A co-authored paper by Silent Eight and RZOLUT

The Problem Isn't the Model. It's the Diet.

The fight against financial crime is increasingly powered by AI. Agentic AI has changed what's possible in investigation, turning work that once consumed thousands of analyst hours into decisions made in seconds, accompanied by documented reasoning. 

However, there's a less glamorous truth underneath the excitement, and it's one that compliance officers are discovering the hard way: an investigative agent is only as good as the data it's given to reason over.

You can build the most sophisticated reasoning engine in the world but if you feed it shallow, stale, or context-free data, it will reach conclusions with the same characteristics. The constraint on better outcomes is not the AI but rather the diet of data. Financial crime data that most institutions rely on today was designed for a different era, a different threat, and a different kind of consumer.

This paper is about that data: how the face of financial crime has changed, why the old data model can't keep up, what "fit-for-purpose" data actually looks like, and how data shaped for the Agentic age transforms what AI investigation can deliver – more false positives closed confidently at source, and more genuine risk found sooner.

The Face of Financial Crime Has Changed. The Data Hasn't Kept Pace.

Financial institutions in the US and Canada alone now spend $61B annually on financial crime compliance (Forrester, 2024) — a figure that has risen for 99% of institutions. Globally, AML spending is projected to reach $51.7B by 2028, with global spending on financial crime compliance (FCC) climbing to over $200B annually. 

The problem is getting bigger every year. Financial crime is faster, more networked, and more adaptive than it has ever been. Sanctions regimes shift in days, not years, and evasion typologies mutate to match them almost in real time. Beneficial ownership is deliberately obscured across jurisdictions. Criminal networks exploit the seams between data sources, jurisdictions, and languages.

Against that backdrop, the prevailing data model shows its age. For decades, compliance data was treated as a static reference asset: buy the broadest list you can, refresh it on a periodic cycle, and screen against it. That model has three structural weaknesses that matter enormously in the Agentic age.

Firstly, it optimises for breadth over depth. A name and a date of birth tell an agent – or an analyst – almost nothing about whether a match is real. Secondly, it treats currency as a scheduling problem rather than a risk problem; a list updated weekly is, by definition, a week behind a threat that moves daily. 

Finally, it is one-size-fits-all: the same generic dataset is shipped to a regional retail bank and a global correspondent institution, despite radically different risk profiles, geographies, and regulatory exposures. The result is the worst of both worlds – too much noise to investigate efficiently, and too little context to investigate well.

A Case in Point: Why Better PEP Data Is Crucial

Consider the challenge of politically exposed persons (PEPs). Many financial institutions screen against generic PEP datasets designed to maximise coverage rather than reflect their own regulatory obligations or risk appetite. The result is predictable: thousands – sometimes tens or even hundreds of thousands – of alerts generated by individuals who ultimately do not meet the institution's definition of a PEP.

For compliance teams, this becomes a familiar cycle. Analysts spend hours reviewing alerts only to document the same conclusion time and again: ‘does not meet our definition of a PEP’. The investigative process is sound. The problem is that it is being applied to alerts that should never have been generated in the first place.

The root cause is not investigator capability, nor the screening rules themselves. It is the underlying data. Generic datasets are designed to be comprehensive, not institution-specific. Without classifications tailored to a firm's regulatory requirements, geographic footprint, and risk appetite, screening systems generate unnecessary work before either a human investigator or an AI Agent has the opportunity to make a decision.

Fit-for-purpose PEP data changes that equation. When classifications reflect the institution's actual compliance obligations, irrelevant records are filtered out upstream, reducing false positives before they enter the investigative workflow. Human investigators spend less time closing predictable alerts, while AI Agents can focus their reasoning on cases where context and judgement genuinely matter.

This is the broader lesson for the age of Agentic AI. Better investigations do not begin with better reasoning alone – they begin with better evidence. The quality of every investigation is ultimately shaped by the quality, relevance, and context of the evidence available to it.

What ‘Fit-For-Purpose’ Data Actually Means

If the goal is to feed an intelligent investigative process, data has to be reconceived not as a list to be matched against, but as evidence to be reasoned over. 

In practice, that means four things:

• Accuracy: Precision at the record level is non-negotiable. Duplicates, misattributions, conflated identities, and unverified entries don't just create work – they actively mislead a reasoning engine. Data that is structured, de-duplicated, and validated at source removes false signals before they ever enter the investigation.

