The agentic AI shift: Why financial services is the highest-stakes testing ground
Agentic AI does not suggest. It acts. Unlike a chatbot that surfaces a recommendation or a copilot that drafts a response for human review, an agentic system executes multi-step tasks autonomously — placing trades, triggering compliance workflows, reallocating portfolio positions, flagging transactions for investigation. In a regulated industry where a single erroneous transaction can generate regulatory scrutiny, client losses, and reputational damage simultaneously, that distinction is not academic. The cost of a wrong answer scales with the autonomy of the system delivering it.
Financial services sits at the intersection of two forces that stress-test AI harder than almost any other sector. Markets generate consequential data by the second — price feeds, liquidity signals, volatility indicators — and that data expires almost as fast as it arrives. Simultaneously, firms operate under one of the world’s most demanding regulatory frameworks, spanning MiFID II, Basel III capital requirements, GDPR, and a growing body of AI-specific guidance from regulators including the FCA and SEC. An AI system that acts on stale data or produces an output it cannot explain to an auditor fails on both dimensions at once.
Yet the infrastructure most firms are working with was built for a different era. Legacy data architectures were designed to support human analysts and batch reporting cycles, not autonomous agents that need authoritative, real-time context to make consequential decisions at machine speed. Steve Mayzak, global managing director of Search AI at Elastic, puts it plainly: “It all starts with the data.”
The urgency to deploy is real. Competitive pressure, cost reduction targets, and client expectations are pushing firms toward agentic systems faster than their data foundations can support. That gap — between the ambition of the AI roadmap and the readiness of the data layer underneath it — is where deployments fail. Not because the models are inadequate, but because the data feeding them is incomplete, inconsistent, or inaccessible at the moment an agent needs to act.
What ‘data readiness’ actually means — and why most firms don’t have it
Data readiness has nothing to do with data volume. A firm can hold petabytes across dozens of systems and still be completely unprepared to run agentic AI. What readiness actually requires is an authoritative, unified context data store — a single layer that AI agents can query reliably, in real time, with results they can trust.
Most financial institutions don’t have this. They have the opposite: fragmented infrastructure built over decades, where trading systems, risk platforms, CRM tools, and compliance logs each operate as isolated silos. None of these systems were designed to serve as a coherent context layer for autonomous agents. They were built to serve specific functions, specific teams, and specific workflows — not to surface a consistent, reconciled picture of reality on demand.
This fragmentation creates a direct operational risk when agentic AI enters the picture. Agents don’t pause to question whether the data they’re acting on is current or complete. They execute. When context is drawn from misaligned sources, agents hallucinate facts, act on stale positions, or receive conflicting signals they have no mechanism to reconcile. In financial services, those failure modes aren’t abstract — they translate into erroneous trades, compliance breaches, and regulatory exposure.
Steve Mayzak, global managing director of Search AI at Elastic, puts it plainly: “It all starts with the data.” That framing matters because the industry conversation has fixated on model sophistication — which agent framework to use, which large language model to deploy. The actual bottleneck is earlier and more fundamental. An agent operating without an authoritative context store is making decisions in the dark, regardless of how capable the underlying model is.
Financial services firms face this problem at unusual severity. They operate under some of the most demanding regulatory requirements of any sector, while simultaneously responding to market data that updates by the second. That combination — high governance stakes, high data velocity — makes a unified, governed, real-time context layer not a nice-to-have but a hard prerequisite. Without it, agentic AI doesn’t underperform. It fails in ways that are difficult to detect and expensive to remediate.
The real-time problem: When stale data meets autonomous action
Financial markets generate pricing updates, order book changes, and risk signals in microseconds. An agentic AI system acting on data that is even a few seconds old is not operating in the present — it is executing decisions against a market that no longer exists. In equities trading, a position flagged as low-risk at 9:31 AM can become a liability by 9:31:03. Autonomous agents have no mechanism to recognize that gap unless the underlying data infrastructure closes it first.
The problem runs deeper than connection speed. Legacy data pipelines in financial institutions were built around batch processing cycles designed for human analysts who review reports every few hours, not for AI agents that query context thousands of times per minute. These pipelines were never engineered to serve sub-second retrieval at the scale agentic systems demand. Feeding a modern AI agent through that architecture is equivalent to routing a high-frequency trading system through a dial-up connection — the intelligence at the endpoint is irrelevant when the data arriving is already stale.
Fixing this requires more than upgrading to faster infrastructure. The data itself must be indexed in ways that support semantic and vector-based retrieval, not just keyword lookup. Before an agent is permitted to act, validation layers must confirm that the data it retrieved is current, complete, and authoritative. As Steve Mayzak, global managing director of Search AI at Elastic, puts it: “It all starts with the data.” Financial services firms operate in one of the most regulated sectors on earth while responding to external events that update by the second — that combination makes data latency not just a performance problem but a compliance and liability risk.
