A Cheaper Way to Train AI Ad Agents — and Plug Privacy Leaks

AI "agents" are programs that don't just answer questions — they take actions. In advertising, that means an AI that can look up sales data, call internal tools, write code, and adjust ad campaigns on its own. Training one of these is tricky, and a new paper accepted at the EMNLP 2026 Industry Track lays out a surprisingly practical fix. The work comes from researchers at Amazon and partner institutions, studying AI ad agents on GPT-OSS 120B, a large open model.

There are two common ways to train such an AI. The first is showing it thousands of examples of good behavior, like handing a new hire a training manual. The second is reinforcement learning — letting the AI try things and rewarding what works, like a sales rep learning from commission checks. Most teams apply the second method to everything. The paper found that's a mistake: it can scramble skills the AI had already learned well. Instead, they sorted tasks into three buckets — ones where examples were enough, ones where practice helped, and ones where practice was essential — and trained accordingly.

The results were concrete. In 18 follow-up tests, their simple sorting rule predicted the outcome correctly 15 times. Targeted training beat a top-tier competitor on 7 of 8 advertiser skills. The biggest jump was in "non-disclosure" — not revealing information it shouldn't — up 11.27 points. Leaks of sensitive information fell from 11.8% to 2.9%, and deliberate attempts to trick the AI into leaking dropped from 22.9% to 6.8%. All this while using 43% less extra computing power, which matters because computing power is money.

For anyone who has ever felt creeped out by an ad that knew too much, the privacy numbers are the headline. The AI stayed just as useful — 85.7% actionable versus 86.2% before — while spilling far less. In plain terms: smarter ad automation, cheaper to run, and less likely to blab.