Marketers put up guardrails as AI agents reshape programmatic buying

"“Would it blow a quarter's worth of budget in a weekend? All of that's well-founded, so that's a fear,” said Henry Webster, svp director of analytics and insight at Kelly Scott Madison (KSM) Media, who spoke onstage during Digiday's Programmatic Marketing Summit (DPMS), May 6-8 in Palm Springs, Calif. “The process of taking baby steps, testing, putting stringent guardrails and checks on agents that would operate in that way - makes a ton of sense.”"

"KSM has an internal agent it calls “the librarian” that acts as a gatekeeper for each client's brand voice. Other agents can ask it for things like client-specific acronyms and campaign names to be more precise with context or targeting."

"As a pharmaceutical company, Bayer has even more tedious guardrails. The brand has spend caps to prevent bots from over using legacy data partners, which would throttle testing with new partners, according to Glenniss Richards, senior director of digital media activation at Bayer. She did not offer a dollar amount for the spend caps."

"“We do put guardrails from a spending perspective to ensure it doesn't conflict with our decisioning or how we want the campaigns to perform,” Richards said, also speaking at DPMS. “There's also guardrails around data anonymization and de-identification before any activations,” Richards added."

AI agents are positioned as autonomous decision makers that can automate campaign creation, optimize ad buys, and draft pitch decks. Marketers are not fully trusting these systems for ad buying because hallucinations can produce incorrect metrics such as CPMs, risking major budget loss. Teams respond by adding guardrails, checks, and incremental testing so agents operate within controlled boundaries. One agency uses an internal “librarian” agent as a gatekeeper to enforce client brand voice and provide context like acronyms and campaign names. A pharmaceutical company adds spend caps to prevent excessive use of legacy data partners, along with guardrails for data anonymization and de-identification before activations.

Read at Digiday