AI in Treasury: Separating Real from Marketing
This is not a technology course. You will not become a data scientist. This is a practitioner's course — click through it one screen at a time, and each module ends with one practical action you can do immediately.
- ✓ Ask the right questions when a vendor pitches you "AI-powered treasury"
- ✓ Recognize the difference between real AI and rebranded Excel or RPA
- ✓ Understand which AI use cases are proven vs. hype in treasury
- ✓ Use the Vendor AI Scorecard (included) in your next vendor evaluation
- ✓ Identify what AI tools you already have access to at zero cost
treasuryease.com
Every vendor claims AI. Not all of them mean the same thing.
There are three distinct categories you will encounter — and the most important skill in this course is identifying which one a vendor's product actually belongs to.
| Category | What it actually is | Treasury example |
|---|---|---|
| Real AI / ML | Algorithms that learn from your data and improve over time. Finds patterns a human or Excel cannot. | Cash flow forecast trained on 3 years of your payment history |
| Statistical Methods | Math formulas. Reliable, transparent, fast. Sold as AI by many vendors. | ARIMA forecast, Holt-Winters seasonality — Excel Forecast Sheet does this. |
| RPA + Rules | Bots that copy-paste data based on fixed if/else rules. Zero learning. Zero AI. | Bot that pulls MT940 files every morning and pastes into Excel |
Why this matters: RPA and rule-based systems are useful. But you should not pay an AI premium for them, and you should not expect them to handle exceptions, learn from new data, or improve without manual reprogramming. The question is never "do you use AI?" — it's what kind?
The AI vocabulary you need — nothing more
| ML (Machine Learning) | Learns from historical data. Used for forecasting, anomaly detection, classification. |
| LLM | Reads and writes text. ChatGPT, Claude. Useful as an assistant, not a forecasting engine. |
| NLP | Natural Language Processing. AI that reads, understands and generates human language. |
| RPA | Automates repetitive tasks. No learning. Valuable but not AI. |
| Deep Learning / LSTM | Advanced ML for complex sequences. Used in serious cash flow forecasting. |
| Generative AI | Creates new content: text, summaries, reports. Not a data model. |
| RAG | LLM connected to your own documents. Answers questions from your policies or contracts. |
Write down one tool your company currently uses that claims to be AI. Which category does it belong to? You'll revisit this in Module 4.
What AI can actually do in treasury
Cash Flow Forecasting
Learns seasonality, identifies anomalies in historical data, incorporates multiple factors (FX, customer behavior, tax cycles). Produces probabilistic forecast ranges, not just one number.
Uses a fixed formula or ARIMA and calls it machine learning. Cannot adapt to new patterns without manual reconfiguration.
Red flag phrase: "Our AI forecasting engine uses advanced algorithms." Ask: which algorithms specifically?
Minimum real answer: Random Forest, Gradient Boosting, LSTM, or Temporal Fusion Transformer — trained on your data specifically.
What AI can actually do in treasury
Anomaly Detection on Payments
Flags payments that deviate from established patterns for that counterparty, amount range, or time of month. Adapts as normal behavior changes.
Checks if amount is above a fixed threshold. That is a rule, not AI.
Red flag phrase: "We alert you when payments exceed defined limits." That is a filter, not anomaly detection.
Minimum real answer: Model trained per counterparty, uses statistical deviation, explains why a payment was flagged.
What AI can actually do in treasury
Document Intelligence (OCR + AI)
Reads unstructured PDFs (bank statements, invoices, contracts), extracts specific fields, structures output without templates.
OCR that requires a fixed template per document format. Breaks when the layout changes.
Red flag phrase: "Supports all major bank statement formats." Ask: what happens with a format not in your list?
Minimum real answer: Model-based extraction, not template-based — can handle new formats within defined document types.
What AI can actually do in treasury
Automated Reconciliation
Matches transactions even with text mismatches (Invoice #123 vs INV-123), learns your specific naming conventions, flags true exceptions for human review. It learns from your corrections: every match you fix makes the next one better.
Exact-match rules. Anything that does not match exactly becomes an exception.
Red flag phrase: "Automated reconciliation with 99% match rate." Ask: on what type of data, and what happens to the 1%?
Minimum real answer: Fuzzy matching, confidence scoring, exception queue with explanations.
What AI can actually do in treasury
Conversational Assistant for Treasury Data
Answers natural language questions against your actual data: "What is our EUR cash position today?" or "Which suppliers have payments due this week?"
Hardcoded chatbot that routes to predefined screens. Cannot answer a question it was not explicitly programmed for.
Red flag phrase: "Our AI assistant answers treasury questions." Ask: can I ask it a question right now that is not in your demo script?
Minimum real answer: LLM connected to live data via API or RAG — can handle novel questions within the data domain.
These five are proven. Everything else needs proof. If a vendor offers something outside them, ask: show me a customer using this in production today. Not a pilot, not a proof of concept. Production.
