AI Agents in Treasury & Finance
From Concept to Implementation — a practical workshop for finance professionals.
What we'll cover: architecture · use cases · technology stack · vendors · implementation. This is a workshop, not a demo — every module ends with something concrete: a comparison to use, a roadmap to plan against, or a checklist to run before you build.
- ✓ Explain what actually makes something an "agent" — not a chatbot, not RPA
- ✓ Name the 4 components of agent architecture and the loop that connects them
- ✓ Identify where agents add real value in treasury, with 2 worked examples
- ✓ Compare LLM providers, orchestration frameworks, and build vs buy vs hybrid
- ✓ Plan an 8-week implementation and name the 5 risks to guard against
treasuryease.com
Not a chatbot. Not RPA. Something new.
- Responds to questions
- Single turn or simple dialog
- No actions, only text
- No memory between sessions
- Pre-defined responses
- Executes scripted tasks
- Follows exact rules
- Breaks on UI changes
- No understanding
- Deterministic output
- Reasons about goals
- Chooses tools dynamically
- Adapts to context
- Maintains memory
- Handles ambiguity
AI Agent = LLM + Tools + Memory + Orchestration
The four components that make an agent work
The agent loop: Observe → Think → Act → Repeat
What each component actually does
The "brain" that reasons and decides. Interprets user intent, plans steps, generates responses. Examples: GPT-4, Claude, Gemini.
Actions the agent can take. Query database, call API, read files, send emails, execute calculations. You define what's available.
Short-term: current conversation. Long-term: user preferences, past interactions, learned patterns. Stored in vector DBs or simple storage.
Manages the loop: receive input → call LLM → execute tools → return result. Handles errors, retries, timeouts. Examples: LangChain, AutoGen, Semantic Kernel.
Where AI agents add real value today
Aggregate balances, calculate exposures, alert on thresholds
Analyze patterns, predict flows, suggest actions
Match transactions, identify discrepancies, explain gaps
Categorize, detect anomalies, optimize timing
Validate against policies, flag exceptions, document
Create narratives, summarize data, answer questions
Key insight: agents excel where judgment + data access + action are needed together.
How the agent works, step by step
User asks or schedule runs
Query bank data, AR/AP, history
LLM finds patterns, seasonality
Generate 7/14/30 day projection
Alert, report, recommend
User: "What's our cash position for next week? Any concerns?"
[Fetches data from 12 bank accounts, AR aging, scheduled payments]
Agent: "Current position: €2.4M across EUR accounts. Projected Thursday: €890K after €1.2M supplier batch. Recommendation: consider delaying Supplier_Group_B payment by 48h or draw €500K from credit facility. Historical pattern shows receivables spike on Friday (+€600K avg)."
Fuzzy matching, with reasoning attached
- Exact match only
- Manual exception handling
- No context understanding
- Hours of manual work
- Same mistakes repeated
- Fuzzy matching with reasoning
- Explains why items match
- Learns from corrections
- Handles edge cases
- Confidence scores per match
Example: agent matches "PYMT-2024-1847" to "Payment #1847-2024" (98% confidence)
Reasoning: "Same amount (€45,230.00), same date (±1 day), reference numbers are transposed format of same ID, counterparty 'Acme Corp' matches 'ACME CORPORATION' in master data."
What you need to build an AI agent
Which model provider fits your constraints
| Provider | Models | Pros | Cons |
|---|---|---|---|
| OpenAI | GPT-4o, GPT-4 Turbo | Best tooling, widest adoption, strong reasoning | US-based, data concerns, cost at scale |
| Azure OpenAI | Same as OpenAI | Enterprise compliance, EU data residency, SLAs | Setup complexity, approval process |
| Anthropic | Claude 3.5 Sonnet/Opus | Strong reasoning, large context, safety focus | Smaller ecosystem, fewer integrations |
| Gemini Pro/Ultra | Multimodal, Google Cloud integration | Less mature for agents, availability | |
| Local (Ollama) | Llama 3, Mistral, etc. | Full data control, no API costs, offline | Lower quality, needs hardware, maintenance |
Recommendation: for enterprise treasury, Azure OpenAI or Anthropic for compliance. For experimentation, OpenAI direct.
