David Zwick, CFO of Billtrust, is a global-minded technology executive with success driving operational and financial growth strategies.
Before every quarter close, I ask some version of the same question: Which of our customers have started paying later than they used to? Nobody can answer it that afternoon. Someone takes the question away, and a few days later a spreadsheet comes back that was accurate as of the morning it was built. By then, the decision that depended on it has usually been made without it.
The distance between the moment a question gets asked and the moment anyone can answer it is the most underrated cost in corporate finance, and it never appears in a report. What it produces is a decision made a week later than it should have been. Or a credit hold on a customer who didn’t warrant one in the first place.
Most of the conversation about AI in finance has skipped past this entirely. The attention has gone to agents that go and do things, releasing payments, working collections queues, closing the books. I understand the appeal, and some of that is coming. Before any of it is worth having, though, a finance organization must be able to ask its own data a question and get an answer inside the same conversation. Very few can.
Fragmentation is the reason, and it is not going away.
Very few large companies run on one ERP. Mergers bring their own systems, regions run whatever they standardized years ago and departments buy their own tools, leaving most enterprises running several platforms at once. Many organizations still operate multiple ERPs, driving up costs, fragmenting data and slowing decisions. Billing sits in one system, cash application in another, relationship history lives in the CRM, and the collections context that would explain all of it is in somebody’s inbox. Answering anything that touches more than one of those means somebody must pull from each of them and reconcile what comes back.
That fragmentation is why most teams cannot see what their customers are doing until well after they have done it. In my organization’s 2026 Economic Headwinds survey of roughly 549 finance leaders, 67% reported that their customers were paying slower than they had six months earlier, and one in five described the slowdown as significant. That’s too broad a shift for a month-end cadence to catch. Most teams explain it after the quarter closes, once the window to act has passed.
Finance leaders are not confused about the stakes. The AFP’s 2025 Treasury Benchmarking Survey found that 73% of treasury practitioners rank cash management and forecasting as their top priority, which is hard to square with how long it still takes to get a straight answer about receivables.
That gap, though, is closing faster than most finance leaders realize. A growing standard called MCP—short for Model Context Protocol—lets AI assistants connect directly to a company’s live systems and pull a real answer instead of a guess. It’s early, but it’s the plumbing that makes “ask your own data a question” something you can actually do this year, not a slide in next year’s road map.
Answering is the first useful thing AI does in a finance function.
Speed is the benefit everyone expects. Decisions get made at the pace the business is moving rather than the pace of the reporting calendar, and the cash trapped in your receivables—earned, invoiced and sitting there anyway—starts getting treated as the live problem it is instead of a number reviewed in arrears.
The effect I find more interesting is what happens to the questions. When an answer costs three days of someone’s time, a finance team rations its questions. It asks the ones it can justify asking, which tend to be the ones where somebody already suspects the answer and wants confirmation. Cheap answers change that calculation. An analyst who can check something in a minute will check it, and then check the thing the first answer suggested, and the useful finding tends to turn up two or three steps in. Plenty of good analysts have dropped something interesting because chasing it cost too much to justify, and no one puts a number on what that costs.
Retrieval also deserves to come before execution for a plainer reason than the one people usually give. An answer assembled from your own records can be traced back to those records. A person can check it, and a decision made on it is one you can defend to an auditor or a board six months later when nobody remembers the details. Execution built on top of the same fragmented data your team already struggles with does the work faster and gives you less visibility into how it got there. Any finance leader who has spent a weekend reconstructing why a number moved understands the difference between those two positions.
Here’s a test worth running this week.
This does not call for a new strategy, and it does not require replacing anything. It requires a fairly boring commitment to making your own data answerable before you ask software to act on it.
So, run the test. Pick a question you actually need answered. Which accounts have moved from paying on time to paying at 45 days over the last two quarters, say, and which of those also changed their order volume? Ask your team for it, and note the time. Note the time again when the answer arrives, and be honest about how much of the delay was people waiting for other people rather than anyone doing analysis.
That number tells you how much of your team’s week goes to assembling information and how quickly you would spot a customer’s behavior changing while there is still something to do about it. Anything more ambitious you have in mind for AI in your finance function will run on the same plumbing.
The information provided here is not investment, tax or financial advice. You should consult with a licensed professional for advice concerning your specific situation.
Forbes Finance Council is an invitation-only organization for executives in successful accounting, financial planning and wealth management firms. Do I qualify?
