The industry does not really have the option of avoiding AI in financial crime operations anymore. The volume, speed, and complexity requirements simply move too fast in the opposite direction.

But the uncomfortable part is this: the more important AI becomes in AML and KYC operations, the more dangerous its weaknesses become too.

And most discussions about trustworthy AI in banking still focus too heavily on the models themselves:

• Model governance.
• Model explainability.
• Model risk.
• Model controls.

Important topics, obviously. But I increasingly think the more fundamental problem sits elsewhere.

Because when you watch how many AI systems are currently being deployed inside large enterprises, what you often see is not really intelligence. What you see is a very sophisticated attempt to compensate for fragmented data environments.

The AI is expected to retrieve information from disconnected systems, interpret conflicting records, understand entity relationships, reconstruct context, reason through ambiguity, identify relevance, and then generate a coherent conclusion on top of all of it.

Which is an astonishing amount to ask from any system.

This is particularly difficult in anti-financial crime operations, where the actual difficulty is rarely a lack of data. Usually it is the exact opposite. There is too much of it, spread across too many systems, carrying too little shared meaning.

You see this very clearly in Ultimate Beneficial Ownership analysis.

Sven Eisermann from Commerzbank recently made an observation I thought captured the issue extremely well. Mr. Eisermann suggested that the challenge with UBO identification is not simply about collecting ownership information. It is about reconstructing and validating ownership and control structures in a way that remains understandable, explainable, and defensible.

And once upstream documentation quality breaks down - incomplete reporting, fragmented ownership trails, inconsistent structures - the burden shifts directly onto the financial institution.

At that point, what should have been a verification process quietly turns into a reconstruction exercise.

Now place AI into that environment.

We expect the system to understand legal entities, ownership hierarchies, indirect control relationships, transactional context, and behavioral significance while simultaneously generating outputs that sound fluent, confident, and actionable.

And this is exactly where the industry keeps colliding with the “plausible sounding answer” problem.

Because modern AI is remarkably good at producing answers that feel right.

That is precisely what makes failures dangerous.

The issue is usually not that the output looks obviously broken. The issue is that somewhere between data retrieval, contextual interpretation, reasoning, and language generation, the system quietly drifts away from reality while maintaining the appearance of confidence.

In low-stakes environments this is manageable.

In anti-financial crime operations, it becomes operational risk:

• Regulatory risk.
• Reputational risk.
• And eventually, under the EU AI Act, governance risk as well.

Which is why I suspect the future winners in high-risk AI for financial crime use cases will not necessarily be the institutions with the most advanced models. They will be the institutions that reduce the amount of chaos those models are forced to navigate in the first place.

For years, data modernization in banking has largely been associated with cloud migration, centralization initiatives, storage strategies, and master data management programs.

But AI changes the definition of modernization.

Because AI systems do not merely need access to enterprise data. They need interpretable context:

• Who is connected to whom.
• What matters.
• Why it matters.
• How relationships evolve over time.
• Which signals carry actual risk significance.

Without that contextual layer, AI spends enormous amounts of effort repeatedly reconstructing meaning from fragmented inputs. Every query becomes its own miniature investigation.

And honestly, this reminds me of something unexpectedly relevant I once heard while visiting Kruger National Park. Rangers had observed certain animals gradually adapting to use the park roads rather than forcing movement through dense bushland. Not because the animals became smarter, but because structured pathways reduced friction and uncertainty.

That analogy has stayed with me because it feels increasingly relevant to enterprise AI. Right now, many banks are essentially asking AI to fight its way through the bushes.

What if instead we gave it the Autobahn?

The Enabler For Superior AI: A Contextual Intelligence Platform

What’s needed is an interpretation layer between fragmented enterprise data and the AI systems operating on top of it. Context that already exists before the model starts reasoning. Relationships already connected. Meaning already established.

Interestingly, this is not a hypothetical idea. Financial crime investigators have already been moving in this direction for years by adopting contextual intelligence platforms

One of the most important shifts contextual intelligence platforms like DataWalk introduced into investigations was not simply speed. It was the ability for investigators to operate directly inside contextual intelligence rather than manually reconstructing it from scratch every time.

That changed the economics of investigations entirely. Instead of spending weeks stitching together fragmented information before meaningful analysis could even begin, investigators could move directly into understanding hidden ownership structures, circular money movements, coordinated fraud activity, and network-level risk patterns.

Now imagine applying that same principle to AI systems. Instead of forcing AI to repeatedly infer context from fragmented enterprise environments, the context itself becomes infrastructure.

The implications are significant. AI hallucinations decrease because the model is no longer compensating for missing connective tissue. Auditability improves because reasoning paths become easier to trace and explain. Operational scalability improves because less compute power is wasted reconstructing meaning over and over again from disconnected sources.

And suddenly the tension between AMLA and the EU AI Act starts looking slightly different.

Because maybe the real challenge was never simply “how do we govern AI?”

Maybe the deeper question is: how do we create enterprise environments where trustworthy AI becomes realistically possible in the first place? Contextual intelligence may hold the key.