The glue that makes agents work
✓ Many integrations
✓ Active community
✗ Complexity, abstractions
✓ Agent-to-agent design
✓ Microsoft backing
✗ Newer, less docs
✓ .NET integration
✓ Azure native
✗ Smaller community
CrewAI (role-based agents) · LlamaIndex (data focus) · Haystack (search + RAG) · Custom (full control, more work)
For treasury: LangChain for flexibility, Semantic Kernel if a .NET shop, AutoGen for complex multi-agent workflows.
Three paths to AI agents in treasury
✓ Full control
✓ Exact fit to needs
✓ No vendor lock-in
✗ Needs dev resources
✗ Longer time to value
✗ Maintenance burden
✓ Fast deployment
✓ Proven features
✓ Vendor support
✗ Generic, not tailored
✗ Expensive (€50K+/yr)
✗ Data leaves your control
✓ Balance of speed & fit
✓ Own your data layer
✓ Use approved LLM APIs
✗ Still needs some dev
✗ Integration complexity
✗ Two things to maintain
For most treasury teams: hybrid wins. Use proven components, customize the last mile.
A treasury AI agent in production
Conversational AI assistant for treasury operations built into TreasuryOS. Users interact naturally: "Show me cash position by currency" or "Generate a forecast for next week."
LLM: OpenAI / Azure OpenAI
Orchestration: Custom (TypeScript)
Tools: ReportCore, CalculatorCore, DataCore
Data: Local SQLite, user uploads
Privacy: Confidential AI (Presidio)
- • Data sanitization before any API call
- • Agent generates applications (calculators, reports) — not just answers
- • Multi-currency aware (EUR, USD, GBP, RON)
- • User stays in control — agent suggests, human approves
From zero to production agent
Total: 8 weeks from start to basic production agent. Plan for iteration after.
Technical, organizational, data, budget
API key (OpenAI/Azure/Anthropic) · Python or TypeScript basics · A dev environment (local or cloud)
IT/Security approval for AI APIs · Clear use case with measurable value · Stakeholder willing to test & iterate
Sample data to start (CSV, Excel) · Access to real data sources (phase 2) · Data classification (what can/cannot leave)
MVP: $50-200/month (API costs) · Dev time: 1-2 people, 8 weeks · Production: varies by volume
1 person with Python basics + 1 OpenAI API key ($20) + 1 clear use case + 1 month of sample data = working prototype in 1 week.
What can go wrong and how to prevent it
Validate outputs, cite sources, human review for critical decisions.
Sanitization layer, approved enterprise endpoints, local processing.
Approval workflows, read-only by default, sandbox testing.
Input validation, system prompt protection, monitoring.
Training, confidence scores, "agent suggests, human decides."
Golden rule: agents assist, humans decide. Never automate decisions without oversight.
Who's building AI agent capabilities for finance
Kyriba, ION Treasury, FIS — adding AI features to existing platforms.
Pro: Integrated · Con: Expensive, slow
Microsoft Copilot, AWS Bedrock Agents — general purpose, need customization.
Pro: Scalable · Con: Not treasury-specific
Vic.ai, Trullion, Digits, TreasuryOS — purpose-built for finance workflows.
Pro: Domain expertise · Con: Newer, smaller
LangChain + OpenAI + your data — full control, fits exact needs.
Pro: Tailored · Con: Needs dev capacity
No single vendor does it all. Expect to combine platforms, APIs, and custom work.
Key takeaways
AI Agents ≠ chatbots. They reason, use tools, and take actions. This is new.
Four components: LLM (brain) + Tools (actions) + Memory (context) + Orchestrator (coordinator).
Technology is ready. LangChain, OpenAI, Azure — production-grade tools exist today.
Data privacy is solvable. Sanitization + approved endpoints + human oversight.
Start small. One use case, 8 weeks, working prototype. Learn by doing.
The question is not "if" but "how" and "when". Treasury teams that learn this now will lead.
AI Agents for Treasury: How They Actually Work
A single agent, traced end to end — the version of Level 3 already built.
Next step · CourseAI for Treasury: Build Your Own Tools
Start smaller — no-code tools, then prompts, before a custom build.
Built by a treasurer, for treasurers. · treasuryease.com