AI Concepts for Treasury Professionals
From enterprise AI frameworks to treasury implementation. This is not the beginner course: if you already know how to separate real AI from marketing, this one gives you the frameworks, structures and decision criteria to actually implement AI in your treasury operations, based on real enterprise deployment patterns.
- 01Evals First
Measure AI performance before deploying it in treasury - 023 Maturity Levels
From basic chat to RAG + tools to autonomous agents - 03Expert Led AI
Why treasurers, not IT teams, should lead AI adoption - 04Prompt Engineering
The 7 step structure, few shot examples, chain of thought - 05Bold Automation
Automate end to end workflows, not just 10% of them - 06Graduation Criteria
When a pilot is ready for production
treasuryease.com
Knowledge compounds. Start now.
projected profit improvement Klarna reported from early AI adoption in customer service. It came from starting early, not from having better technology than everyone else.
AI benefits compound over time. The earlier your treasury team starts, the faster the learning curve flattens and the more sophisticated the use cases become. A typical progression:
Cash position queries via AI
FX hedging recommendations
Liquidity stress testing automation
Start six months late and you don't just lose six months, you lose the compounding. The team that started in month one has twelve months of learning by the time you are at month six, and the gap only grows. So the answer to “when should we start?” is always: now. Even if imperfectly. Even if small.
Measure before you deploy
The single most important concept in enterprise AI deployment, and the one most treasury teams skip entirely. Compare AI output with manual results on historical data. If it is more accurate, move to production. If not, iterate.
Compare your AI cash forecast against actuals for 90 days. If the AI forecast is worse than your Excel model, you do not deploy it. You iterate until it beats the benchmark.
Measure the time to generate a board pack with AI versus manually. This is often where AI shows the most visible ROI: time savings are easy to measure and hard to argue with.
Run AI output through your regulatory checks. In treasury, a compliance failure is not a minor issue, so test this rigorously before anything goes to production.
Morgan Stanley reached 98% advisor adoption of its AI tools, but only after running rigorous evals first. Document access jumped from 20% to 80%. The evals built trust, and without trust, adoption does not happen.
From simple chat to autonomous agents
Most treasury teams are at level 1, and that is fine. The goal is not to jump to level 3 tomorrow. It is to understand the path.
“What is our cash position today?”
Simple Q&A, single turn, no memory, no access to your systems. Summarizing a report or drafting an email. Useful, but it is the floor, not the ceiling.
AI accesses bank APIs, pulls live data, runs calculations.
Multi turn conversations with context and memory. Instead of a generic answer, it connects to your data and gives you the real number. This is where the real productivity gains happen.
AI runs reconciliation autonomously, flags exceptions.
Complex multi step workflows with decisions inside defined boundaries. You define the rules, the AI executes, you review exceptions.
Where is your treasury team today, and where do you want to be in 12 months? Most teams should aim to go from level 1 to level 2. That is where the biggest jump in value happens.
The treasurer knows best
The people closest to a process are best placed to improve it with AI. A treasurer doing manual reconciliation every day knows the patterns, the exceptions and the rules better than any IT team.
custom AI assistants created by BBVA employees in 5 months, after the bank gave them direct access to AI tools. Not IT. Not consultants. The people doing the actual work.
Builds an agent for daily cash position reporting with exception flagging.
Knows what exceptions look like, because they see them every day.
Creates an agent for exposure calculations and hedge ratio analysis.
Knows the edge cases from years of dealing with them.
Configures an agent for SWIFT message parsing and reconciliation.
Knows which fields matter and which are unreliable.
The hybrid approach: the domain expert plus AI is more powerful than a developer without domain knowledge. You do not need to be a programmer. You need to know your domain deeply, and then use AI as your tool.
The 7 step prompt structure
The most practical module in the course. Most people use steps 1 and 5 and wonder why the results are mediocre. The difference between a mediocre prompt and a great one is steps 2–4 and 6–7.
- 1Role & Task
Tell the AI who it is and what you want it to do. This sets the context for everything that follows.
“You are a senior treasury analyst…”
- 2Context Data
Give it the data it needs. The AI cannot read your mind or access your systems unless you feed it the information.
“Bank balances across 5 accounts…”
- 3Rules
Set boundaries. These are your guardrails. Without them, the AI will make assumptions you do not want.
“Never recommend transfers without checking credit limits…”
- 4Examples
Show it what good looks like. One of the most powerful steps (more on the next screen).
“Here are 2 correct reconciliation outputs…”
- 5User Input
The actual question or task.
“Classify this transaction…”
- 6Think Step by Step
Where the accuracy gains happen. Ask the AI to reason before answering and it catches errors it would otherwise miss.
“Verify currency, date, counterparty, then classify…”
- 7Output Format
Specify exactly what you want back. Otherwise the AI decides, and it decides differently every time.
