Can AI run my Facebook ads? A real company's 30-day public test

Most "AI runs your ads" claims come with a demo account. We wanted a real one. On 1 October 2026 we handed Upcar, a real peer-to-peer car-rental marketplace (run by Zavi's founder, so the money and the results are real), to Zavi, our AI growth agent, with three rules:

• $25,000 of real ad money for 30 days (1 to 30 October).
• Hands off: we do not pause or override it. If a person steps in, it goes in the public log.
• It can only place ads and move budget on the accounts Upcar has connected. Elsewhere it researches and drafts.

The scoreboard and a daily log are public at zaviagent.com/live/upcar, with one page for each published day.

What the agent actually does each day

It reads every campaign and ad set, decides what to change, and writes down why in plain words. Some days it changes nothing and says so. Examples from its own log:

• Day 2: changed seven budgets across Google and Meta. (day 2)
• Day 7: raised a Meta budget from $40 to $80 a day. (day 7)
• Day 9: changed three budgets across Google and Meta, and had a further raise refused because the month's spend limit did not have room for it. (day 9)
• Day 11: spotted one ad that brought in paying renters for less and spent its whole budget every day, and asked to back it with the last room left under the monthly limit. (day 11)

The numbers so far

As of day 12 (12 October 2026, 18:23 UTC). The live page is the number of record and keeps moving:

Two honest caveats. "Normal pace" is Upcar's average over the four weeks before the test, not a controlled experiment, so part of the gap could be season or luck. And both the booking total and the normal pace include the same estimate (×1.4) for bookings Upcar's data feed misses, so they compare like for like, but neither is a raw count.

What it got wrong or could not do

The log keeps the misses too: proposed changes that failed when applied, a raise refused by the spend limit, days where it only researched. That is the point of doing this in public: an agent's bad days tell you more than its good ones.

Should an AI run your Facebook ads?

It depends less on the AI than on the guardrails around it:

1. Approval first. Start with the agent proposing changes and you approving them. Widen what it does alone only after you have watched it for a while.
2. A real spend limit. The limit should be checked by the software before a budget change goes out, not just written into the AI's instructions.
3. A written reason for every change, in words you can check against the ad account.
4. Your own normal pace to compare against, so a good season is not mistaken for a good agent.

By default, Zavi asks before it spends. The Upcar test is the exception we chose on purpose, inside a hard $25,000 limit, to show what it does when nobody is steering.