AI for Treasury Practitioners
A practical guide to understanding and using AI. No theory. No hype. Just what works.
Thirteen concepts a working AI product actually rests on — from how LLMs think to how to work effectively with Claude day to day, told through one running example: TOSY, a real treasury AI agent in production.
- ✓ Explain what LLMs are actually good and bad at — and why that matters for your job
- ✓ Write clear prompts and system prompts using a repeatable structure
- ✓ Tell agents, agent orchestration, MCP, and RAG apart — and know when each applies
- ✓ Explain why evals and caching matter, even if you didn't build them yourself
- ✓ Work more effectively with Claude and Claude Code, day to day
Alina Turungiu · TreasuryOS · treasuryease.com
How LLMs Actually Work
Treasury analogy
Imagine a new hire who has read every treasury textbook, every bank statement format, every regulation. They can predict what comes next in a sentence because they've seen millions of similar patterns. But they have no real experience. They haven't made a payment. They haven't reconciled an account. They pattern-match brilliantly but don't truly "understand."
LLM = Large Language Model. Trained on text to predict the next word.
They are extremely good at:
- • Summarizing, writing, analyzing
- • Following complex instructions
- • Generating code and structured output
They are bad at: math, real-time data, remembering past conversations, being 100% accurate.
Your job: give clear instructions and verify the output. AI does the heavy lifting. You do the thinking.
Prompt Engineering
How you talk to AI determines what you get back
"Give me a bank extract."
Result: AI guesses format, currency, date range. Output is generic and wrong.
"Generate a sample MT940 for account RO49AAAA, Oct 2025, 15 transactions, RON, with salary payments."
Result: Exactly what you need.
Who are you talking to?
"You are a treasury systems expert."
What do you already know?
"We use Coupa and SAP."
What do you want?
"Build a cash position dashboard."
How do you want it?
"Return JSON with currencies."
System Prompts
The job description for your AI agent
Treasury analogy: When a new employee joins your team, you give them a document: who we are, what we do, how we communicate, what we never do. They refer to it for every decision.
A system prompt is exactly that. Instructions the AI receives before any conversation starts. The user never sees them. But they determine everything about how the AI behaves.
This is the most important part of any AI product. A mediocre system prompt = a mediocre AI assistant, regardless of how smart the model is.
Who the AI is, its expertise
What it must and must not do
Domain-specific context
How to structure responses
Safety and boundaries
AI Agents
AI that doesn't just answer, it acts
You ask: "What's my cash position?"
AI says: "I don't have access to your data."
It can only talk. It cannot do.
You ask: "What's my cash position?"
AI queries your database, pulls balances, converts currencies, and returns the answer.
It thinks, decides which tools to use, and acts.
User asks a question
Decide what tools to use
Query database, call API, calculate
Return structured answer
Agent Orchestration
Multiple specialized agents working together
Treasury analogy: You don't have one person doing everything. Cash manager handles cash. FX dealer handles FX. Risk manager handles risk. A Head of Treasury routes work to the right person. Agent orchestration is the same: specialized AI agents coordinated by a router.
Decides which agent handles the request
Cash position, balances, statements
FX rates, exposure, conversions
Dashboards, forecasts, exports
You already built this in TOSY.
MCP (Model Context Protocol)
SWIFT for AI: a standard way for AI to connect to tools
Treasury analogy: SWIFT is a universal protocol for banks to communicate. MCP is a universal protocol for AI to communicate with tools and data sources. Instead of building custom integrations for each tool, MCP provides one standard. Created by Anthropic (the company behind Claude).
In your product: MCP lets TOSY connect to your users' data sources.
RAG (Retrieval Augmented Generation)
AI that looks things up before answering
Treasury analogy: When a colleague asks you about a bank agreement, you don't answer from memory. You go to the folder, find the document, read the relevant section, and then answer. RAG is exactly this for AI.
"What are my bank fees for Q3?"
AI searches your documents and data
Finds the relevant sections/records
Answers based on actual data, not guessing
Why it matters: Without RAG, AI makes things up ("hallucinations"). With RAG, AI answers from your actual data.
