The Intelligence Layer: AI Agents Still Depend On The Data Beneath Them

Alex Ford, CRO, Encompass. Alex drives global business growth, working with customers and industry partners to transform KYC with CDI.

getty

​Recently, Anthropic convened a briefing with JPMorgan CEO Jamie Dimon in the room. The event announced 10 pre-built AI agent templates aimed squarely at financial services. The day before, Anthropic announced a partnership with FIS, one of the payments infrastructure companies sitting at the center of how money moves for thousands of financial institutions worldwide. The message was unambiguous. Anthropic wants Claude to be the operating layer for regulated financial work.

It is a bold ambition, and it reflects where the industry is, with 61% of financial services firms actively using or assessing generative AI, with 89% reporting that AI has already increased revenue while lowering costs. The real migration now underway is from single-agent task automation to coordinated, multi-agent systems that can handle end-to-end workflows, with a human only in the approval loop.

But that shift introduces a fundamental tension: The governance models and human-in-the-loop controls that financial institutions rely on today were not designed for this scale. When AI agents execute complex, multistep workflows autonomously across data sources, counterparties and jurisdictions, manual review checkpoints become bottlenecks rather than safeguards. Effective oversight at speed requires governance built into the architecture itself: automated audit trails, explainable decision logic and trust established at the data layer before any agent acts on the data.

But none of it works without the data layer beneath it.

A Crowded And Increasingly Sophisticated Ecosystem

To understand what is happening in financial services AI, you need to see the full stack.

At the foundation, large model providers are staking out territory. Anthropic's 10 agent templates run on Claude Opus 4.7, which leads the Vals AI Finance Agent benchmark, though even Anthropic's own documentation notes every output requires qualified human review. Agents produce drafts. They do not sign off on decisions.

Above that sits a generation of purpose-built vertical platforms. Rogo, founded by former Lazard and J.P. Morgan bankers, has built a workflow orchestrator designed specifically for investment banking, now used by more than 25,000 bankers and investors. Hebbia’s Matrix product runs specialized sub-agents across massive document sets, backing every result with clickable source citations. Both sit on the Forbes AI 50 list.

Then there is the layer that often goes underreported: the banks themselves. Institutions at scale are not simply buying platforms. They are building internal AI programs combining proprietary and third-party models. The most sophisticated institutions are treating external platforms as components of a larger, internally governed architecture.

What The KYC Screener Gives You—And What It Doesn't

The KYC screener agent is the template that has generated the most interest in compliance circles. It sits natively inside analyst workflows via M365 add-ins, reasons across documents rather than simply ingesting them, and escalates edge cases rather than silently passing them. That is a different paradigm from the rules-engine generation of know your customer (KYC) tooling.

Practitioners who have tested it reach a consistent conclusion: useful, but not production ready. What is in the box is a structured starting point: document parsing, a configurable rules-engine skill, a screening step that calls out to sanctions, PEP, and adverse media providers, and a case file output. What is not in the box is what turns a shell into a system: the firm's own calibrated risk rules, live data subscriptions and integration with real screening providers, connection into existing case management systems, and the audit evidence a regulator will want to see. The hard part, production-grade integration, regulator-grade explainability, ongoing monitoring, still needs to be solved.

Guardrails And Stage Gates

Responsible deployment therefore requires controls that sit above the model layer. Domain policy guardrails define what an agent can and cannot do: An AML agent can flag, summarize and route, but it cannot make a filing decision; a KYC screener can identify discrepancies and generate a risk score, but it cannot approve onboarding. Stage gates enforce checkpoints at which a qualified human must review and sign off before the workflow can advance. The sophistication lies in the design of where those gates sit, what information the reviewer receives and what escalation logic governs edge cases.

Two further risks underline why these controls are load-bearing rather than optional. The same case has been observed scoring differently across separate runs, non-determinism that is a structural property of large language models, not a defect to be patched. This reinforces a foundational conclusion: The quality and provenance of the data an agent reasons over are not background considerations. It is the ground on which defensible outputs are built.

Agents operating on corporate identity data automatically assembled from privately held information, authoritative public registries, verified legal entity structures and continuously refreshed beneficial ownership data produce outputs that can be traced from source to conclusion. That is what regulators will require. When a regulator asks how a decision was made, the answer cannot begin at the model. It must begin at the data, and that data must be verified before the agent ever touches it.

The Data Layer Remains The Differentiator

Model providers are commoditizing fast. Benchmark scores are converging. Agent templates are proliferating. What is not commoditizing is the quality, verifiability and provenance of the data that agents reason over.

Perpetual KYC infrastructure, which replaces static periodic review cycles with intelligence-led, always-current customer profiles, does not simply make compliance faster. It makes AI-driven compliance possible.

Institutions that have invested in a verified, structured and continuously maintained corporate data foundation are better positioned to withstand regulatory scrutiny. They can onboard faster, identify beneficial ownership risk earlier, and produce AI outputs that can be defended to an auditor, a regulator or a board.

The AI is the engine. The data is the fuel. The quality of that fuel and the governance of the engine determine whether it runs or stalls midway through the journey.

The technology leaders who will look back on this period with satisfaction are the ones who asked the harder question first: Not "which agent platform should we choose?" but "is the data we're feeding our agents actually good enough to be trusted?"

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?