The AI Knows What Loan Yield Is. It Doesn’t Know Yours.

I spent 10 years as head of finance at a publicly traded community bank, and I was one of a handful of people trusted to write reports. There was a reason for that. The numbers mattered, and so did the definitions behind them. When our leadership team or board looked at loan yield, deposit growth, or any other measure of performance, we needed to know exactly what went into the calculation. 

I also remember what it was like to be one of the people everyone came to for answers. A lender needed a different view of a portfolio. A market president wanted to understand what was happening in a branch. Someone had a follow-up question after a leadership meeting. They weren’t asking out of curiosity alone, they were trying to run their part of the bank. 

Often, the quickest way to keep the numbers consistent was to limit who could create them. It gave us control, but it also meant bankers had to wait for the answers that could help them make decisions. I understood why we worked that way because I was responsible for the numbers, but I also understood the frustration because I was the one fielding the requests. 

That experience shapes how I look at AI today. With tools like Microsoft Copilot and natural language analytics, we are getting closer to a world where a banker can simply ask a question about their institution and get an answer without needing to know which report to run or whom to call. That is an incredible opportunity. But having spent years making sure the numbers were right, I know there is another question we have to answer first: Can the banker trust what comes back? 

Knowing banking is not the same as knowing your bank 

One prospect told our team recently they planned to connect their reporting database straight to Copilot and let people start asking questions. Our response was a question of our own. How many definitions of the critical KPIs your bank uses to measure success do you want floating around? 

At KlariVis, we’ve been testing natural language analytics internally, and one thing showed up early. Ask an AI model what yield on loans is and you’ll get a textbook answer. What it doesn’t know out of the box is how your bank calculates it. It doesn’t know whether you use average or ending balances, whether loan fees are in or out, whether participations count, or how nonaccrual loans are handled. Your finance team made those decisions years ago, and most of them live in someone’s head or a spreadsheet formula, not anywhere a large language model like Claude can read. Change any one of them and the number changes, which is how two people can quote yield on loans in the same meeting and both be right by their own math. 

We saw the same thing with a simpler question. We asked an AI assistant to show deposit balances by branch. It came back fast, clean, and confident. It was also wrong. It left out CDs and dropped every branch that carried a zero balance that day. The model knew what deposits are. It didn’t know that bank’s data, and every bank’s data is different from the next bank’s. That’s the gap. Out-of-the-box expertise gets you the textbook. What the model needs is a scout, something that knows where your bank keeps what and points the model to the right place for each question a user asks. 

AI doesn’t fix the problem of competing numbers. It scales it. Instead of two or three versions of yield on loans in a meeting, you get a new one every time someone phrases the question a little differently, and every one of them sounds certain. 

The goal is not to eliminate different questions 

The first job is giving the AI your bank’s KPI definitions and a map of your data before it ever answers a question. That’s where much of our own work is focused right now, and I think it’s the right place to start. There’s a limit, though. Even with the right definitions in place, nothing stops a market president from asking for yield on loans without the fees, or only on the loans their team originated. They’ll get an answer. And sometimes that’s exactly the right question to ask. The risk comes when that variation walks into the board meeting looking like the official number. 

If I were still in the finance seat, I’d want three things from any AI tool my bank put in front of its people. When someone asks for yield on loans, they get the bank’s yield on loans, every time, labeled as the standard measure. When someone asks for a variation, they get that too, with the bank’s number right beside it and the difference explained. And a variation never borrows the official name. It’s labeled “yield on loans excluding fees”, not just “yield on loans”. 

I’d also want whoever owns the definition to see which variations people keep asking for. If the same one keeps coming up, that may be a sign the bank needs a second official measure, and finance can decide to add one. That’s governance doing its job. People get to explore the numbers, and the bank still knows which ones are official. 

This is a leadership responsibility, not just an IT project 

The old model protected the number by limiting the people. It gave the bank a single source of truth, even if it was a slow one. A lot of banks will be tempted to do the same thing with AI and keep it to the few people who really need it. That’s the old model with a new tool, and it defeats the purpose of putting data in more people’s hands. Access and trust are different problems. This better model opens up the questions to everyone and keeps the definitions owned. That’s a banking data governance decision for your CFO and leadership team as much as for IT, and it’s worth making before chaos sets in, with people all over the bank pointing Copilot at your data. 

The definitions aren’t the only thing that has to carry over. Banks control who sees what for good reasons, and those reasons don’t go away because a question came in through a chat window. Someone asking about deposits by branch should get the answer their role allows, the same as they would from a report. Any bank connecting AI to its data should know how access, privacy, and security are handled before anyone types the first question. 

What excites me about where this is headed 

When I think back to my years in banking, I remember both the responsibility of producing reliable information and the long line of people who needed it. I would have welcomed a way for those bankers to get more of their own answers, provided I could still trust the meaning of the numbers. 

That is the opportunity I see in natural language analytics. We can make it easier for lenders, branch managers, market presidents, and executives to understand their business without asking them to become reporting experts. But we must build on a foundation that reflects how their bank actually works. 

AI already knows what loan yield is. The harder and more important work is making sure it knows what loan yield means at your bank. Because the goal is not simply to give everyone an answer. It is to help more bankers make decisions with information they trust.