Beyond Chat: Architecting Intelligent Agentic Workflows for Finance Systems in Azure AI

Transcript

1. Beyond Chat: Architecting Intelligent Agentic Workflows for Finance Systems in Azure AI Sam Ayo - Senior AI/ML Engineer https://linkedin.com/in/sam-ayo https://x.com/officialsamayo Building Decisi on-Gra de Intelligence for H igh -Sta kes Fina nc ial Op erat ions
2. • Academic background: Economics, Math, Statistics, Financial Modelling. • Ar e as of Interest: Inferencing, NLP, Audio AI, probabilistic models, Decision science, Research, AI Science • Contributions: Pinecone Vector DB • AI Papers: Notepad Like No Other Sam Ayo Senior AI/ML Engineer Husband to Pharm. Ri Certified Microsoft AI Engineer Associate | ARTIBA Meet Sam Ayo • Programming Languages: Python, C++, C#, Golang. • Recent work: Adaptive fraud monitoring, Federated Analytics Systems for financial reconciliation, Multi-agentic insurance processing system and others.
3. Finance Requires More Than Conversations - It Requires Cognition. Sa m 6Qo OofficialsamaQo Rule Engines Fail Static rules can't handle evolving fraud patterns or contextual nuances Disconnected Systems Siloed data sources prevent holistic risk assessment
4. Why Beyond Chat? • Chatbots = narrow Q&A systems • Finance is a high-stakes, compliance-heavy domain. • Need for adaptive intelligence employing semantics • Need for autonomous, auditable intelligence • Accuracy, traceability, and audit logs are as important as intelligence. Sa m Ayo Officia lsa mayo
5. The Problem with Traditional Systems • Rule engines fail with uncertainty or context gaps due to everchanging fraud patterns • Disconnected systems are often characterized by slow compliance workflows • Manual handoffs where transaction cost and risk are high • Rule engines are not adaptive Sa m Ayo Oofficialsamayo
6. Autonomous Agentic Systems Agent = autonomous decision-maker Agentic AI = reasoning + retrieval + action loop. Sa m Ayo Officia lsa mayo LLM Prompt Tools Missing component Missing component
7. Autonomous Agentic Systems An Agent is an LLM-powered decision node with access to memory, tools, and goals. Agent = autonomous decision-maker Core components: LLMs, RAG, Orchestration Azure AI provides: foundation models, connectors, compliance-grade infra Sa m Ayo Officia lsa mayo LLM Prompt Tools Guardrail Memory
8. Automation General Agents Financial Agents Low risk Controlled risk High risk predefined Handles narrow data Handles multimodal data Has no nature Unilateral in nature Adaptive/reactive in nature Nope Tools are mostly APIs Tools are usually internal ML models Very Low randomness Medium randomness High randomness Automation vs Gen.Agent vs Fin. Agents
9. Finance-Grade Agentic Workflow • Financial workflows are not linear they branch, loop, and depend on domain judgments. • Each stage has its own logic, context, and risk profile. • Multi-agent design mirrors human financial teams: specialized, autonomous, yet coordinated. • Financial systems are multimodal RAG + ML + core SQL What We discovered • Autonomy: Each agent owns its decision scope. • Accountability: Every action is logged and explainable. • Interoperability: Agents can exchange structured messages (JSON/Graph state). • Recoverability: Failures revert to last safe checkpoint. The Discipline We Adopted
10. Finance-Grade Agentic Workflow • How do agentic systems actually work in finance?
11. Finance-Grade Agentic Workflow Four Layer Guardrail Strategy The Challenge: LLMs are probabilistic but Finance demands determinism 1. Input Validation Schema enforcement PII detection & masking Prompt injection detection Business content validation 2. Reasoning Constraint Structured outputs Temperature < 1 for classification Few-shot examples
12. Finance-Grade Agentic Workflow Four Layer Guardrail Strategy The Challenge: LLMs are probabilistic but Finance demands determinism 3. Output Validation Pydantic is your best friend Agent should checkmate agent Business rule validation 4. Audit & RollBack Human in-the-loop What happened – Azure Monitor How often did it happen? - metric Why it happened that way - Trace
13. KYC Automation: End-to-End Architecture Azure Services • Azure AI Document Intelligence • Azure Model catalogue – GPT5 • Azure Container Stack • Azure Key Vault • Azure Monitor Sa m Ayo Officia lsa mayo
14. Sa m Ayo officialsamayo Why Accuracy and Auditability Are Non- Negotiable • We can’t tell customers AI stole their money • Financial systems can’t ‘hallucinate’ - every prediction must be traceable. • LLMs must run under guardrails: deterministic reasoning > creative responses.”
15. Responsible AI & Compliance Sa m Ayo Officia lsa mayo
16. QUESTIONS