Building AI that holds: lessons from financial services practitioners at DAIS 2026

Zennify Team

At the Databricks Data + AI Summit, a senior vice president from a regional bank stopped the room. The AI hype cycle is producing more noise than signal, he said. Most vendors are selling the same silver bullet dressed up differently. The institutions actually making progress have stopped chasing the pitch and started building foundations.

Nobody pushed back. Because everyone in the room had lived it.

Zennify and Moody's hosted a private financial services lunch, midway through DAIS. The panelists were practitioners who've moved past the pilot stage. A CTO consolidating platforms at a mission-driven bank, a data intelligence director from a credit union who came up through internal audit, a senior VP who built his own agent orchestration layer to stay platform-independent, and the Moody's team redefining what decision-grade data looks like at scale.

Here's what they said.

The institutions moving fastest aren't chasing every tool

Mark Angler, Senior Vice President at TowneBank, was direct: every vendor arrives with a silver bullet, and none of them have one. The institutions moving fastest resist the pitch cycle, get clear on the outcome, and build deliberately toward it.

At TowneBank, that meant codifying engineering standards, naming conventions, and best practices into AI skills, then using those to build what the institution actually needs. The result is an internal agent orchestration layer that wraps foundational models and centralizes capabilities from disparate sources, without locking the bank into any single vendor's roadmap. They called it architectural independence, the ability to choose what to buy, what to build, and what to integrate on their own terms.

Adrian Glace, CTO of Amalgamated Bank, thinks in four layers. Experience, intelligence, data, and integration, with governance running across all of them. When teams clearly define and own each layer, platform decisions get easier because the principles are already in place. You're evaluating how well something serves the whole architecture, not just what it claims to do in isolation.

Both drew the same line on differentiation. Table-stakes platforms are table-stakes because every institution runs them. The custom work, the capabilities specific to your business model and risk profile, that's where real competitive advantage lives.

Most institutions are still carrying a data preparation burden that's already been solved

Jin Oh, Senior Director of Innovation and GenAI at Moody's, opened with a story about mailing physical discs to clients early in her career. Every week, a disc would go out with Moody's default probability time series. Clients would call when the mail ran late. The room laughed. The parallel to workflows many of them are still running today landed harder than the joke.

Moody's now covers 594 million entities with firmographic data, 210 million companies with financials, and 2 billion ownership links, sourcing over a million articles daily from 28,000 news sources. But the scale isn't the point. Clients working with that data through Databricks are saving 60-80% of the time they used to spend cleaning, auditing, and verifying before any analysis could begin. A well-built data foundation changes the job itself, not just the speed.

Jin's advice on where to start was simple. Begin with the decision you need to make, then work backwards to the data that decision requires. If your credit memo process takes days, that's the problem to solve. If your early warning signals only run quarterly, they're not early. Find where velocity is visibly broken and measure from there.

The people problem doesn't come with a product fix

Alicia Estrada, Data Intelligence Director at Randolph-Brooks Federal Credit Union, brought the conversation back to people, the dimension no platform solves on its own.

One of her engineers told her directly he was worried he was building the thing that would replace him. She reframed the role. Expert judgment should go toward work that actually needs it, away from pipeline troubleshooting and toward problems only that person can solve. And then the honest version: if you don't adapt, someone else will.

Champions showed up where she least expected them. A VP in compliance who'd never written a SQL query started building dashboards in Databricks and eventually created over 200. What shifted the culture was that VP explaining the value to his peers, not the data team evangelizing the tools. Peer credibility reaches people internal advocacy never does. And when those same users ran into conflicting numbers across dashboards, they became motivated to align on shared definitions. The governance conversation got easier because they'd lived the problem firsthand.

The advice that kept converging

The panel closed with one question: what's the single piece of advice you'd give to institutions earlier in their AI build-out?

Mark: Get your foundation in place first. Naming conventions, data dictionaries, reconciliation frameworks, quality checks. Speed without structure is just faster debt accumulation.
Adrian: Start with one real business problem. Solve it end-to-end. Make it credible, visible, and trusted. Every subsequent use case gets easier once you've proven the approach works once.
Alicia: Treat it as a change management program. Find the skeptics and get them on your side. A skeptic who becomes a believer carries more weight than anything the data team can say.
Jin: Start where velocity is broken. Credit memos that take days. Monitoring that only runs quarterly. KYC that isn't real-time. Find the bottleneck everyone already feels, apply the technology there, and let the outcome make the argument.

Four different answers. The same underlying logic: the foundation has to be real before the scale can be trusted.

What ran underneath all of it was a thread that started with a quote from a bank CEO shared at the open. The real prize of AI isn't headcount savings. It's giving people their time back to do work that actually matters. Jin described clients reclaiming 60-80% of the time previously spent cleaning data. Alicia described engineers freed from pipeline troubleshooting. Mark described a team that used to maintain ETL pipelines now building the capabilities that make the bank competitively distinct. The theme wasn't efficiency. It was reclaiming time for meaningful work for bankers, engineers, and the data leaders in between.

What comes next

The practitioners in that room weren't anti-vendor. They were anti-hype, and there's a meaningful difference. Zennify's financial services team has sat in the same seats, navigating the same decisions, the same skeptical stakeholders, the same pressure to show outcomes. If this conversation sounds familiar, we'd like to have it with your team.

Let's talk