Which guardrails should an AI agent for paid media have?
A guardrail is a hard limit the system cannot cross at any autonomy level. Three kinds cover most of the risk:
Guardrails you configure
When something breaks
Conversion feed stale for 6h (Meta CAPI). Autonomous actions paused themselves and the team was paged in Slack. The machine stops itself before it can optimize on a broken signal.
From our product demo.
Compare the ±15% limit with the change log below. Google Search-FR was pacing 31% ahead of the advertiser's weekly plan. Google paces spend against the calendar month and has no view of that plan, so it won't pull a 31% lead back on its own. Its daily budget came down from €300 to €260: the 13.3% cut sits inside the limit, and the log shows it ran under guardrail G-2. Meta Prospecting-DE's rise from €400 to €520 is +30%, twice the limit, and the log shows A. Weber as approver. On a €400 budget the guardrail allows anything between €340 and €460 without a person.
The monthly ceiling sits across Google, Meta, TikTok and Amazon Ads together, so no combination of moves on separate platforms can add up past it. Blocked campaigns stay at alert-only whatever level the rest of the account runs at.
What should the audit trail record?
Every change needs a receipt: the trigger, the reasoning, the confidence, who approved it and what happened next. Rejected proposals belong in the log too, because they show what the system wanted to do and why a person said no.
| When | Action | Outcome | Trigger & reasoning | Conf. | Reviewer | |
|---|---|---|---|---|---|---|
| 08:41 | Meta · Prospecting-DE daily €400 → €520 | Approved | CPL 22% under target for 6 days | 88% | A. Weber | Revert |
| 03:12 | Google · Search-FR daily €300 → €260 | Autonomous | Pacing 31% ahead of the weekly plan | 94% | System · guardrail G-2 | Revert |
| Yesterday | TikTok · Spark-DE rotation 3 ads → 5 ads | Pending | Frequency 6.1, CTR down 34% in 14 days | 76% | M. Rossi | Review |
| Yesterday | Meta · Prospecting-DE audience 1% LAL → 3% LAL | Rejected | Rejected: creatives not ready for colder traffic | 62% | A. Weber |
From our product demo.
The log in Nupact is queryable, exportable and permanent, with a Revert control on every row. Five months later, when someone asks what changed in March and whether it worked, you filter the log by month and campaign and read the outcomes. It also survives handovers: a new hire or a new agency inherits the reasoning behind every past change, and the system learns from the approvals and rejections it records.
If you cannot reconstruct what changed and why, do not let it change anything.
What happens when the data breaks?
An optimiser fed a broken conversion signal does damage quickly. If the Meta CAPI feed stops sending purchases, every campaign's CPA looks infinite, and a system that trusts the number cuts the budgets that are working. In Nupact, a stale feed pauses autonomous actions on its own and pages the team in Slack. In the demo, the Meta CAPI feed had sent nothing for six hours.
Ask any vendor to show this case on a test account: switch off a conversion feed and watch what the system does next.
Where do Advantage+, Performance Max and Smart+ fit?
Meta Advantage+, Google Performance Max and TikTok Smart+ are AI inside one platform. Google describes Performance Max as "a goal-based campaign type that allows you to access all Google Ads inventory" (Google Ads Help). Each optimises toward the goal you set, using that platform's own measurement, and none of them can move a euro to another platform.
A control layer sits above them. It decides how much each one gets, and it judges each one on tested returns as well as attributed ones. The example on our incrementality page shows how far the two can sit apart:
| Campaign | Your attribution, return per euro | Tested | Against €2.40 break-even |
|---|---|---|---|
| Meta Retargeting-DE | €7.60 | €1.90 ±0.40 (audience holdout, March) | Below, even at the top of the range |
| Advantage+ Sales-DE | €2.10 | €3.10 ±0.50 (Meta lift study, May) | Above, even at the bottom of the range |
From our product demo.
Your attribution credits Meta Retargeting-DE with €7.60 per euro, more than three times the €2.10 it gives Advantage+ Sales-DE. The tests reverse the order. Retargeting-DE returned €1.90 in a March holdout that Nupact ran on the advertiser's own customer list. Advantage+ Sales-DE returned €3.10 in a May lift study, which Meta ran and Nupact checked against the advertiser's revenue.
