What I Check Before Letting an AI Agent Touch Live Ad Spend
Before any AI agent touches live ad spend, it gets a spend cap, a change log, and a human who checks it daily for the first two weeks. No exceptions. I don't care how good the demo looked.
The short answer
Daily review for the first 14 days, then weekly. That's the cadence. The reason is simple. Agents don't fail loudly. They fail by quietly scaling a bad ad set at 2am because the early CTR looked good. A human checking weekly will catch a pattern. A human checking monthly will catch a disaster after it already cost five figures.
What changes the timing
Not every account needs the same level of hovering. A few things move my review schedule earlier or looser.
- Budget size. On a $2.2B infrastructure project I marketed, a $10M media budget meant a bad week was real money. Anything over $50K a month in spend gets daily checks, period.
- Account maturity. A brand-new account with no conversion history is noisy. Agents misread noise as signal. I give new accounts 30 days of manual-only decisions before an agent touches bids.
- How much write access it has. Read-only reporting agents get checked weekly. Agents with bid, budget, or audience edit access get checked daily, no matter the budget size.
- Seasonality. Black Friday week, a product launch, anything with a hard deadline, I pull agents back to recommend-only mode. Automation and high-stakes windows don't mix well.
The guardrails I actually use
This is the part people skip because it's less exciting than "the agent optimized my ROAS." Here's what sits between the agent and the live account.
- Hard spend caps per campaign. Not account-level. Per campaign. I set it at 20% above the trailing 7-day average daily spend. If the agent wants to go past that, it needs approval.
- No same-day budget increases over 25%. A DTC brand I scaled from $100K to $3M in revenue taught me this the hard way. A 40% budget jump on a Tuesday because Monday's ROAS looked great is how you torch a week's margin on Wednesday.
- Change log with a diff, not a summary. I want to see exactly what changed: old bid, new bid, old audience, new audience. Summaries hide mistakes. Diffs don't.
- Kill switch tested monthly. If I can't pause every campaign the agent touches in under 60 seconds, the setup isn't ready for live spend.
- Attribution window locked before launch. Agents will optimize toward whatever window you give them. If it's set to 7-day click when your real sales cycle is 21 days, the agent will happily scale the wrong ads.
- A second account, not the main one, for the first 30 days. Run the agent on a duplicate campaign with 10% of real budget before it gets near the account that pays the bills.
Signs you are overdue for tighter controls
- You can't explain why a bid changed without opening three dashboards.
- CPA swung more than 30% week over week and nobody flagged it until the invoice.
- The agent made more than 10 changes in a day and you reviewed none of them live.
- You're trusting a single metric (usually ROAS) and not checking it against revenue or margin.
- Nobody on the team can tell you the current spend cap off the top of their head.
The most common mistake
People treat the approval step as a one-time setup. Turn it on, write a nice prompt, walk away. That's backwards. The approval process should get looser over time as the agent earns trust on your specific account, not stay static or get skipped entirely after week one. I've seen teams spend three weeks building a tight review process, then stop checking in month two because nothing bad happened yet. Nothing bad happening isn't the same as nothing bad being possible. Platforms change their auction dynamics. Competitors enter. The agent that was safe in March can start making bad calls in June because the whole landscape shifted under it.
What happens if you wait too long
I've watched an account burn 18% of a monthly budget in four days because an agent kept scaling a campaign that had a tracking bug inflating its conversion numbers. Nobody caught it because the weekly review had slipped to "whenever we get to it." That's roughly $9,000 gone on a $50K monthly budget, for zero real return. The bigger cost isn't even the wasted spend. It's trust. Once a founder or a client gets burned by an agent with too much rope, they pull automation out of the whole stack, including the parts that were working fine. You lose months of progress because one guardrail was missing on one account. Spend caps and daily checks feel like friction when everything's going well. They're not friction. They're the reason you get to keep using the agent at all.
Practical takeaway: before you give any agent write access to real ad spend, write down three numbers on a sticky note: the daily spend cap, the max single-day budget increase, and who gets pinged if either gets hit. If you can't write those three numbers down right now, the agent isn't ready for live spend yet. The setup is.