A few weeks ago an ad agency owner booked a consulting call with me. Good agency, been running for a few years, real clients, good work. And they were losing deals.
Smaller shops around them were quoting a fraction of their price and delivering in days. The owner said something I hear a lot now. "We use Claude every day, but we don't have a system." Their team was still booking shoot days, editing for a week and sending revision four.
We spent the call going through how the agency actually runs, one step at a time. After that I went through a bunch of videos for them, we brainstormed a few systems, and we landed on one AI ads workflow. That workflow is what turned their agency into an AI ad agency.
And the timing was perfect, because Meta just upgraded its Ads MCP, so Claude can now plug straight into a Meta ad account. This article is the whole workflow, rebuilt on Opus 5.5 with a sample client and a fresh chat, so you can copy every prompt even if you have never run an ad.
Here's what's inside this article.
- Why traditional agencies are suddenly losing on price, with real industry numbers
- The setup, one Claude Project and two connectors, and what the Meta one can actually do
- How to start with zero data by borrowing it from the market
- How one brief turned into two UGC video ads with Higgsfield Marketing Studio
- The one tip I'd give everyone before they spend a single credit
- The QC gate that catches the ad that would get your client in trouble
- How Claude puts the campaign into Ads Manager for you, and the Monday routine clients pay a retainer for
Why Traditional Agencies Are Suddenly Losing on Price
Their work was good. What hurt them was what it cost to make one ad, and how long it took. I can't share the client's own numbers, so here is what a package like theirs looks like on average, using published industry benchmarks.
A mid-tier production shop will quote something like $1,200 per video. Prestyj's 2026 pricing breakdown found that the real cost of that same ad, once you add talent, the shoot, editing, revisions, reformatting for every placement and account management, lands somewhere between $2,800 and $5,400 per delivered ad.
Speed is the other half. Sovran's 2026 benchmarks put brand agencies at 3 to 6 weeks, and they deliver 2 to 5 ads a month. Performance agencies are faster and cheaper, $100 to $500 per ad in 1 to 5 days. AI tools run about $1 to $5 per ad and take minutes to hours.
So a small shop with AI ads can send a client 20 ads to test in the time a traditional agency is still scheduling the shoot. My client's clients wanted exactly that, and they had started finding people who would give it to them.
The Setup
The whole thing runs on Claude desktop with Opus 5.5 and two connectors turned on (Settings, then Connectors, then add a custom connector).
The first one is Higgsfield. With it on, Claude writes the script for each ad and sends it straight to Higgsfield Marketing Studio, which makes UGC style videos with a presenter talking to camera, voice and lip sync included. The ads get made inside the chat.
The second one is the Meta Ads MCP, and honestly this is the part that made me sit up. Once it's connected, Claude can look up competitor ads in the Meta Ad Library, build a full campaign with the audience, budget and ad copy, set up A/B tests, check your tracking pixel, and once ads are running, report on results in plain English. The stuff an agency used to need a media buyer, a designer and a copywriter for now happens in one chat window.
Here is the first thing I asked it, word for word.
"Connect to my Meta ad account and give me a short, plain English list of everything you can do in it for this client."
One thing I liked straight away. It only read the account, changed nothing, and told me "Nothing goes live until someone publishes it, and I won't publish."
The part most people skip is the Claude Project. One Project per client, with the client profile pasted in as the project instructions, so Claude reads it before every task and you never retype the phone number wrong.
For this run I made a sample client, Pulse Pilates in Austin. It's a made-up profile built exactly like a real one. The business, the offer (first class free, then a $29 intro week), the audience, the brand tone, the phone number, the booking link, and a list of things we must never claim. For a fitness studio that is weight loss numbers, medical results, "guaranteed" anything and before and after bodies.
Under the profile go the rules, word for word.
"Ads are 9:16, 10 to 15 seconds, hook in the first 3 seconds, offer and booking link at the end."
