How AI Agents Combine Rules & Machine Learning for Smarter Banking Decisions
Banks face a dilemma—customers demand conversational interfaces powered by AI, but regulators require consistent, explainable decisions. Discover how leading institutions automate loan approvals using multi-method agentic AI that blends the flexibility of large language models with the reliability of rules-based systems.
The Limits of LLMs in Banking
Large language models revolutionized customer interactions with their ability to understand natural language, but banks discovered critical limitations when using them for core processes. While LLMs excel at interpreting customer requests and answering policy questions, they struggle with three requirements essential to regulated financial services:
Consistency: Every loan applicant must be evaluated using identical decision logic—LLMs can't guarantee this as their outputs vary even with identical inputs.
Auditability: Regulators demand complete records showing exactly why each decision was made—LLMs provide explanations but not deterministic audit trails.
Precision: Financial decisions require exact calculations using verified data—LLMs sometimes hallucinate numbers or make arithmetic errors.
The breakthrough: Banks now combine LLMs for customer-facing interactions with rules-based systems for core decisions—getting natural conversations without sacrificing compliance.
Multi-Agent Architecture for Loans
A complete loan automation system coordinates several specialized agents, each optimized for different tasks:
Chat Agent (LLM): Handles initial customer conversations, interpreting messy natural language requests ("Can I get a loan for my new boat?") and converting them into structured intents.
Orchestration Agent (LLM): Routes requests to appropriate specialist agents based on the customer's intent—policy questions go to policy agents, loan applications trigger workflow agents.
Policy Agent (LLM + RAG): Answers questions about lending policies by retrieving information from vectorized document stores using retrieval-augmented generation.
Loan Application Agent (Workflow): Manages the stateful loan application process across multiple sessions, remembering where each customer left off.
Eligibility Agent (Rules Engine): Makes yes/no/maybe decisions based on business rules, generating audit logs that explain each determination.
Document Ingestion Agent (LLM): Extracts structured data from unstructured documents like boat brochures with handwritten price notes.
The Conversational Layer
The customer experience begins with natural language interactions powered by LLMs. At 2:15 in the video, we see how the chat agent handles varied customer inputs:
"Can I borrow money for a boat?" might be phrased as "What's your policy on marine loans?" or "I need financing for a 32-foot cabin cruiser." The LLM normalizes these variations into a standardized loan policy inquiry.
For application requests, the orchestration agent identifies the need to launch a loan workflow instance. Crucially, the LLM doesn't process the application itself—it merely initiates the proper backend process.
Key design principle: LLMs handle what they're good at (language understanding) and delegate what they're not (decisions, calculations) to specialized systems.
Workflow Orchestration
Loan applications require state management—customers often start applications then return hours or days later. Workflow systems (like those using BPMN) provide this persistence:
Process Definition: Visual models map each step—eligibility check, document collection, final approval—with decision gates between phases.
Instance Tracking: Each application gets a workflow instance tracking progress. If abandoned, the system remembers where to resume.
Agent Coordination: The workflow calls other agents as needed—asking the eligibility agent for a determination, triggering document collection when required.
At 7:30 in the video, we see how the workflow maintains state even when customers leave mid-process—a capability LLMs lack without significant engineering.
Rules-Based Decision Agents
Core lending decisions use business rules management systems rather than LLMs because:
Deterministic Logic: Rules engines apply the exact same criteria to every applicant. If two customers with identical profiles apply, they get identical decisions.
Transparent Reasoning: The system generates decision logs showing exactly which rules fired and why—critical for regulatory exams and customer disputes.
Performance: Rules engines evaluate thousands of conditions in milliseconds—far faster than LLM inference for complex decision trees.
The eligibility agent exemplifies this—it takes verified customer data (income, existing debts) and applies policy rules to return a clear yes/no/maybe with documented rationale.
Document Processing Agents
Loan applications require documents that traditionally meant manual data entry. Now, document ingestion agents:
Extract Structured Data: From messy inputs like stapled boat brochures with handwritten prices—LLMs excel at interpreting these varied formats.
Validate Completeness: Check if extracted data meets loan requirements (boat make/model/year/price) before passing to decision systems.
Feed Decision Systems: Provide clean, structured data to rules engines—avoiding the reliability issues of having LLMs make decisions directly.
At 12:45 in the video, we see how this combines LLMs' flexibility with rules engines' precision—the LLM interprets the document, but the rules engine uses the extracted data to make the actual lending decision.
