Back to Projects

Case Study

CRM Agentic AI Architecture Evolution

From prompt-based API execution to reusable enterprise AI skills.

This case study summarizes how an enterprise CRM automation system evolved through multiple architecture stages and is now exploring a reusable skill-based direction for repair, dispatch, scheduling, notification, and route-planning operations.

Architecture focus

Agentic AI execution

System domain

Enterprise CRM workflows

Key direction

Reusable skill systems

Architecture

Evolution Timeline

Each stage solved one set of problems while revealing the next architecture constraint.

Stage 1

Prompt-Based API Execution

Pattern

User Input → Prompt → LLM → API Call → Response

Strengths

  • Fast to prototype
  • High semantic flexibility

Limits

  • Prompt complexity grew quickly
  • Execution order was hard to control
  • Business logic depended too much on LLM behavior

Stage 2

Multi-Agent Architecture

Pattern

User Request → Master Agent ├─ Query Agent ├─ Validation Agent ├─ Notification Agent └─ Route Agent

Strengths

  • Clearer responsibility boundaries
  • Easier prompt ownership

Limits

  • Coordination became more complex
  • LLM cost and latency increased
  • Flow control was still too agent-driven

Stage 3

Workflow-Based Execution

Pattern

Workflow Orchestrator ├─ LLM Node ├─ Function Node ├─ API Node ├─ Condition Node └─ State Node (dynamic path based on workflow state)

Strengths

  • More deterministic execution
  • Better debugging visibility
  • Stronger fit for enterprise processes

Limits

  • Workflow definitions became larger
  • Repeated logic started to appear
  • Reusable capability boundaries were still unclear

Stage 4 / In Progress

Skill-Based Architecture Direction

Pattern

Agent → Candidate Skill → Workflow / API / LLM → CRM Operation Result

Strengths

  • Reusable capabilities under exploration
  • More modular platform direction
  • Clearer separation between agent intent and execution details

Limits

  • Architecture is still being validated
  • Higher abstraction complexity
  • Debugging needs stronger observability across layers

System Model

Candidate Skill-Based Flow

The current direction explores how to separate agent intent from stable execution capabilities.

Agent

Understands task intent

Skill

Exposes reusable capability

Workflow

Controls execution path

API / LLM

Runs tools and reasoning

CRM Result

Completes operation

Comparison

Before And After

The architecture moved from flexible but fragile prompt behavior toward reusable, controlled execution.

Before
After
Prompt controls flow
Workflow controls flow
LLM decides execution details
Agent routes toward candidate skills
Logic repeated across prompts
Capabilities are being shaped into reusable units
Hard to debug failures
Node and skill boundaries can improve traceability
Prototype-oriented structure
Moving toward enterprise-ready execution architecture

Core takeaway

The important engineering shift is not adding more prompts. It is moving toward an execution architecture where agents can choose reusable capabilities, while workflows and backend services handle the deterministic parts of enterprise operations.