your data truly AI-ready?
There tends to be a gap between an organization’s legacy data and the need for upgraded data that is accessible, enriched, and structured for AI workloads. Most large asset managers, asset owners, and wealth managers are housing fragmented internal data: documents and datasets that are not consistently classified, tagged, or linked through a coherent ontology. Without that foundation, even the most sophisticated AI model will produce unreliable outputs.
Preparing data for AI consumption requires cleaning, metadata attachment, semantic tagging, and entity resolution at scale. It is an ongoing practice to treat as a strategic, proprietary knowledge asset unique to your firm.
2. Do you have a controlled environment for agents?
Agentic AI is a step change in workflow capabilities. Unlike a chatbot or a retrieval-augmented search tool that a user initiates, agents act autonomously, run continuously, access multiple systems, and trigger downstream processes. It’s both powerful and, without the right governance framework, a real operational risk.
The concept of agent sprawl is emerging in early enterprise deployments. Agents that run unchecked burn tokens, create audit-trail gaps, and have the potential to produce outputs that conflict with each other or with firm policies. The solution is to build a controlled and transparent agent-management environment from the outset: agents are registered, their access is defined, their triggering logic is explicit, and their activity is visible to the enterprise.
The triggering framework matters as much as the agent itself. The most efficient agents are those that know precisely when to act. Financial services firms already have rich experience with event-driven signaling, such as market data changes, corporate actions, and earnings releases. That same logic, applied to agent activation, dramatically improves both efficiency and cost management.
3. Are you willing to reimagine workflows for AI?
Firms’ current workflows were designed on technology that was modern at the time. It was assumed that analytical tasks required significant human time because there was no alternative.
As a result, today there is a natural tendency to look at AI tools and ask how to apply them to existing workflows.
Earnings analysis is a good example. On any given earnings day, a buy-side analyst might spend hours digesting transcripts, updating financial models, reviewing sell-side reports, and reconciling those views with internal positioning. That workflow was rational given the tools available five years ago.
Today, a well-architected multi-agent system can handle the bulk of that information processing automatically in real time to reveal signals that matter, flag anomalies, and update relevant models. The analyst's time is freed for the strategic, judgement-intensive work that requires human expertise.
A practical architecture will enable firms to advance their AI capabilities (and competitive differentiators) at scale. We have built FactSet Intelligence, a layered approach that addresses each of these challenges in sequence. It combines a data layer, an agent infrastructure, and intelligent workflows.
The foundation is the data layer, where data is enriched at scale to ensure you receive the right answers and audibility across your solutions. This involves both curating and structuring extensive FactSet datasets across 150+ sources that are delivered through our proprietary model context protocol (MCP) server and offering document intelligence services that bring our enrichment capabilities to clients' own internal data.
AI tools without a verified, structured data connection can produce analyses and visualizations that look credible but are drawing on inferred, unsourced, or opaque data. The output looks right, but the provenance is unclear. In our industry where investment decisions, regulatory compliance, and client reporting depend on data integrity, the ability to trace every output back to a verified source is a design requirement.
Enrichment is quite involved. When a document is ingested, it is broken into chunks, which are the discrete pieces of content an AI model will use. Those chunks are then enriched. For example:
Named entities are identified and resolved so that variations of the same entity name are understood as identical and linked to a consistent identifier.
Financial concepts are tagged, enabling complex conceptual queries rather than simple keyword search.
Relationships between entities and concepts are mapped into a graph structure, so that when a complex question is asked the connections between suppliers, companies, countries, and events are understood and can be brought into play.
This is the same capability that has powered search and analysis within the FactSet Workstation for years. It is now available as a service for your firm's own documents—research notes, internal reports, client correspondence, proprietary data—bringing that same enrichment depth to your private knowledge base.
Model context protocol is the mechanism that makes data and capabilities plug-and-play across environments. It standardizes the way data is distributed to models, agents, and tools. Whether you are working within the FactSet Workstation, building your own solutions, or integrating with a frontier lab tool, MCP ensures that FactSet data flows in with full precision and traceability and that you always know exactly where your answers are coming from.
The middle layer is the agent infrastructure, which provides a controlled environment for building, registering and managing agents—both FactSet-built and client-built for specific workflows—and third-party agents from partners in the ecosystem. Because the agent infrastructure is built on top of the FactSet Workstation, it integrates naturally into existing user workflows rather than requiring wholesale adoption of a new environment.
What users see as a single agent interaction is, in practice, a more complex orchestration. Behind the interface, multiple sub-agents run in parallel with each handling a distinct part of the workflow to compose their outputs into a coherent result. This is what enables genuinely complex, multi-step processes to be completed at speed. Building and managing that orchestration requires a robust agent infrastructure: one where tools, skills, and sub-agent relationships can be configured, monitored, and updated as workflows evolve.