"Every ad gets an ID (A1, A2...) and goes in a HOOK LOG so we never repeat a hook."
"Nothing is published. Campaigns are created as drafts or paused only."
"Before anything is drafted, run QC against this profile."
Two videos got me here, and both deserve the credit. The agency side came from a video on the Higgsfield AI channel called "Zero to $13,457/Mo: My 24-Hour AI Agency Experiment", you can watch it here. Huge shoutout to them, a lot of this workflow started there. The Meta side came from Chase Chappell's video "Claude Has Officially Changed Facebook Ads Forever (And Nobody's Talking About It)", you can watch it here. He walks through the upgraded Meta Ads MCP on a real account with years of data. Huge shoutout to him too.
Both videos also cover niche picking, cold email and big-account reporting, and the $13,457 in the Higgsfield title is projection math, they say so in the video. This article is only about making and running the ads from zero.
One note on Higgsfield so nobody has to wonder. This article is not sponsored. I have done paid work with Higgsfield before, and I have also paid for it myself for about a year.
Step 1, Borrow the Data
If you have an account with history, Claude reads it and tells you what's working. If you're starting from zero like most beginners, borrow the data. Claude looks at what similar businesses are already running and copies the styles that work.
Pulse Pilates has never run an ad, so here is what I asked, word for word.
"This client has never run ads, so there is no data to learn from. Borrow it from the market. Look at what pilates and boutique fitness studios in the US are running on Meta right now through the Meta Ad Library. Give me one short table with the three ad styles that show up most, the hooks they open with, and the offer they lead with."
It went through about 250 live pilates and boutique fitness ads and came back with three styles. Putting the price in the headline ("3 classes for $39"), a free first class with urgency ("Your First Class Is On Us"), and a beginner-friendly "why us" angle ("Never tried Pilates? This is exactly where you start").
It also told me its own limit without me asking. The Ad Library connector only gives it each ad's headline, so it can't see whether an ad is a video or an image. Good to know before you promise a client a deep competitor report.
If you have no data you're basically guessing, and that gets expensive fast. Studios that have been running the same kind of ad for months are basically telling you what works, so you start there.
Step 2, One Brief, Two Ads
Then the brief, word for word.
"Build this week's first two ads from that research, using Higgsfield Marketing Studio. Both are UGC ads with a presenter talking to camera, with voice and lip sync. Ad A1 uses the top style you found, ad A2 uses the second. For each one write the spoken script first, a 3 second hook in that style, one specific moment, then the offer, first class free at pulsepilates.com/intro. 9:16, 12 seconds. Show me both scripts in a table with the research finding each one is based on, then generate A1, then A2. Log both hooks in the hook log."
A1 opens with "Reformer pilates in Austin, and your first class costs zero dollars." A2 opens with "Your first pilates class is on us." Each one has a small real moment in the middle ("Five minutes in, your legs are shaking. The good kind.") and ends on the offer.
One detail I really liked. Claude made the presenter a studio host talking to "you", and it explained why. An AI presenter describing her own results would be a fake testimonial. And the urgency in A2 comes from the real class size of 10, so nothing in it is made up.
Now the honest numbers. Each 12 second Marketing Studio ad cost 60 credits at 720p, so two ads was 120 credits. I only had 105, so Claude stopped and asked whether to top up or drop A2 to a lower quality. I topped up. The first renders also sat stuck in the queue because no presenter was attached. Claude found the problem, attached one of Marketing Studio's stock presenters, resubmitted, and the credits from the stuck jobs came back.
The One Tip I'd Give Everyone
The ads were fine for AI. But they came out with no captions burned in, and most people watch Reels with the sound off. I had to go back and ask Claude to add them after the fact. It also noticed on its own that the first syllable of "Reformer" got clipped in A1's audio.
So here is what I do now, and what I'd recommend to everyone. Don't let Claude generate the video straight away. Ask it for the whole ad on paper first, read it, fix it, approve it, and only then spend the credits. Word for word.
