CampaignPilot: Cross-System Marketing Intelligence Agent

When CPA spikes, Google Ads says "optimize bids." CampaignPilot checks your website, products, and support tickets — and tells you it's your CDN that crashed. 18 ES|QL tools, ELSER search, 11 indices.

Comment

Inspiration

Every marketing team runs the same playbook when CPA spikes: check bids, check creatives, check audiences. But the real cause is often invisible — a CDN outage crashing your landing page, a product price increase killing conversions, or a wave of support complaints driving churn. These root causes live in completely separate systems that no ad platform connects.

We built CampaignPilot to answer one question: What if an AI agent could reason across ads, website, product, AND support data to find the real root cause automatically?

What it does

CampaignPilot is a cross-system marketing intelligence agent that monitors multi-channel ad campaigns in real time. When anomalies occur, the agent doesn't just analyze ad metrics — it automatically investigates website health, product catalog changes, and customer support sentiment to deliver a root cause diagnosis with actionable recommendations.

Example: CPA spiked 130% → Agent checks ads (✅ normal) → checks website (🔴 CDN degraded, 6800ms load time) → checks support (✅ normal) → Diagnosis: WEBSITE-LEVEL problem, route to engineering, not marketing.

How we built it

  • 11 Elasticsearch indices with 22,000+ documents covering campaigns, channel metrics, creative assets, audience segments, website events, product catalog, and support tickets
  • 18 custom tools in Agent Builder: 11 ES|QL query tools for structured analytics, 3 ELSER semantic search tools for natural language search, and 4 platform tools
  • LOOKUP JOIN for cross-index campaign correlation
  • semantic_text fields (×7) enabling natural language search across competitors, creatives, and support tickets
  • 3 Elastic Workflows for automated campaign optimization, budget reallocation, and daily briefing generation
  • Custom Streamlit frontend with real-time dashboard, anomaly scanner, agent chat, and action log
  • MCP Server + A2A Protocol for external integration

Challenges we ran into

The main challenge was designing the agent's reasoning flow. A naive agent stops after checking ad metrics and concludes "creatives are fine." We needed the agent to reliably cross system boundaries — from ads → website → product → support — every time. We solved this by designing a structured system prompt with an explicit cross-system root cause analysis step and a classification table (AD-LEVEL / WEBSITE-LEVEL / PRODUCT-LEVEL / CUSTOMER-LEVEL) that guides the agent's reasoning path.

Accomplishments that we're proud of

  • The agent can correctly diagnose that a CDN outage (not bad creatives) caused a CPA spike — a conclusion that requires reasoning across 3 separate data sources
  • 18 tools working together through Agent Builder's orchestration, with the agent autonomously selecting the right tool at each step
  • A complete Streamlit dashboard that gives marketers a single-pane view across all campaigns, anomalies, and agent conversations

What we learned

  • ES|QL is remarkably powerful for time-series anomaly detection with STATS + EVAL + CASE patterns
  • ELSER semantic search transforms support ticket investigation from keyword matching to intent understanding
  • The real differentiator isn't more tools — it's teaching the agent when to cross system boundaries
  • Agent Builder + ES|QL + ELSER is a surprisingly complete stack for building production-grade AI agents

What's next for CampaignPilot: Cross-System Marketing Intelligence Agent

  • Real data integration — connecting to live Google Ads, Meta Ads, and Google Analytics APIs
  • Multi-agent architecture — specialized agents for each data domain (ads, website, support) coordinated by an orchestrator agent
  • Automated remediation — moving from "recommend actions" to "execute actions" with human-in-the-loop approval via Elastic Workflows
  • Alerting pipeline — proactive anomaly detection using Elasticsearch Watcher that triggers the agent before a human notices

Built With

Share this project:

Updates