“Table: Bank | Balance | Currency | Action”
Use all seven steps, every time. A good prompt takes two extra minutes to write and saves hours of fixing bad outputs.
Show the AI what good looks like
A zoom on step 4, one of the most effective techniques in prompt engineering: provide 2 to 3 examples of correct output before asking the AI to perform the task.
“Classify this transaction.”
Generic classification, inconsistent format, misses edge cases, low accuracy. A slightly different structure every time you run it.
“Here are 2 correct reconciliations. Now classify this one.”
Matches your expected format, handles edge cases because it has seen how you handle them, high accuracy, consistent output quality.
In my experience, adding just 2 to 3 examples improves output quality by 40 to 60%. It costs nothing except preparing the examples once. After that, you reuse them.
Start an example library for your most common tasks: reconciliation formatting, cash position reports, FX exposure summaries. Include the examples in your prompts every time.
Automate end to end, not 10%
This is where most treasury teams think too small. A typical treasury workflow has five steps:
Most teams automate step 4 only, and a human still collects, reconciles, generates positions and distributes. The goal: automate all 5 steps. Not to eliminate humans, but to free them for exceptions, decisions and strategy instead of data entry and formatting.
Klarna: resolution time with its AI assistant
Novo Nordisk: clinical report generation
Lonely Planet: content creation with GenAI
Order-of-magnitude changes come from rethinking the whole workflow, not automating one step. So don't ask “which task can AI help with?” Ask: what does the ideal end to end workflow look like if AI handles the routine and humans handle the exceptions?
When a pilot becomes production
Your pilot works in testing. Before it graduates, it has to meet three categories of criteria.
- Cash forecast accuracy >95% for 30 consecutive days
- Reconciliation match rate >90% without human help
- Report generation time <10 minutes
- System uptime >99.5%
- Support documentation complete
- Team trained on AI assisted workflows
- Data protection audit passed
- Compliance review signed off
- Rollback procedures tested
30 consecutive days, not one good week. Miss the numbers and the pilot stays in testing.
Operational readiness is the part most people skip: a great model nobody knows how to use, with no documentation and no playbook when it breaks, is not production ready.
Tested means you actually ran the rollback, not that you wrote a document saying you could. You need to be able to revert to the manual process within hours.
Write these criteria down before you start the pilot. If you define success after seeing the results, you are just rationalizing.
Success metrics by treasury function
A reference list. You don't need all of them, but pick at least 3 to 4 per function for your use case and track them consistently.
- · Forecast variance %
- · Prediction horizon
- · Data source coverage
- · Trend detection rate
- · Cash buffer accuracy
- · Seasonality fit
- · Auto match rate
- · Exception count
- · Time to reconcile
- · False positive rate
- · Manual intervention %
- · Processing volume
- · Rate accuracy
- · Exposure coverage
- · Hedge effectiveness
- · Cost savings on spreads
- · Alert precision
- · Scenario completeness
- · Generation time
- · Data freshness
- · Distribution success
- · User satisfaction
- · Compliance score
- · Template reuse rate
Two principles that apply to everything
Run AI processes alongside existing manual workflows until performance and reliability are proven.
For treasury: keep your Excel-based process running in parallel with AI for 1 to 3 months. Compare outputs daily. Switch only when confident. You are not replacing anything yet, you are proving the new approach works. This eliminates the “big bang” risk that conservative treasury teams fear most.
Teams give up on prompt engineering too quickly and jump to fine tuning. A few hours of prompt optimization often solves what teams think requires months of custom model training.
Fine tuning is expensive, requires data, and locks you to one model. Good prompts are portable, cheap and fast to iterate. Invest in prompting first; fine tune only when prompting hits a ceiling. In my experience, 90% of treasury use cases can be solved with good prompts.
Key takeaways
- 01
Start with evals. Measure AI against manual benchmarks before deploying. No exceptions.
- 02
Knowledge compounds. Every month of delay is a month of lost learning. Start now, even if imperfectly.
- 03
Put AI in the hands of treasury experts, not just IT teams. The domain expert plus AI beats a developer without domain knowledge.
- 04
Master prompt engineering. Role, context, rules, examples, output format. Use all seven steps.
- 05
Use few shot examples and chain of thought. Two techniques, and the biggest accuracy gains.
- 06
Automate end to end. Not just the report: the entire workflow, from data collection to distribution.
- 07
Define graduation criteria before you start. Performance thresholds, operational readiness, risk management. Write them down.
- 08
Run parallel deployment. Prove AI alongside existing processes before switching. Eliminate the big bang risk.
AI Glossary for Treasury
AI concepts in plain language, plus the AI technologies behind treasury use cases.
Next step · ToolTreasury Prompt Builder
Put the prompt structure into practice on your own treasury tasks.
Built by a treasurer, for treasurers. · treasuryease.com