For scaling brands that are able to run at least $10k+/month in paid media.
The human-led, AI-multiplied ad system, stage by stage.
"We'll just use AI for our ads" is the new "my nephew does our website." Here is the operating loop that actually works: six stages, who owns each one, what gets produced, and the decision gates in between. Copy it.
The order is the whole system. Direction first, multiplication second.
Most brands run it inverted. They automate the judgment and hand-craft the production. The judgment was the only part that needed a human.
Below is the loop as it runs on a real account. Each stage lists the owner, the concrete output, and the gate that must pass before the next stage starts. The gates are the system. Skip them and you have activity, not strategy.
Research before spend. Human directsAI reads
Before a single ad runs: customer language from your reviews, competitor complaints, support logs, and community threads. AI does the reading, an hour instead of a week. A human sets the extraction targets: first-person pain, trigger moments, objections, identity phrases, failed alternatives.
OutputAn evidence base: clustered verbatim quotes, each cluster written as one belief statement in the form "this customer buys when they believe X."
GateNo production starts until the belief statements exist in writing. Prompting a copy model with a product description instead is how accounts fill up with ads that contain zero customer psychology.
The angle decision. Human only
The research yields a dozen candidate angles. One gets chosen for the next test. The choice weighs quote frequency, competitor gaps, and account history: which belief the market voices most, which nobody is selling against, which already died here and why.
OutputA one-page decision note: the chosen belief, the evidence behind it, what would prove it wrong. Written down, because in three months nobody remembers why.
GateOne new belief enters testing at a time, so results stay readable. But the account as a whole should carry several distinct, proven beliefs concurrently: Meta's current delivery system rewards genuine conceptual diversity and collapses near-duplicate ads into one entity. Diversity across the account is fuel. Seven half-formed value props inside one test is noise.
Production at volume. AI multipliesHuman rejects
Now, and only now, AI earns its keep. One belief becomes a batch: multiple hooks on the same angle, statics and video variants, platform cuts. What used to take a production week takes an afternoon.
OutputA named, tracked batch. Every creative logged with its angle, hook, and hypothesis in one tracker (a simple Airtable does it), so results attach to beliefs, not to file names.
GateA human reviews every asset against the research and rejects the confidently mediocre ones. AI iterates winners. Humans originate angles and kill weak output. Flip that and volume goes up while learning goes to zero.
Launch with spend floors. Human sets rules
Every ad gets a minimum spend before any verdict, sized against your target CPA, because an ad killed at pocket-change spend has produced activity, not evidence. The floor is decided before launch, in writing, so nobody kills a slow starter out of boredom or scales a lucky spike out of excitement.
OutputLaunch rules: spend floor per ad, kill criteria, scale criteria. Three lines. Most accounts have none.
GateNo mid-flight rule changes. Changing verdict criteria after seeing early results is how learning data gets polluted account-wide.
The weekly read. Human decidesAI compiles
Once a week, not once an hour. Automation pulls the numbers into one view; a human makes three kinds of calls: scale what cleared the floor and the CAC ceiling, kill what failed with the floor spent, iterate what shows a strong hook but weak close, or the reverse.
OutputA decision trail: what was scaled, killed, iterated, and why, one line each. This is the account's memory. It is also the answer to "which angle is working and why" in one sentence, which most accounts cannot produce.
GateVerdicts reference the CAC ceiling, not last month and not blended ROAS. An ad can look profitable and still be building the wrong customer base.
The angle pipeline. Human owns
While the current angle prints, the next one is already in research. The target state is a rotation: one proven angle live, one challenger in testing, one researched candidate on deck. Then fatigue is a scheduled hand-off instead of a crisis rebuild under pressure.
OutputBack to Stage 1 with new data: the winners' performance feeds the next research pass. The loop closes.
GateThe second angle starts development while the first still works. After the plateau is too late, and fatigue windows have compressed to weeks, not months. The rotation is not optional anymore.
The failure pattern in fully automated accounts is always the same stage: not production, not launch. Stages 2 and 5. The decisions. Hundreds of launches, no spend floors, no decision trail, and a founder concluding "Meta doesn't work." Meta works. Unowned decisions don't.
Run the loop check on your account.
- Do written belief statements exist for your live ads, with quotes behind them?
- Can anyone name the beliefs currently live, and the one being tested?
- Does every creative trace to an angle and a hypothesis in a tracker?
- Are spend floors and kill criteria written before launch?
- Is there a one-line decision trail from last week?
- Is the next angle in research while the current one still works?
Six questions, six stages. The ones you answered "no" to are your system gaps, and they are all fixable in-house. This is the loop we run for clients; we publish it because the constraint was never the secret, it was the discipline.
The loop assumes the economics work. Check that first.
Step 1: run your numbers through the calculator. Two minutes for your aMER, your CAC ceiling, and the monthly leak. Stage 5's verdicts depend on that ceiling.
Step 2: want the loop installed and run on your account? Request the audit and we start at Stage 1.
Run your numbers Request the full audit