Eric Picard: agentic ad platforms need deterministic guardrails around real budget spend

August 18, 2026 · AdExchanger

Ask a large language model the same question five times and you can get five different answers, even with identical context and instructions, Eric Picard writes in AdExchanger's Data-Driven Thinking column. That variability makes LLMs useful for drafting, summarizing and open-ended reasoning, and it makes them a poor fit for any task where a wrong call spends real money, he argues. Every major ad platform runs on operational rhythms an API document never spells out: how audiences get structured, how budgets pace, how automated bidding wants to be fed. "The difference between legal and wise is where campaigns either perform or waste money," Picard writes, adding that buyers learn it only after years and billions of dollars of real spend.

Campaigns running at high volume leave no room for a human to check every change, and Picard is equally clear that automation should not make a decision nobody approved. His fix is to set the boundaries before any code runs: which actions are allowed, which need a human sign-off, which execute automatically, and then let a deterministic system operate inside those lines. The LLM's job shifts to the interface, turning a buyer's stated goals into instructions the platform can act on. Budget-touching execution stays on deterministic, rule-governed infrastructure built over years of live spend, so agencies and brands get intent translated into action without an agent stepping outside the limits they actually approved.