What We Took Away From the FSI Forum in New York

Earlier this week, we brought a room of banking and FSI Finance leaders together at the Kimpton Hotel Eventi in New York for our FSI Forum. Half a day, five sessions, and a lot of sharp questions from people who run Finance inside real banks. One idea kept coming back no matter who was speaking: in banking, AI is only worth as much as the financial core underneath it. If you couldn’t be there, here’s what you missed.

Our opening keynote, “Forward Finance,” didn’t start with AI. It started with the job. Banks are being asked to do more and do it faster, with examiners and boards wanting more evidence more often and rate pressure leaving little room to absorb a miss. Then came the question everyone had been circling: where does AI actually fit?

Most AI efforts in banking stall before they pay off, and usually not because the model is weak. They stall because the data is scattered across systems that never fully reconcile, and the business-critical logic lives in spreadsheets no one controls or can trace. A number you can’t explain to a regulator doesn’t help you, however clever the model behind it. So the case was plain: get the core right first. One platform, one governed source of truth, with close, planning, and reporting running off the same numbers. Do that, and the AI you put on top is something you can stand behind.

Deloitte took the long view. Finance has always run on whatever calculating tool was newest, from the abacus to the spreadsheet to the ERP, and AI is the next step on that line. What’s changed is the pace. The stat that stuck with me: 44% of the fundamental skills a Finance worker needs are expected to change by 2030. The Finance team of the near future looks less like scorekeepers and more like advisors to the business, with the routine execution handled by agents.

They were careful about one point, and right to be. AI reasoning is probabilistic and Finance is deterministic, and the two don’t safely mix on their own. You need a clear boundary where agents reach financial data only through governed connections, never by pulling raw tables. That boundary is what keeps everything above it defensible.

Aun Merchant and Kyle Webb from CrossCountry Consulting got specific, and this was where the trusted-core idea turned into something you could watch happen. Their argument: forecasting a bank’s balance sheet at the product-category level is where accuracy goes to die. A 30-year fixed and an ARM don’t behave the same way, so lumping them together widens your error band before you’ve started.

Then they showed the alternative, live. A forecast built at the instrument level, running roughly 220,000 individual CD accounts through one automated pipeline inside OneStream, with loans and investments modeled the same way down to the CUSIP. Prepayment speeds and maturity behavior computed in the platform and rolled straight into the balance sheet. Every number traced back to the instrument that produced it, one click away. If you’ve ever spent a close reconciling a forecast to actuals by hand, that’s the part that lands.

Our own session on governed Finance AI opened with a number worth sitting with: only 7% of CFOs say they see strong impact from their AI spending, and 88% of pilots never reach production. That has little to do with the technology. Teams don’t trust generic AI enough to hand it work that has to hold up to an examiner.

So we showed what governed looks like. SensibleAI runs on the same OneStream data model Finance already uses, and now reaches into the tools your teams already have open. The demo people leaned in for: a Finance manager asking Claude, in plain English, for net interest margin by business line, actuals against plan. A governed connection opened through the bank’s own single sign-on and access controls, pulled the numbers, and answered: blended NIM beat plan by six basis points, split by line of business. Anyone can demo a chatbot answering a Finance question. What made this one usable is that the answer came straight from governed OneStream data the bank can source and defend.

PwC closed the content with the part most teams get wrong. From their client work, about 90% of Finance functions are stuck at what they call Horizon 1: handing individuals AI tools for personal productivity, where the gains never add up to anything the enterprise can measure. Getting past that plateau is a strategy problem more than a tooling one.

Two ideas I’ll be borrowing. First, they build backward. Start from the business outcome you need and the metric that proves it, then work back to the use case, the enablement, and the data, instead of starting with whatever the technology happens to do. Their own numbers back it up: the blockers companies name most often aren’t compute or latency, they’re finding the right use case (46%) and proving ROI (42%). Second, the way they explained an agent finally made it click for a Finance crowd. Think about onboarding a new analyst. You give them a role, you train how they reason, you hand them tools and system access, and you expect them to remember what happened last time. Role, reasoning, tools, memory. That’s an agent.

We closed with a customer fireside chat, the one part a recap can’t recreate: a Finance leader talking in plain practitioner terms about what changed once their numbers lived in one place.

Here’s the thread that ran through the afternoon. Every speaker, including the independent ones, got to the same place from a different direction. Deloitte from the future of the profession, CrossCountry from the forecast, PwC from the operating model, us from the platform. The banks that pull ahead on AI won’t be the ones talking about it the loudest. They’ll be the ones that did the unglamorous work of getting their financial core trustworthy first, so that when they put AI on top, they get to keep it.

If your core isn’t there yet, that’s the conversation worth having. Come find us.