Data's radical new role in the era of agentic AI

The era of agentic AI is upon us. A 2026 survey of 628 global financial institutions found that more than half (52%) of respondents are already piloting agentic AI or at more advanced deployment stages.^[1] Another study focused specifically on the capital markets found that building customized agents ranked top of the list in terms of increased IT spending, with 84% of firms seeing AI agents as a new layer of enterprise capability.^[2]

The potential is transformative. But we can't simply insert agents into existing systems and processes. We need to rethink the fundamentals, starting with the way we design and deploy our data.

Rethinking data fundamentals

Data doesn't take action. But data drives action. And the quality of that data predicts the efficacy of those actions. For decades, the world has designed data to be ingested by software and used by humans. Now we need to design it to be used by agents.

An investment in agent-ready data is one of the most valuable an organization can make, especially in the capital markets, where data errors and shortcomings can trigger significant financial, reputational, regulatory risks. But what makes data agent-ready?

Agent-ready data is: contextual

Humans are gifted at understanding context, which enables us to fill in the blanks or resolve discrepancies when data is inconsistent or incomplete. But agents can't handle ambiguity, which is why an agent-first system needs data that is preloaded with its own context and supplemental information.

For example, imagine a dataset for the private markets, where free cash flow is defined as "dry powder" in one instance, "uninvested capital" in another, and "highly liquid assets" in a third. Humans can easily reconcile these terms, but agents can't reconcile differences in terminology unless the data carries instructions for doing so. The same is true of factors such as provenance and timestamps, all of which can help agents find what they need to perform an action effectively, but only when that additional information is embedded within that data.

At S&P Global, we support agent-first systems aligned to industry ontology with a rigorous taxonomy that minimizes ambiguity by standardizing entity types, identifiers and relationships across datasets.^[VK1]

Agent-ready data is: fluid

Siloed systems have always hindered efficiency in the capital markets; in the agentic era, the effect is amplified. Agents are at their most valuable when they can complete workflows that span multiple systems. For example, an agent might need to pull a position from one system, capture a trade in another, and check settlement status in a third.

At S&P Global, we describe this as the "multiple front doors" principle. In the early days, data was generally accessed through a single point, such as a web interface. Then we shifted to a more flexible model using feeds and APIs. In the age of agentic AI, we need to go even further to ensure that a wider range of data can flow and commingle seamlessly, traverse multiple systems and channels, and track back to source frictionlessly, all while retaining the rich, contextual information agents need to act intelligently and decisively.

This means designing data with complex read capabilities that can be exposed to any system through universal connectors such as MCP (Model Context Protocol). This requires rigorous data cataloging, a query interface that enables agents to request specific records at any time in any environment, taxonomy, metadata, and a universal taxonomy that enables cross-system workflows without human intervention.

Agent-ready data is auditable

Audit trails have always been critically important in the context of the financial markets. As we turn more of the workflow over to agents, we need to be able to trace any data point back to its original source and context. The regulatory, legal, systemic, and operational risks demand that agentic processes support the highest degree of transparency and accountability. The capital markets are a complex, interconnected system, and agents that operate in this environment need both quality metadata and strategically deployed human checkpoints to safeguard system integrity.

Quality metadata, including information such as freshness timestamps, confidence scores, and lineage, help agents know when to escalate and their human managers know how to resolve the issue.

Data that can defend itself

Ultimately, the agentic era has created the need for a different kind of data—data that can defend itself. While humans can piece together the context that informs a specific data point, agents can't. Going forward, no operation can function without data that embeds its own context as a matter of record. 

S&P Global is one of the world's leading providers of data for the capital markets, and agent-ready, defensible data is our priority. Whether that data flows through our software ecosystem of portfolio management, corporate actions management, settlement, client lifecycle management, and market intelligence solutions or through our clients' bespoke systems, it is designed to support agentic intelligence and action, minimize the need for humans in the loop, and ensure that when humans do need to step in, they can do so efficiently and confidently.

Stay tuned for the next post in this series, which examines the vital role of an organization’s system of record in the agentic AI era.

^{[1] https://www.jbs.cam.ac.uk/wp-content/uploads/2026/04/ccaf-2026-04-28-global-ai-in-financial-services-report.pdf}

^{[2] https://marketingassets.microsoft.com/gdc/gdc4Gxmxi/original}

^{[VK1]I think we shoudl refer to industry aligned ontology that we are building here - which gives a consistent meaning for the attributes}