Agentic AI Won’t Fix Your Data Problem. It Will Scale It.

Financial institutions are investing in AI on the expectation that it will help them do more with the people and systems they already have. In Financial Crime, agentic AI is the obvious next step: rather than automating a single task, agents can coordinate a whole investigation or customer review, pulling evidence from different systems, linking it to the customer and preparing the next steps for an analyst.

Every one of those steps depends on data, and data is exactly where most banks are struggling. In conjunction with 1LoD, we recently published the 1LoD 2026 Financial Crime Benchmarking Survey, which found that 92% of banks cite data constraints as a barrier to AI adoption and 85% point to legacy integration, while only 23% say a lack of AI expertise is holding them back. Banks broadly believe the technology works and that the skills can be bought in, but a joined-up view of their own data is not something they can buy.

There is no shortage of appetite. 85% of banks want agentic AI, 82% want network or link analysis and 69% want entity resolution and a single client view, although the last two are really data capabilities in their own right. Before firms can get real value from agents, they need data that agents can actually work with.

Fragmentation, not quality, is the real issue

Ask most firms about their data problems and the conversation soon turns to accuracy, completeness and out-of-date customer records. Those things matter, but the survey points somewhere else: asked to name the single data issue that most holds back AI adoption, well over half of banks pointed to some form of fragmentation or linkage across systems, and only 23% named internal data quality.

That makes particular sense in Financial Crime, where the work depends on connecting transaction records, KYC documents, adverse media, SAR narratives and investigator notes that are often held in different places. Most firms have accurate data, in other words; what they lack is a reliable way to connect it.

For large banks, the problem has usually built up over decades. As one AML and sanctions lead in the US put it:

“A lot of the brick-and-mortar bigger banks have been built through acquisitions and mergers, so they have data everywhere. And not all that data is talking.”

External data can help in specific areas such as PEP and adverse media screening, but it can’t fix records that don’t link up inside the bank.

Why agents raise the stakes

Analysts have been working around data gaps for years, often without realising it, whether by noticing that two records belong to the same company or that an ownership chart looks out of date. Agents take a lot of that human judgement out of the loop.

An agentic workflow pulls evidence from several sources, links it to the customer, works out what’s missing and prepares the next steps for review, often before anyone has looked at the underlying records. As the report points out, a model doesn’t recognise gaps in its data; it simply produces a confident answer from whatever partial picture it can see. A 1st line onboarding and KYC lead put it more bluntly:

“You might buy the fanciest car, but crap fuel in it is only going to take you that far, which is nowhere. That for me is data.”

Periodic KYC refresh shows how this plays out in practice. It’s one of the strongest use cases for agents, since many banks carry large backlogs of low-risk customers and an agent can build most of the file while flagging only material changes for review. Picture an agent pulling transaction history from one platform, ownership data from another and screening results from a third, then preparing a risk recommendation, when the ownership data in one system was never reconciled with a more recent update in another.

A poorly designed agent won’t notice the gap, and its recommendation will look just as convincing as if nothing were wrong. A better one will spot the mismatch and send it to an analyst rather than burying it in a decision log. The agent doesn’t need much independence to be useful here; what matters is whether it can tell when the data doesn’t add up.

That is also what regulators want to see. 54% of banks cite regulatory caution as a barrier, which in practice means concerns about defensibility, explainability and human oversight, and defending an AI-supported decision starts with being able to show where the data behind it came from.

Policy is data too

There is a second kind of data that agents depend on, and it gets far less attention: the rules they are meant to apply.

In many firms, KYC and CDD policy still lives in Word documents and PDFs that are read differently depending on who picks them up. The survey describes CDD and onboarding as rules-heavy and document-intensive, and notes that traditional KYC software simply enforced procedures the bank had written itself, which meant it could be tested against the policy manual. Agentic AI changes that, because when the critical output is an agent’s judgement, the rules behind that judgement need to be explicit.

An agent can’t reliably follow a policy it has to interpret. Encoding policy as structured, machine-readable logic gives it rules it can apply consistently, and gives the firm something it can test and explain to a regulator.

Fixing the data without waiting for perfection

None of this means firms have to fix every data issue before they start using AI. For large institutions, with data built up across platforms, acquisitions and generations of technology, that day would never come, and launching a big data-cleansing programme and waiting for it to finish isn’t the answer either.

Fragmented data is usually a symptom of fragmented processes, and those need fixing too, because cleaning the data without changing the process that produced it tends to mean the same issues come back under a different name. A 1st line CDD lead at a UK-headquartered global bank, whose AI results are among the strongest in the survey, put it this way:

“Before you do anything, you have to re-evaluate the process, start to end, and pull it apart. If all you do is put AI or agents over the top of a broken system, it will reinforce the brokenness. It won’t sort out the inefficiencies. It will just automate nonsense or speed up nonsense.”

AI can help with this part as well. The report describes banks using narrow AI, which copes well with messy inputs, to build the linkage and structure that a joined-up data layer needs, so that more ambitious use cases can sit on top of it later.

What AI can’t do is settle the organisational questions, such as agreeing common identifiers, assigning data ownership, defining the golden record and deciding whose version wins when sources disagree. Those are decisions the business has to make for itself.

The banks that have made progress didn’t wait for perfect data. They started where data is already contained and repeatable, in areas such as document processing, sanctions alert triage and screening, and built out from there.

Five data questions to ask before deploying an agent

1. Which sources is it drawing from, and has anyone decided which one wins when they disagree?
2. Can it tell which records belong to the same customer, across every system it touches?
3. Can every piece of information it uses be traced back to where it came from?
4. Is the policy it applies written as explicit rules, or as a document it has to interpret?
5. If a regulator asked how a specific output was reached, could you show the data and the rule behind it?

Starting with the data, not the platform

The firms making real progress aren’t the ones with the most advanced models, but the ones that decided early which sources they trust, who owns a customer record and how records link up across systems. That’s the work we do with clients before any agent is deployed: finding out where customer data actually lives, which sources should be treated as authoritative, and where a contained use case can prove the model before it’s scaled.

KYX is our agentic KYC capability, designed to be built with clients inside their own environments, leveraging their existing data, systems and policies. It ingests and connects data and documents from across the organisation, creating a unified view of customers, entities and ownership structures. Policies are translated into machine-readable rules, ensuring consistent decision-making rather than relying on individual interpretation. Where data sources conflict, agents surface the discrepancy for human review rather than making assumptions, with every action fully traceable and auditable. Because KYX integrates with existing CLM, screening and onboarding ecosystems, it enhances current capabilities with AI-driven automation rather than requiring wholesale replacement of platforms that already work.

Agentic AI will only ever be as effective as the data, controls and policies that underpin it. The real opportunity today is not simply deploying agents but preparing the foundations that allow them to operate safely, consistently and at scale.

Want the full findings? Download The 2026 Financial Crime Benchmarking Survey & Report.