Agentic Multi-Agent Workflow for Marketing Personalisation

The Problem

True personalization requires more than inserting a {{First_Name}} tag. A "Luxury Buyer" needs a narrative of exclusivity, while a "Cart Abandoner" needs urgency and reassurance.

• The Bottleneck: Writing unique, strategic copy for dozens of micro-segments is time-prohibitive.

• The Risk: Marketers do not trust "black box" AI generation. If an AI hallucinates a discount that doesn't exist, it causes financial and reputational damage.

The Solution: A 5-Agent Hierarchy

I designed a system where specialized agents hand off tasks sequentially, separating Strategy from Execution to prevent "strategic drift."

1. The Orchestration Layer (The Manager)

• Role: Parallel Processing Manager.

• Function: Instead of processing segments linearly, the Orchestrator spins up simultaneous workflows for all 12 target segments (e.g., Lapsed Users, High-Value, Cart Abandoners), managing dependencies and data flow.

2. The Strategic Analysis Layer

• Analyst Agent: Ingests raw segment data to build a psychological profile (e.g., "This segment is price-sensitive and urgency-driven").

• Strategy Agent: Converts the profile into a creative brief. It defines the objective, offer logic, and constraints before any copy is written.

  • Constraint Logic: "For Luxury Buyers, strictly forbid discount language. Focus on status."

3. The Creative Execution Layer

• Creative Agent: Receives the locked brief and generates the email copy.

• Self-Correction Loop: The agent assigns itself a Confidence Score. If the score is below 70%, it autonomously regenerates the copy up to three times before escalating, ensuring only high-quality drafts reach the next stage.

4. The Guardian Layer (Quality Assurance)

• QA Agent: Acts as the "Brand Guardian." It audits the output against the original brief and brand guidelines.

• Logic Check: It verifies offer accuracy (e.g., "Did we include the free shipping code mentioned in the strategy?").

• Rejection Protocol: If an error is found, it rejects the draft and sends specific feedback back to the Creative Agent for a re-write.

The "Human-in-the-Loop" Dashboard

Trust is built through transparency. I designed a prototype dashboard that visualizes the agents' reasoning in real-time.

• The "Red Bucket" Protocol: The marketer does not review every email. They only review the "Edge Cases"—drafts that failed the QA Agent's checks multiple times.

• Scale without Loss of Control: This reduces days of manual copywriting to minutes of review, allowing marketers to focus on strategy rather than execution.

Technical Competencies Demonstrated

• Multi-Agent Orchestration: Designing hierarchical agent interactions with clear hand-off protocols.

• Autonomous Self-Correction: Implementing "Reflexion" loops where agents grade and fix their own work.

• Guardrail Engineering: Using QA agents to enforce brand safety and offer logic.

• Human-Computer Interaction (HCI): Designing interfaces that expose AI reasoning to build user trust.