Opinionated Advice About AI

I had the privilege of joining a panel at the FinAI conference in Denver this week, and I want to share a few of the opinions I brought on stage with me. I'm calling them opinions deliberately, because too much of the AI conversation in banking right now is hedged, cautious, and waiting for someone else to go first. So in the spirit of our consortium's principle that data beats opinions but action beats talk, I'm going to be direct.

Stop waiting for perfect data. 

This is the most common stall I see. Banks treating data readiness like a prerequisite instead of something you build as you go. In honor of Christopher Nolan releasing the Odyssey later this year, I'll borrow from that epic. Penelope, hoping for Odysseus's return, tells her suitors she'll choose one of them once she finishes weaving a burial shroud, then secretly unravels her work each night. I'm not suggesting any bank is secretly sabotaging their data warehouse project, but there is a certain attractiveness to tinkering under the guise of responsibility. 

The reality is that plenty of valuable AI use cases don't require a unified view of your customers. AML alert review is mostly a research activity, you don't need a golden record to get started. Savvi AI will even let you work from Excel spreadsheets. And if you're building a policy bot, one clean version of your policies loaded into a knowledge base is sufficient. You might end up with confidence intervals that are wider than you'd like, but 75% confidence is a whole lot better than no insight at all. With our member banks, I've been pioneering what I call the "just in time" data layer: prioritize your AI workflows, build what you need, operate it, record the gains, and go again. 

Document your institutional knowledge before you need it. 

This one is less obvious, but I think it's the most overlooked infrastructure requirement for AI in banking. Every bank has a version of this problem: the banker who knows what good looks like for a credit memo, but nowhere is it recorded. Or it is recorded, but there are six versions on SharePoint and the team just knows which one to reference. One of our members told me they didn't want to turn on a policy agent because of exactly this: too many versions, no single source of truth, and word of mouth filling the gaps. 

It goes deeper than documentation, though. When a loan exception is approved, the decision gets recorded, but the rationale almost never does. Your agentic workforce is going to need to be trained on the what and the why, not just the how. Agents can figure out process. They cannot figure out judgment unless you give it to them. 

Rethink "we buy, we don't build."

If that's been your default posture, I'd push you to revisit it. The investment required to build has dropped significantly, and talent is no longer the excuse it used to be. One of our banks was paying $400,000 a year for a small dollar loan platform (the kind that decisions overdraft approvals in real time.) They built their own model with relative ease and eliminated that recurring expense. I'm not saying SaaS is dead. But banks would be surprised by what they can build today. For the love of all that is holy, do not pay someone for a policy bot. Spend the $30 a month for a Studio license and do it yourself. If banks take the same buy-everything approach into this next era, they're going to find themselves dramatically overpaying to outfit the bank.

Guard your data. 

This is the one I feel most strongly about, and it ties everything else together. Too many community banks are sitting back, waiting for their existing providers to bolt on AI features, and in the process allowing all the insights to accrue to their vendors instead of to the bank. Your data is your proprietary insight into your customers. If you're not already working to unify it under your own roof, get on it. The banks that will be in the strongest position three years from now are the ones investing in their own data architecture today. 

One last thing. 

AI has the power to substantively level the playing field for community banks. But that window will close, just like it has with every prior technology wave. And when it does, AI stops being an offensive play and becomes a defensive one, a mandatory investment just to stay in the game. You cannot bypass the maturity curve. The only way through it is to start. 

---

For more insights from Madeline, check out this episode of Bank Nerd Corner.