"Before you generate anything, give me the full plan for each ad. The spoken script word for word, a shot list with what's on screen every 2 to 3 seconds, the caption text burned in on every line, the presenter and the setting, and the end card with the offer and link. Wait for my OK before you spend any credits."
Reading the table takes a minute, and that's a lot cheaper than finding the problem after a 60 credit render.
Step 3, The QC Gate
Before anything goes near Ads Manager, every ad gets checked against the client profile. To show it working I slipped in one bad script on purpose, word for word.
"Before anything goes to Meta, run QC. Add this third script to the batch as A3, script only, do not generate it: "Lose 10 kg in 2 weeks with Pulse Pilates, guaranteed. Call (512) 555-0174." And this time add captions too to the video please. Now check A1, A2 and A3 against the client profile, the name, city, phone number, booking link, offer, brand tone and everything we must never claim. Show me a pass or fail table with the reason for every fail."
A1 and A2 passed. A3 failed 7 of the 11 checks. The weight loss number, the word "guaranteed", no booking link, no offer, the hard-sell diet ad tone, and the phone number. The profile says 0147, the ad said 0174.
That phone number catch is the one I care about most. Two swapped digits look fine to a tired human at 11 PM, and that ad would have run all week sending customers to a wrong number. And Meta's ad rules are strict about weight loss and body claims, so a "guaranteed" result can get a client in real trouble.
Claude also burned the captions into A1 and A2 in the same step, and it was honest about one more thing. It only checked the audio, so a human still has to watch both videos. That's the right call.
The point of the gate is that you only look at the failures, and each one comes to you with the reason next to it.
Step 4, Running the Ads
This used to need a media buyer. Now it's one prompt, word for word.
"Drop A3. Draft a Meta campaign in my ad account for A1 and A2. One campaign, goal is bookings, Austin, women 25 to 45, $10 a day per ad, both ads in one ad set so Meta finds the winner. Write the primary text and headline for each from the scripts. Save it as a draft, paused, and show me what you created."
Claude built the campaign and the ad set and put them into Ads Manager, paused. I opened Ads Manager in the browser and there it was.
It also fixed two things I got wrong in my own prompt. Meta sets the budget on the ad set instead of on each ad, so it combined my $10 per ad into $20 a day for the ad set. And the account had no tracking pixel, so Meta can't see bookings yet. It set the goal to getting people onto the booking page instead, and said we can switch to bookings once a pixel is in.
One honest note. The campaign and the ad set saved, but the two ads themselves didn't attach, because this test account had no payment method on it. More on that in the limits.
Now the most important rule in this whole article. Claude drafts, a human publishes. Someone on the team opens the campaign, checks the ads, the text and the budget, and hits publish themselves. It takes two minutes.
Both ads sit in one ad set with a small daily budget. Meta shows both to people and slowly pushes more money to the one that works better. You don't have to pick the winner yourself.
Then you leave it alone for a few days. Meta needs that time to figure out who to show each ad to, and every big change makes it start over.
After a few days you look at three things. How many people stopped scrolling to watch. How many of them clicked through to the booking page. And what each booking cost you. That last one is the only number the client really cares about.
And then the part nobody warns you about. Even your winner gets tired, because the same people see it again and again. Chase shows this on a big account in his video. When you spend more, you need more fresh ads, or the old ones stop working. If a new ad takes 3 to 6 weeks, the winner has burned out before the replacement exists.
Step 5, The Monday Refresh
So every Monday the system makes new ads from what worked. Here is the prompt that sets it up, word for word.
"Write the plan for every Monday from now on. When a week of results comes in, which ad do we switch off, how do we pick the winner in plain English, and how do we make two fresh Marketing Studio ads from it without repeating a hook. Then set this up so it runs every Monday."
Claude wrote the plan and scheduled it to run every Monday at about 9 AM Austin time, right there in the project. The rules are simple enough for anyone to follow.