Game changer for TOSY: imagine users asking "What does my credit agreement with Bank X say about covenants?" and getting an accurate answer.
Evaluation (Evals)
How you know if your AI is actually working
Treasury analogy: Treasury reconciliation. You compare what should be there with what is there. Evals do the same for AI: compare AI's answers with correct answers. Like a treasury audit, but for your AI.
"It seems to work."
"A user complained it gave wrong FX rate."
"I think the last update made it worse but I'm not sure."
You are flying blind.
"TOSY answers 94% of cash queries correctly."
"FX conversion accuracy dropped 3% after last update."
"Reconciliation agent hallucinates on 6% of edge cases."
You have numbers. You can improve.
Start simple: 50 questions you know the answer to. Run them through TOSY. Count how many it gets right.
Caching
Pay once, use many times
Treasury analogy: Instead of calling the bank every time someone asks for the balance, you check once in the morning and use that number all day. Caching stores repeated AI instructions so you pay once, not every time.
TOSY has a system prompt (~4,000 words).
60 users × 10 messages/day = 600 API calls/day.
Without cache: you send 4,000 words × 600 times. You pay for all of it.
With cache: you send 4,000 words once. The next 599 calls cost a fraction.
Prompt caching can reduce API costs by up to 90%.
You already implemented this. Now you know why it matters.
Claude Code
Your AI development partner in the terminal
Claude Code is a command-line tool where AI reads your entire codebase, understands context, writes code, runs tests, and fixes bugs. It's not a chatbot. It's a co-developer who sees everything.
Connectors & Skills
Extending what Claude can do in claude.ai
In Claude.ai settings, you can connect external services:
- • Google Drive
- • Gmail
- • Google Calendar
- • Notion
- • Slack, Jira, and more via Zapier
Once connected, Claude can read your Drive files, search emails, create calendar events, etc. These are MCP servers running behind the scenes.
Skills are instruction sets that teach Claude Code how to do specific tasks well:
- • Create Word documents (.docx)
- • Create presentations (.pptx)
- • Create spreadsheets (.xlsx)
- • Create PDFs
- • Frontend design, and more
Think of skills as best-practice recipes. Claude Code reads them before starting a task to ensure quality output.
Both are examples of giving AI access to tools and knowledge it doesn't have by default.
Working Effectively with Claude
Practical rules for getting better results every day
"I need a 13-week cash forecast" is better than explaining the steps.
Show Claude what the result should look like. A screenshot, a sample, a format.
Ask: "Are you sure? What could be wrong here? Attack this." Claude gets better when pushed.
Think with Claude AI. Build with Claude Code. Don't mix the two.
Don't ask for everything at once. Phase 1, Phase 2, Phase 3. Verify each before moving on.
When something works perfectly, save the prompt. You'll need it again.
Claude's second or third attempt is often better. Iterate. Refine. Push back.
Your AI Stack
What you already built (and didn't know it had a name)
6 out of 8 core AI concepts: already in production.
You don't need to learn AI theory.
You need to label what you already do.
"I don't teach AI theory.
I build AI systems in production.
Then I explain what I built."
Alina Turungiu | TreasuryOS
Key takeaways
LLMs pattern-match brilliantly but don't truly understand. Give clear instructions, verify the output.
A good prompt has 4 elements: Role, Context, Task, Format. A system prompt is the same idea, set once.
Agents act, not just answer. Orchestration routes work between specialized agents, the way a Head of Treasury routes work between people.
MCP connects AI to tools like SWIFT connects banks. RAG grounds answers in your real data instead of guesses.
Evals tell you if your AI actually works. Caching makes running it affordable. You probably need both, eventually.
You don't need to learn AI theory. You need to label what you already do.
AI Concepts & Orchestration
The deeper glossary: LLMs, fine-tuning, RAG, embeddings, agentic patterns, and how the layers connect.
Next step · CourseAI Agents in Treasury & Finance
From concept to implementation — architecture, use cases, roadmap, risks.
Alina Turungiu · TreasuryOS · treasuryease.com