Human-in-the-Loop Escalation
When rules engines return "maybe" decisions, the system escalates to human loan officers with AI assistance:
Companion Agent: Surfaces relevant customer data (credit reports, application details) so reps don't waste time searching multiple systems.
Explainer Agent: Translates the rules engine's technical decision log ("Rule 1427 triggered: debt-to-income ratio exceeds policy maximum") into natural language the rep can use with the customer ("Your current loan payments consume 43% of your income—we can't approve additional debt above 40%").
Resumption Handling: After human resolution, the workflow automatically retries the decision—no manual re-entry required.
Hybrid advantage: Humans handle edge cases while AI ensures 80%+ of applications complete automatically—dramatically improving efficiency without sacrificing quality.
Watch the Full Tutorial
See the complete system in action—including how the document ingestion agent processes a boat brochure at 12:45 and how the explainer agent helps human loan officers at 15:20.
Key Takeaways
Modern banking automation requires blending different AI approaches—using each where it performs best while maintaining regulatory compliance.
In summary: Combine LLMs for natural customer interactions, workflow systems for process state management, rules engines for auditable decisions, and document AI for data extraction—connected through agent orchestration. This multi-method approach delivers both great customer experiences and reliable, compliant operations.
Frequently Asked Questions
Common questions about AI agents in banking
Large language models excel at natural conversations but struggle with consistency and auditability. Rules engines provide deterministic decision-making that meets regulatory requirements.
By combining them, banks get the best of both worlds—natural customer interactions through LLMs with reliable, auditable decisions from rules engines.
- 80%+ automation rates for standard loan products
- Full audit trails for every decision
- No more "black box" AI concerns
Workflow systems manage the state of each loan application, remembering where customers left off if they abandon the process mid-flow.
They coordinate between different agents (chat, decision, data) while maintaining a complete audit trail. This is crucial since loan applications often require multiple sessions to complete.
- Track incomplete applications for weeks/months
- Resume exactly where customers left off
- Maintain compliance throughout multi-step processes
LLMs can't guarantee consistent decisions across all customers—a regulatory requirement in lending. They also can't provide the detailed explanations regulators demand.
Decision agents built on rules engines apply the exact same logic to every applicant while generating audit logs that explain each decision in business terms.
- 100% decision consistency across all applicants
- Ready-made documentation for audits
- No regulatory risk from unpredictable outputs
Document ingestion agents extract structured data from unstructured documents like boat brochures or handwritten notes.
While LLMs power this extraction, the resulting data feeds into the rules-based decision system. This combines LLMs' ability to handle messy inputs with the precision of rules-based processing.
- 70% faster than manual data entry
- Processes photos, scans, PDFs equally well
- Validates data completeness automatically
When the rules engine returns a 'maybe', the system escalates to human agents with context. Companion AI helps call center reps by quickly surfacing relevant customer data.
Explainer AI translates the rules engine's technical decision rationale into natural language the rep can use when talking to the customer ("Your debt-to-income ratio exceeds our policy by 3%").
- 50% reduction in call handle times
- Consistent messaging across all reps
- Faster resolution of edge cases
The system combines several technologies: LLM platforms for conversational interfaces, workflow/BPM systems for process orchestration, business rules management systems for decisions.
MCP (Model Contact Protocol) connects these components as interoperable agents. RAG (Retrieval Augmented Generation) enhances policy knowledge bases with up-to-date documents.
- Open standards like MCP prevent vendor lock-in
- Modular design lets you upgrade components independently
- Cloud-native deployment options available
Yes, this multi-agent architecture applies to any financial process requiring both customer interaction and regulated decision-making.
Common use cases include account opening, fraud investigation, insurance underwriting, and investment advisory services—anywhere you need both human-friendly interfaces and auditable decisions.
- Same architecture adapts to multiple products
- Centralized rules management across all processes
- Consistent customer experience throughout
GrowwStacks designs and implements multi-agent AI systems for financial services, combining conversational AI with workflow automation and rules-based decisioning.
We help banks, lenders, and insurers automate complex processes while maintaining compliance. Our team handles everything from architecture design to integration with your existing systems.
- Free 30-minute consultation to assess your needs
- Proven templates for common financial workflows
- Ongoing support and optimization
Ready to Automate Your Financial Processes?
Manual loan processing costs banks an average of $250 per application and creates frustrating customer experiences. Our multi-agent AI systems typically automate 70-80% of applications while improving compliance.