For each of the five use cases, write one question you would ask your current TMS or bank portal vendor. You'll use these in Module 4.
Vendors use specific phrases that sound impressive but signal nothing
| Phrase you'll hear | What to ask instead |
|---|---|
| "AI-powered forecasting" | Which algorithm? Trained on my data or a generic dataset? |
| "Self-learning system" | What does it learn from? How often does it retrain? Metrics before/after? |
| "Intelligent automation" | What happens on an untrained exception — does it decide, or escalate? |
| "Advanced analytics" | Advanced compared to what? Descriptive or predictive? |
| "Explainable AI" | Show me an actual explanation — a live output, not a slide. |
| "No black box" | Can I see feature importances? Explain one flagged payment? |
| "Proprietary algorithm" | Translation: we won't tell you what it does. What type of model is it, in plain English? |
| "Next generation" | Compared to what? Show me the benchmark. |
| "Cognitive" | Borrowed from the IBM Watson marketing era. What does the system actually do? |
| "Smart" anything | Smart matching, smart analytics: vague by design. Smart how, specifically? |
"Can you show me this feature running on data that looks like mine, not your prepared demo dataset?"
Or, more bluntly: "Can I upload my own data right now and see the output live?"
A vendor with real AI will say yes. A vendor with hardcoded demos will change the subject.
- • Does my data leave my environment to train your model?
- • Is your model trained on aggregated data from multiple clients?
- • Who owns the trained model? If I leave, do I retain it?
- • What is your data retention policy after contract termination?
Before your next vendor call, write down the three phrases from this module you've already heard from that vendor. Prepare the matching questions.
Score your next vendor demo, live
Answer each question 0–2 as you'd rate the vendor's actual answer. A total below 6/10 is a fail.
What specific AI algorithm do you use for this feature?
Acceptable: names a real algorithm (Random Forest, LSTM, XGBoost, Transformer) — not "advanced AI."
Is the model trained on my data, or a generic dataset?
Acceptable: trained on client-specific data after onboarding; generic pre-training only as a starting point.
Can you demonstrate this on realistic data, not a demo dataset?
Acceptable: agrees and demonstrates — does not redirect to slides.
Does my data leave my environment? Where is it processed?
Acceptable: clear answer on cloud region, data residency, and whether data trains other clients' models.
What happens on a pattern the AI has never seen — how is it escalated?
Acceptable: describes human-in-the-loop escalation — doesn't claim the AI handles all cases.
Score every question to see the verdict — 9–10 real AI, 6–8 pilot it, 0–5 walk away.
- • Is the AI component priced separately or bundled? Bundling is often a way to hide that the "AI" is minimal.
- • Can you name three or more clients using the AI features in production? Not pilots. Production.
Before paying any vendor, check your current toolkit
| Tool | Where to find it | What it does for treasury |
|---|---|---|
| Excel Forecast Sheet | Data menu in Excel | Seasonality-aware forecast — not AI, but better than a manual trend line. |
| Power BI Anomaly Detection | Analytics pane, any line chart | Flags statistical anomalies in your data automatically. |
| Power Automate Desktop | Pre-installed, Windows 10/11 Pro | Automates repetitive copy-paste — not AI, but eliminates manual work. |
| AI Builder (Power Platform) | Power Automate / Power Apps | Invoice processing, text classification — included in many M365 licenses. |
| Copilot in Excel | Excel, with M365 Copilot license | Answers natural language questions about your spreadsheet data. |
| Power BI / Tableau | Your existing BI license | Forecasting visuals and natural language Q&A. Most treasurers never turn these on. |
| Python / Google Colab (free) | colab.research.google.com | A full ML pipeline (pandas, Prophet, scikit-learn) at zero license cost; Colab needs no install. Don't upload confidential data. |
| ChatGPT / Claude | Browser, free tiers | Summarize reports, draft analysis, explain complex financial data in seconds. Sanitize before you paste. |
The honest limitation: free tools require you to set them up. Vendor tools are pre-built but require you to evaluate them critically. Both require your time. Neither works without your data and your judgment.
Log into Power BI with your existing license and enable Anomaly Detection on one cash flow chart. Two minutes, immediate value.
What you know now
Vendor says "AI-powered" and you evaluate the UI
→ You ask which algorithm, trained on what, with what governance
AI feels like a black box you cannot challenge
→ You have a vocabulary and a scorecard for any claim
You pay for features you cannot verify
→ You negotiate based on what the product actually does
You assume you need a big vendor to use AI
→ You know what you already have and where to start today
AI Use Case Canvas
Turn a vague AI idea into a concrete, testable experiment.
Next step · CourseChoose a Task Before Choosing a Tool
Find a genuine AI use case in your own treasury work.
Built by a treasurer, for treasurers. · treasuryease.com