• Contextual: A match is meaningful only in context. Relationships, beneficial ownership, corporate hierarchies, geographic footprint, the nature and provenance of an adverse media report, the basis and recency of a PEP classification – this is the connective tissue that lets an investigator distinguish a true hit from a coincidence of names. Context is what turns a data point into a decision.

• Fit for purpose and client-aligned: The data an institution screens against should reflect the risk it actually carries – its markets, products, customer base, regulatory obligations. Tuning coverage, sources, and thresholds to the client rather than to a generic average is what makes the difference between a relevant alert and a wasted one. Data should be shaped to the question being asked.

• Timely: When sanctions and risk move in real time, periodic refresh cycles are a liability. Data needs to be updated on a timely, ideally near-real-time basis, so that the picture an agent reasons over is the picture as it is now – not as it was at the last batch update.

• Provenance: The source of the data determines the degree to which it can be trusted. A PEP gathered from a government website is a solid record; one that is classified on the basis of a news article is not so solid. The crucial lens to apply is – what quality will pass regulatory scrutiny?

This is the discipline RZOLUT has built its platform around: a single connected compliance data engine spanning sanctions, PEPs, global watchlists and enforcement actions, and entity relationships across hundreds of jurisdictions and scores of languages, structured and maintained so that it is fit to be consumed by AI Agents, not just read by people. 

Where Data Meets the Agentic Investigative Process

Here is where the two halves come together. Silent Eight's AI Agents conduct investigations the way an expert analyst would — gathering the relevant information on an alert, weighing names, contexts, geographies, and relationships against the real-world ambiguity that defeats rules-based systems, and reaching a decision accompanied by a written, auditable rationale. That process is powerful in its own right and becomes truly transformative when the data underneath it is accurate, contextual, fit-for-purpose, and current.

When an Agent can reason over rich, well-structured evidence rather than a thin name-and-date record, two things change at once.

The first is the impact of false positives. The overwhelming majority of alerts in any screening programme are not real risk — they are coincidences of name, transliteration, or partial match. 

With deep context available, an Agent can confidently establish that a match is a non-match and close it at the source, with a documented explanation that stands up to audit. Better data means more unnecessary alerts can be eliminated before investigation begins, while those that remain can be resolved confidently and automatically. This is where institutions begin to see meaningful reductions in investigative workload, faster customer onboarding, and lower compliance costs.

The second is the impact on genuine risk. The same context that lets an Agent dismiss a false match lets it recognise a real one. Imagine a sanctions alert involving a corporate customer. On the surface, nothing appears unusual. The entity has no obvious sanctions exposure and the initial match looks weak. However, richer data reveals a beneficial ownership link three layers deep to a sanctioned individual, alongside adverse media available only in a local-language source. 

Viewed in isolation, neither signal is decisive. Together, they provide the evidence an AI Agent needs to identify genuine risk where a rules-based approach might otherwise see only another false positive.

 Better data and better automation are thus not competing investments that trade off against each other. Deeper data makes the Agent more decisive; a more decisive agent extracts more value from the data. Each strengthens the other, creating a compounding effect across the investigation process.

A Shared Conviction — And a Partnership To Act On It

This is the thinking that has brought Silent Eight and RZOLUT together. We come at the problem from two ends of the same pipeline — RZOLUT shaping financial crime data to be accurate, contextual, fit-for-purpose, and current; Silent Eight building the Agentic AI that reasons over it to reach confident, explainable decisions — and we have arrived at the same conviction. The next leap in compliance performance will not come from data alone, nor from AI alone, but from data and AI designed in concert.

That shared conviction is what brought Silent Eight and RZOLUT together. Rather than attempting to solve the same problem, each organisation focuses on a different part of the investigative pipeline – RZOLUT on evidence-grade compliance data, Silent Eight on explainable Agentic AI reasoning. 

Together, we believe institutions should no longer have to choose between better data and better automation.

The Bottom Line

The face of financial crime will keep changing, and it will keep changing quickly. The institutions that stay ahead of it will be the ones that stop treating data as a static list to be matched and start treating it as living evidence to be reasoned over — accurate, contextual, aligned to their actual risk, and current to the day. 

Feed an intelligent investigative process data shaped that way, and the results follow: fewer false positives surviving past the source, more real risk caught and caught sooner, and compliance teams free to spend their judgment where it genuinely counts.

The model is ready. It's time the diet caught up.

Silent Eight and RZOLUT are partnering to bring evidence-grade compliance data and explainable Agentic AI investigation together. To learn more, get in touch at:

sales@silenteight.com