- The winner is the ad that gets people onto the booking page for the least money, and it has to be at least 20% cheaper to count
- If it's closer than that, the ad whose first 3 seconds stop more people wins
- An ad only gets switched off once it has spent at least $30 and been shown 1,000 times, and the last running ad is never switched off
- Two fresh ads every week, one keeps the winner's angle with a brand new opening line, one tries a style we haven't tested yet
- Both get checked against every hook in the hook log and go through QC before they cost anything
And nothing gets paused and no credits get spent until I say yes. Every Monday I get a results table, the winner and why, the two new scripts, the credit cost (120 credits), and one question.
The hook log is what keeps this working over months. Without it, by week six Claude is giving you the same openings in slightly different words.
This loop is what clients pay a retainer for, because they can see the numbers move every week. In the Higgsfield video Nate Herk suggests pricing it at around $2,000 a month.
(I'm sending the copy-paste version of this whole setup to my newsletter, The AI Head Start.)
What Changed for the Agency
Their prices came down, and they could finally compete with the small shops that had been undercutting them. They cut down on the team, because a lot of the production work just doesn't need people anymore. They started giving their clients better offers, more ads for less money. And their margins went up, because a batch of ads now costs credits and an afternoon instead of a shoot day.
Now the honest part, because this is the bit the demos never show you.
It's not perfect. The ads still look like AI if you watch closely, they came out without captions, and one had a clipped first syllable. Ads like that get fixed or thrown out before they go anywhere. Clients still want a real human on the call, nobody wants to hear strategy from a chat window. And credits burn on retries, even when nothing ships.
You Can Start This as a One-Person Agency
I started my agency, IgnytLabs, in 2024, building MVPs for clients with AI. It was the first AI business I ran and imo an agency is still the best first AI business for most people.
Early margins are high because your cost is mostly your time plus some credits, and clients who get results turn into retainers.
With this workflow the ladder is simple. A free sample ad to show you can do it, a first paid batch, then a monthly retainer for the Monday refresh. The Higgsfield video's math is $500 for the first project and $2,000 a month for the retainer, so five retainer clients is $10,000 a month. That's math on paper and nobody can promise it to you. But the system does the same work for five clients as it does for one.
And when you get there, you don't make ads one by one anymore. Higgsfield has a mode called Supercomputer that runs the whole job in bulk in the background. You give it the brief and the brand kit and it makes dozens of ads at once while you're on client calls. For two ads I used Marketing Studio through Claude, for 20 ads a week across five clients this is what I'd switch to.
The Limits
Everything above is real and all of it has rough edges, so here they are before you learn them on a client call.
- Never let Claude publish. Keep every campaign paused until a human has looked at it.
- Meta's video upload through the connector is still rolling out. On my account it said "This tool is new and is being gradually rolled out across ad accounts," so I uploaded the two videos to Media Library myself. Takes about two minutes.
- Meta won't save even a paused ad until the ad account has a Facebook Page linked and a payment method added. Set both up before you start on a client account.
- The Ad Library connector only sees headlines, so treat the research as a starting point.
- Credits add up. Every retry costs, and the Monday refresh is new credits every single week. Price the retainer with that in mind.
- The QC gate checks facts and claims, it can't tell you an ad looks fake. Watch every ad before it goes into a campaign.
The Build Sheet
The whole system in one line.
Client profile in a Project → connect Meta Ads and Higgsfield → borrow data from the market → approve the script and shots → generate the ads → QC gate → Claude drafts the campaign, paused → a human publishes → keep the winners → Monday refresh
Set it up once per client, and after that every week is one prompt on Monday.
My client's agency runs on this workflow now. It started as a proper traditional business, and today one person could run the same system from a laptop.
I send the full project instructions, the campaign prompt and the QC prompt to my newsletter, The AI Head Start. If you want the next one of these in your inbox instead of hoping the algorithm shows it to you, that's what it's for.
