At Money20/20 in Amsterdam, Elastic's Tim Brophy and fintech thought leader Dr. Efi Pylarinou delivered a clear message: the real challenge in AI is not the model, but the context behind it. The pair discussed how financial services companies need to focus less on hype and more on the data, workflows and governance that make agentic AI usable in practice.
From RAG to agentic AI
The vocabulary around AI is shifting fast. Concepts like retrieval augmented generation are being reframed as part of a broader move toward agentic AI and what Brophy calls "context engineering." That shift is at the core of AI Agents & Automation in regulated industries.
"We're maturing as an industry," Brophy said. "People have evolved to start talking about context engineering. This is essentially the same thing as RAG - however the consequences are higher than simply providing an informed chat response. It's about providing really precise, highly relevant context into an agentic process so that an LLM can make good decisions based on what's happening in any given workflow scenario."
For Elastic, this evolution reinforces the importance of unstructured data and real-time retrieval. Brophy added that anyone who is "just saying 'AI-enabled blank' is behind" in the current messaging.
Context over models
While much of the industry debate centres on foundational versus domain-specific models, Elastic's approach focuses on the layer beneath. "Even across all of these three model types, every single one of those still needs the information about the process that's running at any given time to give that model what it needs to make a good decision in terms of the outcome," Brophy said.
Pylarinou pointed to a widening gap between hyperscalers and the rest of the market. "What worries me is that hyperscalers not only have that rich data, but they have the resources - human, technical and everything - that they can go ahead and experiment with foundational transactional models," she said. "But what about the 99% of banks and financial services players?"
Elastic's answer is flexibility. Institutions do not need to commit to a single model strategy. "You can pick and choose," Brophy responded. "And if you have the resources to train your own model to be so domain-specific that it is aware of your own processes and your own data, that is ideal. But, of course, not everyone has that luxury."
Start small, scale impact
Despite the scale of AI ambition across financial services, Brophy advocated a pragmatic starting point. "Customers come to us and ask, 'Where do I start?'," he said. "What I like to say is: think really, really small. Don't try to boil the ocean with broad AI and have the approach of just deploying AI to everything."
Pylarinou agreed, pointing to focused use cases. "I like the example of taking anti-money laundering and saying that we'll tackle that, or onboarding a customer - KYC - and really go deep there," she said. A practical example is fraud detection, where Gen AI can augment existing models and human capability. Brophy noted that using AI to summarise alerts, find related alerts and derive an initial case analysis lets the human analyst avoid repetitive work.
Data architecture as the differentiator
As banks explore new AI-driven interfaces, Elastic argues that front-end innovation is secondary to back-end readiness. "If your data layer is fragmented in silos - with different query languages across all of these silos of data - that process is going to be cumbersome and painful," Brophy said. "It doesn't matter what the interface is."
Pylarinou added that for businesses, the real work lives behind the scenes. "We are going beyond the back office deep into the plumbing. We can't kick the can down the road on data architecture - it has to really be intelligent," she said. Elastic's platform combines APIs, dashboards and emerging agentic tooling, all built on the principle that speed-of-light retrieval and context generation depend on a unified data layer.
Compliance and explainability
With the EU AI Act set to introduce stringent requirements for high-risk use cases, explainability and auditability are becoming central to AI deployment. Pylarinou highlighted the stakes: penalties for non-compliance can reach 7% of annual revenues or €35 million, whichever is higher.
Brophy warned about the "black box" nature of models. "There's an element of this process that you cannot track and audit," he said. The answer is to log every interaction in the workflow. "What was the question? What were the prompts? What were the search results that powered the prompt? And then what was the answer?"
That logging, he said, ties back to RAG in the context engineering sense. "I really feel that we're in a very good position not only to power these agentic workloads, but also to provide the assurance of how these processes run and the explainability of their outcomes."
Why this matters for finance professionals
For banks and financial institutions, the message is clear: the next wave of AI won't be won by chasing the largest model, but by engineering the context that feeds it. Teams working on AI for Finance should start with narrow, high-impact processes like AML, KYC, or fraud alert triage, and ensure their data architecture supports real-time, auditable retrieval. The regulatory pressure from the EU AI Act makes explainability a business requirement, not a nice-to-have. Those who build the plumbing now will be best positioned when agentic AI moves from pilot to production.
Your membership also unlocks:
