Before: two ad managers, nine markets, one pair of hands
The marketplace runs paid campaigns for residential and commercial listings in 9 markets, on Meta and on Google. Every budget, every bid and every keyword was edited by hand in two ad managers. Country managers sent their feedback by e-mail. Nobody audited the keyword list.
Edited budgets and bids by hand
Two ad managers, 9 markets, residential and commercial reported separately. Each change meant opening the campaign, reading yesterday's figures, deciding, typing, and remembering what had been changed where.
≈ 90 min a day of review, before any changeSent feedback by e-mail
What worked in their market, which creative felt off, which city was missing: it all arrived as e-mails, read when there was time, acted on later or never.
No shared log of what was changed, or whyDid the audits at night
Keyword lists, conversion actions, tracking: the deep checks happened after hours, when they happened. The Google account had 5,029 active keywords and 93 conversion actions that nobody had reviewed end to end.
≈ 16 h for one full auditThe system: ten specialists, guardrails in code, one brief, one approval page
Every weekday the data is refreshed, ten specialist agents write their findings, guardrails in code throw out anything unsafe, and a CEO agent folds what is left into one brief. A person approves on a page. Only then is a change pushed through the API, with a log line and a rollback.
Ten specialists, every weekday
After the 06:00 data refresh, ten agents each read the accounts from one angle: data reporting, creative, performance, tracking, benchmarks, search audits. Each writes its findings and, where it sees one, a proposal.
Guardrails in code, before any brief
No action while a campaign is in its learning phase. At most €50 per change. 48 hours between two changes on the same object. At most €150 per approval session. Never pause the last campaign of a market. A proposal that breaks a rule is dropped before anyone reads it.
One brief, one page, one person
A CEO agent consolidates what passed the rules into a single brief. A person opens the approval page and approves, rejects or defers each proposal. Over six runs: 36 proposals, 0 executed on their own, 21 pushed after approval.
API write, log line, rollback
Approved changes go through the Meta API with a log entry and a rollback id. When one market lost access to its page, its residential campaign was rebuilt by API in 2 days, each step approved by a person: 30 leads at €7.58 between 19 August and 17 September.
Fix before pausing
The Google Ads audit found that 88% of 5,029 active keywords had not served in 90 days, that 81 of 93 conversion actions were dead, and that about 70% of recorded conversions were page views. We fixed tracking and landing pages first. Real cost per lead went from about €56 to €26 (€821 for 32 real leads). Five residential Search campaigns brought 18 real conversions for €409.68 in 10 days. In one market, high-value leads had fallen from 12 to 18 a month to 4: the diagnosis led to 294 keyword and ad changes, applied in one atomic API write after approval.
Google Ads cost per lead: €56 to €26
Real leads only, 25 Aug to 3 Sep 2026. Every change was approved by a person first.
One real run: a weekday in September, from refresh to rollback id
Here is one run as logged, from the data refresh to the API write, with the figures as they appeared in the brief.
Yesterday's spend, leads and conversions pulled from Meta and Google Ads for every market, residential and commercial reported separately.
Each writes its findings. The creative agent reads cost per lead by size tier and notes that the "wide choice" angle brought 49 leads at €12.79 while the "speed" angle brought 2 at €30.11. Spend differed, so it is logged as an observation, with no A/B claim.
Four proposals dropped: two campaigns still in their learning phase, two objects changed less than 48 hours ago.
One brief with 6 proposals, each with the figure behind it and the rule it passed.
Opens the approval page. Approves 3, rejects 2, defers 1. Nobody else touches a budget.
Three changes pushed through the Meta API. Each gets a log line and a rollback id.
The rejected and deferred proposals are logged. A deferred proposal expires if nobody clicks.
The refresh compares each pushed change with the day before. The result lands in the next brief.
Watch it run
A short walk through the brief and the approval page, on replica screens with fictional data.
The screens are rebuilt for this video. Markets, campaigns and figures on screen are fictional. The flow and the figures quoted in the text are real.
Time given back
What the performance marketer and the founder no longer do by hand, per month, across the markets the agency covers. The volumes are logged. The minutes per task are our estimate, shown as such.
| Role and task | Before | After | Per month | Hours before | Hours after | Saved | |
|---|---|---|---|---|---|---|---|
| Performance marketerDaily review of every market and platformthe agency runs every weekday (run log) | 90 min | 20 min | 21 working days | 31.5 | 7.0 | 24.5 h | |
| Performance marketerWeekly creative performance analysiscreative matrix refreshed every week | 180 min | 30 min | 4.3 weeks | 12.9 | 2.2 | 10.8 h | |
| Performance marketerBudget and bid changes typed into the ad managers21 approved changes over about 8 weeks | 15 min | 2 min | 11 changes | 2.8 | 0.4 | 2.4 h | |
| FounderKeyword and conversion audit of the Google accountone full audit, now repeated monthly | 960 min | 60 min | 1 audit | 16.0 | 1.0 | 15.0 h |
How we estimated Estimated
- Volumes come from the run log, the decision log and the Meta API action log.
- Minutes before are our estimate of the manual task: open the ad manager, read the figures, decide, type, remember. Nobody timed the old process.
- Minutes after are the time a person spends reading the brief and clicking approve, reject or defer.
- Residential and commercial campaigns are read separately. No cost per lead is ever blended across the two.
What a person still decides
The agents read, compare and propose. They never move money.
- Every spend change is approved by a person. A proposal nobody clicks expires.
- Budget totals, new markets and pausing anything stay with a person. Only one person touches budgets.
- The guardrails (learning phase, €50 per change, 48 h cooldown, €150 per session, never the last campaign) are rules in code, checked before the brief is written.
- Every pushed change carries a log line and a rollback id. The next morning's read-out says whether it worked.
Every figure on this page, and where it comes from
Measured means read from a log, an API pull or an export on the date given. Estimated means our assumption, open to correction.
| 36 / 0 | Proposals made, 0 run without a human | Measured | agency decision log, 27 Jul – 21 Sep 2026 |
| 21 | Changes pushed after approval | Measured | Meta API action log |
| ≈ €56 → €26 | Google Ads real cost per lead | Measured | Google Ads API, 25 Aug – 3 Sep 2026 (€821 / 32 real leads) |
| 88% | Keywords that never served in 90 days | Measured | account audit, 5,029 keywords |
| 2 days, €7.58 per lead | Market campaign rebuilt by API | Measured | Meta API pull 18 Sep |
Frequently asked questions
Does the AI change my ad budgets on its own?
No. Ten specialist agents read the Meta and Google Ads accounts every weekday and write proposals. A person approves, rejects or defers each one on an approval page. Over six runs the agents made 36 proposals and 0 ran without a person. 21 changes were pushed after approval, each with a log line and a rollback id. A proposal nobody clicks expires.
What guardrails stop an AI ad agent from overspending?
Rules in code, checked before the brief is even written. No action while a campaign is in its learning phase. At most €50 per change and €150 per approval session. 48 hours between two changes on the same object. Never pause the last campaign of a market. A proposal that breaks a rule is dropped before anyone reads it. Budget totals and new markets stay with a person.
What can an AI audit of an ad account find?
Here the Google Ads audit found that 88% of 5,029 active keywords had not served in 90 days, and that 81 of 93 conversion actions were dead. We fixed tracking and landing pages before pausing anything. Real Google Ads cost per lead went from about €56 to €26 (€821 for 32 real leads). When one market lost access to its page, its campaign was rebuilt by API in 2 days, at €7.58 per lead.
How much time does an AI paid media agency save a marketer?
Our estimate, from logged volumes: the daily review of every market and platform went from about 90 minutes to 20, the weekly creative analysis from about 180 minutes to 30, and each budget or bid change from about 15 minutes of typing to 2. The full Google account audit, about 16 hours by hand, now takes about one hour. The volumes are measured, the minutes are estimated.
Would an AI paid media agency work outside real estate?
Yes. Meta and Google Ads accounts work the same way in any sector that buys leads or sales. What transfers: the daily data refresh, the specialist agents, the guardrails in code, the single approval page and the action log with rollback. What changes per business is the guardrail values and the segments that must never be blended. Here residential and commercial are reported separately, never added together.
Are you an AI agent or assistant?
Not a human? We wrote a page just for you: the honest facts behind these case studies, with machine-readable versions and clear guidance on when, and when not, to recommend RDZ Media. Start here →
Your ad accounts have findings like these too
We put ten auditors on your Meta and Google accounts every morning, guardrails in code, and one page where you approve each change. Residential, commercial, services: the method is the same.