I Let an AI Agent Run Live Ad Ops for 30 Days

I let an AI agent run live ad ops for 30 days. Here's what broke - and the one save no human would've caught.

Everyone's posting the 5-minute demo right now. Agent checks creatives every 15 minutes, scales winners, pauses losers while you sleep. Looks like magic.

The demo is easy. The part nobody films is week 3, when the loop has enough rope to do real damage.

Here's what I actually ran. A Claude loop wired into the Meta and Google APIs, watching a live Shopify account in the mid-five-figures a month. Read performance, propose changes, execute inside guardrails. I built it because I wanted to find where it breaks before a client account did.

It failed in four places. Same root cause every time: the agent optimizes perfectly against the signal it can see, and it can't see the context that lives in my head.

1. It tried to pause a "losing" campaign that was actually our retargeting floor. Low ROAS on platform, high incrementality in reality. The agent doesn't know which conversions would've happened anyway.

2. It scaled a winner 80% in a day off three days of data. The math looked great. The variance was noise. A human waits for the sample. The agent wanted to act now.

3. It kept proposing budget shifts into branded search dressed up as "high ROAS." The same cannibalization trap I strip out of every account by hand.

4. It had no concept of margin. A 4x ROAS SKU at 10% margin loses money after COGS and fees. The platform doesn't pass margin, so the agent never saw it.

So I added four guardrails. Daily budget change capped at 20%. No pause without a 7-day window and an incrementality flag. Branded search locked out of the decision set. The margin table passed in as a hard constraint, not a suggestion.

After that, it earned its keep. Late one night it caught a tracking break I'd have missed. Conversions flatlined on one campaign at 2am, it flagged the anomaly and paused spend before the day's budget leaked into nothing. I was asleep. Someone reviewing in the morning eats the full loss.

That's the actual product. An agent that runs 24/7 inside constraints a senior operator wrote, and escalates the calls it can't judge.

The people winning with this in 2027 won't be the ones who automate the most. They'll be the ones who know exactly which four levers to never hand over. I run this across the brands at Tegra because I'm technical enough to build the loop and burned enough not to trust it raw.