Agentic advertising is moving fast on the platform side too. Google's official Google Ads MCP server lets AI assistants query account data, and its documentation says the current release is "strictly read-only" (Google Ads API documentation, updated 30 September 2026). TikTok launched Symphony Agent at Cannes Lions in June 2026 for ad creation and creator discovery (ANA). As these agents gain write access, each one needs the four controls described above.
Plain Claude on an ad account works like a smart junior hire with no memory. It remembers that you like short answers. It forgets why you killed a creative three weeks ago and the budget rule you agreed.
What should you ask an AI media buyer vendor before it touches your budgets?
These questions work for any AI agent for Google Ads or Meta, Nupact included. Ask to see each answer on your own accounts.
- Can I set autonomy per campaign and per action type? An account-wide on/off switch forces you to trust everything or nothing, when you probably want Brand-Search locked while pacing corrections run on their own.
- Which guardrails are hard limits? Ask for the maximum daily change, the spend ceiling across platforms and the list of blocked campaigns, and whether any autonomy level can override them.
- Show me the log for one change. It should name the trigger, the reasoning, the confidence, the approver and the outcome, with a revert control. Ask whether rejected proposals are kept.
- What happens when my conversion feed goes quiet? The right answer is that automation stops and someone is paged.
- Which platforms can it change, and which does it only read? A clear split is a good sign. Nupact acts on Google Ads, Meta, TikTok and Amazon Ads and monitors The Trade Desk, DV360 and the other connected platforms.
- Whose numbers does it optimise on? Attributed returns and tested returns can point the same budget in opposite directions, as the table above shows. Ask which one drives the budget moves.
- Where is my data processed? For European advertisers, ask about EU hosting, GDPR and the EU AI Act.
- How do I start without risk? A good vendor starts you on alert-only and earns the next level with proposals you would have approved.
A performance marketing director told me in August: "We spent two months building. Now we double-check every important output." The teams I speak with now ask for four things first: access management, a change log, memory the whole team shares, and campaign changes without opening the platform UI.
What else do people ask about AI agents for paid media?
What is an AI agent for paid media?
Software that reads your ad accounts, decides on changes such as budget moves, bid changes or creative rotation, and proposes them or applies them through the platforms' APIs. The term covers everything from a chat assistant to a fully autonomous buyer, so ask what it may change without approval.
Can I use ChatGPT or Claude as an AI agent for Google Ads?
For analysis, yes. Google's official Google Ads MCP server lets an assistant query your account, and it is read-only in its current release. To make changes you would connect write access yourself, and the guardrails, the change log and the stop on broken data would then be yours to build.
Which AI agents work across Google, Meta, TikTok and Amazon Ads?
Ask each vendor which platforms it can change and which it only reads. Nupact acts on Google Ads, Meta, TikTok and Amazon Ads and proposes budget moves between them, each with a forecast. It monitors The Trade Desk, DV360, LinkedIn Ads, Microsoft Ads and the other connected platforms.
Does every change need my approval?
Only the ones you choose. At Level 2 every change waits for a click in Slack. At Level 3 changes inside your guardrails apply on their own and appear in the log with a revert control, while anything outside them still waits for a person.
Will budget changes reset Meta's learning phase or unsettle Smart Bidding?
On Meta they can. Meta's Help Centre says frequent budget changes can send an ad set back into the learning phase, which usually ends after about 50 results in the week after the last significant edit, and that ad sets in learning are less stable and usually have a higher CPA (About the learning phase). Google is different. It names a new bid strategy, a changed bid strategy setting and campaigns or keywords added to or removed from it as the reasons Smart Bidding shows "Learning" (Google Ads Help). In some rare cases a large change in targets can bring a conversion cycle or two of volatility (How our bidding algorithms learn). In Nupact a guardrail caps how far a daily budget moves without approval, ±15% a day in the example above, so routine corrections stay small. A bigger move, like Meta Prospecting-DE's +30%, waits for a person, and every change can be reverted from the log. Autonomy is set per campaign and per action type, so while a Meta campaign's ad sets are still in learning you can keep its budget changes on propose-and-approve, and each one goes live only after a click.
Is Nupact aligned with GDPR and the EU AI Act?
Nupact is built and hosted in the EU, aligned with GDPR and the EU AI Act, and every action is audited. Nupact GmbH is based in Berlin.
Bring the account you would least want an AI to touch.
On the call we set out the guardrails you would put on it and which action types would stay on approval.