Almost one year ago, I started an experiment.
After 10+ years in treasury, managing cash, negotiating with banks, implementing TMS platforms, building endless Excel models, I had a simple question: could AI actually help with real treasury work, or was it just too soon?
I wasn’t looking to replace my job. I wanted to know if AI could make treasury less painful.
Here’s what I found and exactly how it led me to build TreasuryOS.
AI as a Thinking Partner
The first thing I discovered: AI is exceptional at structured thinking. And treasury is nothing but structured thinking. I started using AI the way I’d use a sharp junior analyst, someone to stress-test logic and catch blind spots.
Here’s an actual example: I was building an analysis across five relationship banks (BNP, JPM, Morgan Stanley, RBC, and Deutsche Bank). I’d done this analysis dozens of times in Excel. I gave AI this input (actual figures & banks are changed!)
I'm calculating wallet share for 5 banks. Total fees paid: €2.1M annually.
Categories: Cash Management (€340K), FX Trading (€520K), Trade Finance (€410K),
Credit Facilities (€680K), Transaction Banking (€150K).
JPM breakdown: Cash Management €95K, FX Trading €180K, Trade Finance €85K,
Credit Facilities €220K, Transaction Banking €45K.
Calculate JPM's wallet share.AI came back with the calculation, but then asked: “Are your FX Trading fees purely spot/forward transactions, or do they include FX-related transaction fees from your Cash Management platform? If your TMS books FX through the cash management module, you may be counting those fees twice.” . I checked. The €180K in FX Trading included €47K of transaction fees that were already counted in Transaction Banking. I’d been double-counting for months. JPM’s real wallet share: 27.8%, not 29.8%. Across a €2.1M wallet, that’s a material error when you’re negotiating fee reductions.
That’s when I realized: this isn’t a chatbot. This is a sparring partner that doesn’t get tired and doesn’t get defensive.
From Pain to Prototype
Here’s the truth about treasury technology: most of us have a backlog of tools we wish existed but will never get built. A custom cash positioning dashboard? Six-month IT project. An automated bank fee analyzer? Not a priority. A consolidation tool that works for your entity structure? Maybe next year.
So we build in Excel. We copy-paste. We create workarounds that become permanent. Next month, copy the previous month’s folder and replicate. I decided to test if AI could change that timeline.
The experiment: I wanted a bank fee analyzer that could compare fee schedules across banks and flag outliers. Something I’d been doing manually in Excel for years. Here’s the prompt I used (actual figures & banks are changed!):
I receive monthly fee statements from 5 banks in different formats.
I need to:
1. Categorize fees into standard buckets (wire fees, account maintenance,
FX spreads, credit facility fees, trade finance fees)
2. Calculate cost-per-transaction for comparable services
3. Flag any fee that's >20% higher than the average across banks
Here’s a sample from Bank A:
– Domestic wires: €8.50/transaction, 340 transactions = €2,890
– International wires: €25/transaction, 89 transactions = €2,225
– Account maintenance: €450/month
– FX spot transactions: 12bps average spread on €4.2M volume = €5,040
Build me a comparison framework.
What AI generated: A structured template with fee categories, formulas for cost-per-transaction, and conditional formatting logic for the outlier flags.
What worked: The categorization logic was solid. The cost-per-transaction formulas were correct. I had a testable structure in 20 minutes.
What failed: AI assumed all banks report fees the same way. They don’t. One bank bundles account maintenance into a “relationship fee.” Another one separates SEPA and non-SEPA wires. I had to add mapping rules for each bank’s specific format.
What I learned: The first output was maybe 60% right. But I could test it immediately. Break it. Fix it. Test again. Each cycle took hours, not weeks. After 4 iterations over 3 days, I had a working fee analyzer. Not perfect. But it caught a €34K overcharge on trade finance fees that I’d missed in manual review.
Patterns Emerged
After daily use, I mapped what AI actually does well in treasury and where it fails.
What AI handles naturally:
- Structure and hierarchy. Treasury is full of entity trees, currency relationships, fee categorizations. I described our legal entity structure (eg. 14 entities across 6 countries with 3 functional currencies) and AI mapped the intercompany flows in minutes. In Excel, that’s a half-day exercise.
- Pattern recognition. I uploaded 12 months of bank statements for one entity and asked: “What’s unusual?” AI flagged that wire fees spiked 40% every quarter-end. Reason: AP team batched customer refunds quarterly instead of processing them with regular payment runs. A process fix saved €8K annually.
- Translation. I described what I needed: “I want to see which entities are holding excess cash above their working capital buffer, and where that cash should be swept to.” AI translated that into logic I could build: threshold rules, sweep hierarchies, currency conversion sequences.
Where AI fails without you:
- Context. AI doesn’t know that your credit agreement prohibits sweeping from the German entity without lender consent. Or that your Brazilian subsidiary has capital controls that affect everything. I had to feed this context manually, every time.
- Judgment calls. “Should I hedge this EUR/USD exposure?” AI can model scenarios, calculate break-evens, show you the Greeks. But the decision, your risk appetite, your board’s tolerance, your view on the market, that’s all yours.
- Operational edge cases. AI doesn’t know that SAP exports dates as DD.MM.YYYY but your bank portal uses MM/DD/YYYY. Or that Bank X sends MT940 statements with non-standard tags that break your reconciliation. Domain expertise matters. You can’t outsource 10 years of operational scar tissue.
The pattern: AI accelerates execution but doesn’t replace expertise. The more treasury knowledge I brought, the better the output.
What This Became: TreasuryOS
After months of these experiments, I had a collection of tools I’d built for myself:
- A fee analyzer that flags outliers across banks
- A cash positioning dashboard by entity and currency
- A wallet tracker with quarterly trend analysis
- An FX exposure calculator with hedge scenario modeling
And a realization: if I could go from frustration to functional prototype in days, other treasurers could too.
That’s why I built TreasuryOS—a platform where treasurers create their own tools without waiting for vendors or IT.
One week after launch: users onboarding daily, building apps organically. Not by developers. By treasury people solving their own problems.
If You Want to Start Tomorrow
Here’s how to start with AI today:
Step 1: Pick one repetitive task
Start with something you do monthly that annoys you. Bank fee reconciliation. Cash positioning. Intercompany balance confirmation. Something with structure.
Step 2: Describe it to AI like you’d describe it to a new analyst
Don’t write code. Write plain language:
Every month I receive fee statements from 5 banks.
I manually enter them into Excel, categorize by fee type,
and calculate total cost per bank.
It takes 4 hours and I hate it.
Here's what Bank A's statement looks like: [paste sample]
Help me build a faster process.Step 3: Test the output immediately
Don’t evaluate if it’s “good.” Evaluate if it’s testable. Run it with real data. See what breaks.
Step 4: Iterate
My first useful tool took 4 iterations over 3 days. Expect the same. The first output won’t be perfect. That’s fine. You’re compressing a 6-month IT project into a week.
The Bottom Line
The gap between “I wish this existed” and “I built this” is smaller than it’s ever been. The technology is ready. The question is whether treasury teams will use it.
If you want to try building your own treasury tools, TreasuryOS is live. No consultants. No 6-month implementations. Just treasury people solving treasury problems.
Enjoy!
