AI Agent Feasibility Research: Email Marketing Automation

Discovery Process

Interviewed product team to understand current manual workflows and customer challenges. Analysis revealed customers manually managing 3-round engagement sequences:

• Send initial email
• Track whether recipient opened the email
• Track whether recipient clicked links in the email
• Repeat with adjusted content for rounds 2 and 3 based on engagement

Additional pain point identified: customers manually personalized each email using known customer data and sender/product information. This personalization process was time-consuming and difficult to scale across large recipient lists.

Technical Experimentation

Built proof of concept to validate technical feasibility of automating the identified workflow.

Initial Approach and Limitations

Initial implementation used monolithic LLM prompts that attempted to encode the entire 3-step engagement sequence in a single prompt. This approach proved fragile in practice - minor changes to prompt wording caused unpredictable variations in output quality and structure. The complexity of maintaining a single prompt that handled all sequence positions and engagement states made the system unreliable and difficult to iterate on.

Refined Approach

Pivoted to a fine-tuned model architecture where the model learned to recognize sequence position (round 1, 2, or 3) and engagement state (opened email, clicked link, or no engagement) to generate contextually appropriate emails. This decomposition of the problem into smaller components produced more reliable and maintainable results.

Integrated Retrieval-Augmented Generation (RAG) to handle the personalization requirement. The system could automatically incorporate customer-specific data and sender/product information into generated emails, addressing the manual personalization pain point without requiring custom prompts for each use case.

Generated synthetic training data using ChatGPT since no real customer sequence examples existed at the project's exploratory stage.

Communication & Impact

Documented technical approach and business value in a white paper for product management and executive leadership. The proof of concept successfully demonstrated that AI agents could reliably automate multi-step email workflows that customers currently performed manually.

This research directly validated the feasibility of AI agents for email marketing automation within the company's product. The work led to development of the customer segmentation agent, which aligned more closely with near-term product roadmap priorities while building on the technical foundations established in this research.

Development Environment

• Python
• Jupyter Lab
• LangGraph
• LangChain
• ChromaDB
• Ollama
• NumPy
• ChatGPT (synthetic data generation)
• Claude