AI is everywhere right now — board meetings, vendor pitches, every conference agenda. And a lot of the excitement is warranted. I use AI tools every day to move faster, think through problems, synthesize information I don't have time to read. The efficiency gains are real.
But when finance leaders at growing companies get pitched on AI-powered treasury tools, I want to ask one question first: do you have what AI needs to actually work?
For a lot of mid-market companies, the honest answer is not yet.
AI doesn't create data. It amplifies what's already there.
This is the part that gets glossed over in the pitch. AI works by recognizing patterns, flagging anomalies, generating recommendations — but only based on what it can see. If the data it can access is incomplete, fragmented, or stale, you don't get better insight. You get faster wrong answers, delivered with more confidence.
For most companies in the growth stage, treasury data is scattered. Cash balances live in bank portals — multiple banks, multiple logins, no consolidated view. Historical reporting might exist in a shared drive or a spreadsheet one person maintains. Covenant agreements are in a PDF from the close of your credit facility. Authorized signers and account structures exist in someone's head.
None of that is accessible to an AI tool. And even when data has been consolidated somewhere, there's a harder question: is it right? If your FP&A forecast is built on assumptions that haven't been pressure-tested, AI will treat shaky inputs as ground truth.
The foundation has to come first
Daily cash operations. To tell you where your cash is going, AI first needs to know where your cash is. If you're logging into 10 or 15 bank portals each morning to piece together a position by hand, there's no consolidated feed for AI to read. There's nothing to analyze. The manual work isn't a process inefficiency — it's a data availability problem.
Cash forecasting. AI forecasting works by identifying patterns in historical cash behavior and projecting forward. That can be powerful — but it assumes your historical data is consolidated and reliable. It also assumes the future looks roughly like the past. For a PE-backed company doing add-on acquisitions, integrating new entities, or navigating a financing event, historical patterns can actively mislead a model. AI doesn't know your business is changing. You have to.
Covenant compliance. Some of the most promising AI applications in treasury involve real-time covenant monitoring — flagging when you're trending toward a breach before it happens. But to monitor covenants, the system needs the covenants. That means credit agreement terms have to be loaded, structured, and maintained somewhere accessible. For most mid-market companies, those agreements live in a legal inbox or a shared drive. No one has built a data feed from those documents into any system, let alone an AI layer on top.
Operational context. Who are the authorized signers on your accounts? What's the approval workflow for a wire above threshold? Which entities are dormant? AI doesn't know any of this. Without it, any recommendation it makes about moving cash or flagging unusual activity is operating without the context treasury professionals carry with them every day.
Getting connected isn't plug-and-play either
Even if your data exists, connecting AI to it is non-trivial. Treasury data is some of the most sensitive in a company. Bank portals are secured for good reason. ERP integrations require real technical work. Credit agreements have confidentiality provisions. Building connections that are secure, auditable, and don't introduce new control gaps takes time and ongoing maintenance.
The pitch deck makes it look seamless. The implementation rarely is.
What it actually looks like when you're ready
I want to be clear: this isn't an argument against AI in treasury. When the foundation is in place, it gets genuinely interesting.
Imagine a tool that looks at your consolidated cash position and says: *"You've been consistently holding $8M in your operating account over the past 90 days with no forecasted deployment — based on your investment policy, here are three options."* Or: *"Entity X runs a consistent surplus by mid-month with minimal activity after that — here are a few ways to repatriate that cash to the parent."* Or real-time covenant monitoring that surfaces headroom erosion before your next compliance certificate is due.
That's real value. That's AI earning its place. But none of it works without the infrastructure underneath it.
So where does that leave you?
A few honest questions worth sitting with: Do you have a consolidated, accurate daily cash position across all entities? Is your historical cash reporting in a format that's actually analyzable — or scattered across bank portals? Have your covenant agreements been structured and maintained somewhere accessible? Do you have documented processes that could be handed off to a system?
If most of those answers are no, the AI conversation is premature. The conversation to have right now is about building the foundation that makes AI worth having.
Get that right first. Then AI has something real to work with.
---
If you're being asked to evaluate AI in treasury and aren't sure your foundation is ready, that's the right place to start. Happy to compare notes — amy@talasadvisory.com
Amy Norman
Founder & Principal, Talas Advisory
Amy is an operator-first treasury leader with over a decade of experience building and running treasury functions in complex, high-growth environments — including a decade at Amazon across cash operations, liquidity management, investments, FX, and banking infrastructure. She founded Talas to bring that same discipline to companies building their treasury function for the first time.