An autonomous compliance agent that triggers a suspicious activity report based on a customer’s account state from six hours ago can create regulatory exposure rather than reduce it. Real-time data access is not a feature of agentic AI in financial services. It is the baseline requirement without which the entire system becomes a liability.
Governance at scale: The compliance dimension that most AI vendors underplay
Financial services firms operate under regulatory frameworks that treat AI decisions as legal acts, not suggestions. MiFID II requires firms to document the basis for investment decisions, including the data inputs used. Basel III mandates demonstrable controls over risk model inputs and outputs. The EU AI Act, now in force, classifies credit scoring and insurance risk tools as high-risk AI systems subject to transparency and auditability requirements. Meeting these obligations means every action an AI agent takes must trace back to a specific, versioned, permissioned piece of data — not a general model state, but a precise record of what the system knew, when it knew it, and who authorized it to act on that knowledge.
Most enterprise data stacks cannot deliver this. They were built for human-speed workflows, where a compliance officer could manually reconstruct a decision trail after the fact. Agentic AI operates at machine speed, executing hundreds of decisions per minute across trading, credit assessment, fraud detection, and client servicing simultaneously. Bolting governance onto a system after deployment — through logging layers, post-hoc auditing tools, or manual review processes — creates gaps that regulators will not accept and that operational risk managers cannot defend.
The data layer itself must enforce access controls automatically. It must track lineage — recording which source fed which decision — and maintain version history so that a query run at 9:47 a.m. can be reconstructed exactly, using the data state that existed at that moment. Role-based permissions must propagate in real time so that an agent handling retail client data cannot access institutional counterparty records, even when both exist within the same underlying store.
Steve Mayzak, global managing director of Search AI at Elastic, frames the challenge directly: the success of agentic AI in financial services depends on the quality, security, and accessibility of the data it relies on — not the sophistication of the model sitting on top of it. Governance built into the context data store is not a compliance checkbox. It is the infrastructure condition that makes agentic AI legally deployable at all.
The authoritative context data store: What the solution architecture actually looks like
The emerging answer to financial services’ data fragmentation problem is a purpose-built authoritative context data store — a unified layer that ingests, normalises, and governs enterprise data from across every system and makes it queryable in real time by AI agents.
This architecture demands something that neither traditional relational databases nor modern data lakes were built to deliver: a simultaneous combination of vector search capabilities, structured data access, and strict governance controls operating within a single platform. Vector search enables AI models to retrieve semantically relevant context — finding information by meaning rather than exact keyword match. Structured data access ensures agents can query precise, transactional records. Governance controls enforce who and what can access which data, a non-negotiable requirement in financial services where regulatory obligations attach to every data interaction.
Elastic positions itself directly in this space. Steve Mayzak, the company’s global managing director of Search AI, is direct about the dependency: “It all starts with the data.” Elastic’s search and retrieval infrastructure is designed to serve as the data backbone for agentic AI deployments, bridging the gap between raw enterprise data distributed across siloed systems and the clean, contextualised, model-ready information that agents actually need to function reliably.
The practical implication is significant. An agentic AI system operating without this kind of store is effectively working blind — pulling context from wherever it can reach, with no guarantee that the data is current, accurate, or permissioned for that specific use. In financial services, where market conditions shift by the second and regulatory exposure is constant, that gap between retrieval and reality is where AI deployments fail. The authoritative context data store closes that gap by making a single, governed, continuously updated data layer the only source AI agents are permitted to query.
What this means for financial services leaders right now
Financial services leaders face a clear decision point: fix the data foundation now or accept that agentic AI deployments will fail publicly and expensively. Model capability is not the constraint. Data readiness is. Boards need to treat this as an enterprise risk issue, not an IT backlog item.
The competitive advantage in agentic AI will not go to the firm running the most sophisticated model. It will go to the firm that can deliver the most reliable, governed, and real-time data context to its agents at the moment a decision is being made. Financial services companies operate in an environment where external events update by the second and regulatory exposure is constant. A slower but data-accurate agent outperforms a capable model operating on stale or ungoverned information every time.
Before any agentic deployment, leaders should audit their data infrastructure against three specific criteria. First, accessibility: can agents retrieve the data they need in real time, or does retrieval depend on manual pulls, batch updates, or siloed systems that introduce latency? Second, reliability: is the data accurate and current at the point of use, not accurate as of last quarter’s reconciliation? Third, governance: is every data interaction auditable and compliant at scale, across every agent action, not just at the input stage?
As Steve Mayzak, global managing director of Search AI at Elastic, states directly: “It all starts with the data.” Firms that skip this audit and move straight to model deployment are building on an unstable foundation. When an agentic system triggers a flawed transaction, misprices a risk, or produces a compliance gap, the failure will trace back to data — not the model. That is the failure that lands in front of regulators and on the front page.
The firms that get agentic AI right in financial services will be those that invested in an authoritative context data store — one that is accessible, reliable, and governed — before they scaled agent capabilities. That investment is the actual competitive moat.