Underpinning our agent infrastructure is an event-driven signaling framework (Intelligence Toolkit) that uses our established market data infrastructure to trigger agents only when conditions warrant, rather than allowing them to run continuously and inefficiently.
This triggering infrastructure is not being built from scratch. FactSet already processes, ingests, stores, and monitors changes across data sources as part of core data operations. The same infrastructure that drives alerts and signals in the Workstation today is what will drive autonomous agent activation tomorrow.
Understanding what an agent actually needs to function is important. At the most basic level, an agent requires skills and tools.
Skills are the instruction set for completing a workflow; skills can be highly customized to a specific firm or standardized across user types.
Tools are the deterministic, repeatable analytical capabilities the agent draws on to execute those workflows.
This distinction matters in practice. Just because a large language model can perform a calculation or generate an analysis does not mean it is the most reliable way to do so. For financial workflows where consistency and auditability are non-negotiable, deterministic tools—engines that consistently produce the same output given the same input—are more appropriate than LLM-generated outputs for many core tasks. Building on infrastructure that exposes those capabilities correctly is what separates a high-performing agent from less capable agents.
The top layer is intelligent workflows, where real value is delivered. It’s where we are investing in forward-deployed teams of engineers and consultants who work directly with client organizations to redesign workflows for the agentic world.
Our Earnings Intelligence agent, demonstrated at a recent client conference, is an example of what this looks like in practice: a multi-agent, orchestrated system that analyzes portfolio exposure, processes earnings transcripts, surfaces relevant signals, and delivers insights in real time.
To best support our clients in such a quickly evolving technology landscape, we are building our agentic approach in a componentized, flexible manor to support dynamic delivery across delivery channels. Whether our client base wants to build on top of our intelligence stack, utilize it in evolving AI ecosystems, or interact in the most integrated way within the AI-native workstation, the experience will be consistent and seamless across all three.
In practice, users will move fluidly between all three channels, sometimes within the same working session. The infrastructure underpinning all three channels needs to be consistent so that data, context, and auditability travel seamlessly regardless of where the work is happening. Building in a way that serves all three channels simultaneously is a core design principle of FactSet Intelligence.
One of the questions we hear from clients is whether they should commit to a specific AI model or frontier lab, or whether they should wait for the landscape to stabilize before making investments.
Our view is straightforward: Waiting is not a viable strategy in the current environment, but committing to a single model is also unnecessary.
The right answer is to build on an architecture that is model agnostic, where data enrichment, agent management, and workflow design are separated from the specific model or frontier lab being used at any given moment. As models evolve, the underlying infrastructure retains its value.
FactSet Intelligence is being built with this portability in mind, supporting connectivity to Anthropic, OpenAI, Google Gemini, and others so that clients can make model choices based on capability and cost without rebuilding their data and agent foundations each time.
For decision makers assessing where to focus enterprise efforts, we recommend the following framework:
Audit your data estate thoroughly. Identify where your most valuable internal data lives, whether it is structured, tagged, and accessible for AI workloads, and where the gaps are. This is often more urgent than evaluating any specific AI tool.
Require auditability from the outset. Any AI solution deployed in a financial services context should be able to trace every output back to a verified source. If it cannot, the output is not production-ready regardless of how credible it appears.
Establish agent governance before you need it. The temptation is to run multiple agent pilots and worry about governance later. That approach enables sprawl and opacity problems that are harder to unwind than to prevent.
Choose infrastructure over individual tools. The specific models and applications available today will look different in 18 months. The data foundations, enrichment pipelines, and agent-management frameworks you build now will compound in value regardless of which models are leading at any given point.
Prioritize workflow redesign alongside technology deployment. The organizations capturing the most value from AI are pairing technology investment with genuine redesign of how work is organized and who is responsible for what.
Partner deliberately. The startup ecosystem in AI is producing remarkable capabilities at pace, but evaluating every new vendor would be a significant cost. The more sustainable approach is to work with a trusted infrastructure partner that has already integrated the most relevant capabilities. Then build on a platform that allows the integrations to evolve without disruption to your core workflows.
The combination of exponential data growth, increasingly capable AI models, and agentic infrastructure is creating a permanent shift in how investment research, portfolio management, and client reporting will be delivered.
Organizations that approach this moment with intention and clarity about their data foundations, governance requirements, workflow priorities, and partnership strategy will find both today’s and tomorrow’s technology genuinely transformative rather than merely disruptive.
Visit factset.com/ai and download our eBook: The Future of Connected AI—Demystifying Model Context Protocol for Capital Markets.
This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.