From 700 Trillion Signals to Action: Building Production AI Agents for Real-Time Media Planning

The goal is not another generic planner, but a defensible, repeatable system that translates hundreds of trillions of consumer signals into strategy, powered by MIQ's data partnerships and an agentic workflow on the Databricks Platform. Organizations evaluating similar approaches can find a structured framework in this enterprise analytics platform evaluation guide that covers the key criteria separating production-ready systems from demos.

The session unpacks why a bespoke agent is necessary, how the team scoped the first release to the pitch stage where inefficiency is greatest, and what it takes to make enterprise data reliable for automated reasoning, SQL generation, and audience insights at scale.

Why media planning needed a purpose-built agent

Generic LLM plans are fast but shallow, inconsistent between sessions, and not anchored in the data that clients expect. Agencies face slow, costly pitches that often depend on individual planner experience. MIQ framed the problem as three linked gaps: uneven planner knowledge, fragmented data and taxonomies, and weak actionability from otherwise strong insights.

The agent targets pitches first, where turnaround is slowest and the payoff is highest. Recommendations are grounded in MIQ's connected signal graph, spanning viewing, browsing, and transactional data, and augmented by trusted external intelligence. Outputs are therefore current, variable by prompt, and defensible to client scrutiny.

From brief to plan: decomposition, tools, and human checkpoints

A ReAct loop sits at the core: the agent observes the prompt, reasons about intent, then selects tools to answer subproblems. Requests are categorized as direct, focused, moderate, or comprehensive before any database call. The agent decomposes a brief into subatomic tasks, proposes a plan of attack, and pauses for a human-in-the-loop review so planners can approve, edit, or add constraints such as retailer splits or exclusions.

Execution then generates a structured response, including an executive summary, tactical persona variants, key takeaways, and activation recommendations, plus interactive breakdowns of browsing behavior, demographics, and temporal habits. Because results are computed against live signals, end-to-end runs take roughly 10 to 15 minutes, a conscious tradeoff for depth and auditability.

Making enterprise data usable for agents

Data that works for analysts often confuses agents. MIQ enriched Unity Catalog metadata with detailed column descriptions, tool capabilities, and example queries so Genie can pick the right tables and write correct SQL. Extensive instructions steer SQL generation, including when to use percentages or indexes, while common definitions and joins live in SQL templates to keep within Genie's roughly 20,000-word instruction limit and reserve space for reasoning guidance.

The agent integrates four tool types: MIQ's historical campaign performance, consumer watching and browsing signals, external panel-driven marketing intelligence via MCP, and a persona builder that synthesizes insights into targetable audience variants. A persistent memory preserves context across sessions, and a curation layer stores local overrides so planners retain control over what is recommended and what is excluded.

Quality control, presentation, and observability

Quality is enforced in loop, not after the fact. A critique layer evaluates statistical significance, checks for bias, validates SQL logic with a secondary evaluator, filters empty or out-of-context outputs, and can trigger targeted retries. Results are then summarized in planner-friendly language with appropriate visualizations and per-pillar summaries, concluding with concrete activation guidance.

Observability was designed in from day one. Every node call, input, output, and tool selection is traceable across development, integration, and production. Golden prompts benchmark each change across pillars such as executive summary quality, insight depth, recommendation quality, coverage, and structural coherence. CI/CD gates prevent merges that regress performance, making behavior explainable and stable as the system evolves.

Production-ready agents are built on data, not prompts

The central lesson is that agent quality depends as much on data readiness and evaluation as on model choice. Grounding in verified signals, transparent reasoning with human control, in-loop critique, and rigorous observability separate demoable agents from tools planners can confidently pitch behind.