Creating Meta ads with AI means two different things that get mixed up constantly. Inside Ads Manager, Meta's own AI is at work: the Advantage+ features build image variations, write up to five text variations, assemble the audience and spread the ad across Facebook, Instagram, WhatsApp, Threads and the Audience Network. An outside AI such as Claude or ChatGPT works before all of that: market research in the Ad Library, sharpening the offer, describing the audience in real words, writing ad copy and test variations. Two things are worth knowing up front: according to Meta's own help centre, text generation is currently available when the primary text input is in English, Portuguese or Spanish, and Meta reports per ad, not per text variation, so you never learn which message won. That is exactly why the decisions that cost money stay with a human: the offer, the audience limits, the budget and the approval.
I run my business with a workforce of AI employees, each with their own role and their own personnel file. One of them, Susi, does nothing but Meta ads, and the workflow below is the one she follows. If you want to see the whole team: My AI Workforce, and everything on the topic lives on the AI employees page.
What Meta's own AI handles inside Ads Manager
Meta bundles its AI features under the Advantage+ name. According to the help centre, some of these enhancements are on by default and can be turned off at any time. That is the most important sentence to start with: if you change nothing, AI is in play anyway.
Meta publishes its own figures for two of these. Both are vendor figures from Meta's help centre as of September 2026 and no commitment for your account: ad sets using Advantage+ placements saw an average 11.7 per cent lower cost per action in one experiment, and for Advantage+ audience Meta cites 7.2 to 14.8 per cent lower cost per result depending on the campaign objective. Meta itself states elsewhere that a high score in its Opportunity Score tool is no guarantee of performance.
Two limitations come with all of this. First, language: Meta writes that text variations with and without personas are currently available when the primary text input is in English, Portuguese or Spanish, and that the features are being rolled out gradually. If you advertise in another language, you cannot rely on text generation inside Ads Manager. Second, industry: financial services, pharmaceuticals and health, plus ads for housing, employment, credit and political or social issues, may per Meta not get immediate access to all generative AI features.
Where Meta's AI stops
There is one line in the help centre that appears in no click-path tutorial and still changes everything: because ad reporting is produced per ad, you cannot see performance details for individual text variations. Meta's text variations may bring you more results, but no insight: afterwards you do not know which promise pulled, so you cannot carry it over to your landing page, your emails or your next offer.
The second spot concerns targeting. If you optimise for link clicks, landing page views, conversations, conversions, app events, app installs or value, the help centre states that Advantage+ detailed targeting is applied automatically and cannot be switched off. Your interests and behaviours are then suggestions, not limits: Meta will also show the ad to other audiences when that is likely to improve performance. The only hard limits you keep are audience controls such as minimum age, location, language and exclusions.
Which leads to the actual point of this article. Meta's AI optimises delivery. It does not optimise your offer, because it does not know it. Flipping the switches in Ads Manager automates the distribution of a message nobody has checked.
The part before that: who you want to reach, and with what promise
Most Meta ads tutorials explain click paths. The click path is not the problem. The problem is that an AI which does not know your offer produces interchangeable ads: the same three hooks, the same question in line one, the same call to action you could paste under every second ad out there. This is exactly the part an outside AI can genuinely take over, provided you feed it.
Audience research with real material. According to Meta, the Ad Library contains all active ads running across Meta technologies, and anyone can open it and search by term, name or page. That puts your competitors' ads in front of you in full text: the words they use to name the problem, the promise they make, how long an ad has been running. That is material, not guesswork. A long run time is an indication, not proof of profitability: the library shows what is live, not what pays off. An AI that summarises twenty such ads and sorts them by recurring phrasing hands you a map of the language in your market within half an hour; how such material becomes an audience map in five steps has its own post.
Write down the offer. Before a single line of ad copy exists, five things need to be concrete: for whom, which problem, which promise, which proof, and what happens after the click. Leave one of them out and the AI fills the gap with a platitude. That is not a model problem, it is a briefing problem.
Ad copy and variations. Only now does the AI write, and not one ad but variations that differ in exactly one thing: the promise; how such an ad text is built and how short it has to be has its own post. Because Meta does not report per text variation, you set these up as separate ads rather than as text options inside one ad. That is what makes each message individually visible in the first place. Visible is not the same as compared, though: Meta's delivery does not spread evenly, it pushes budget and impressions onto a few ads early. A comparison you can rely on therefore needs three things fixed before launch: the one metric you judge by, a minimum amount of delivery per variation, and a structure that does not starve the weaker variations straight away, which in case of doubt means its own ad set with its own budget per variation.
If you feed the AI real customer emails, survey answers or chat logs, the usual privacy handling applies: what goes into an AI tool should be a decision you made beforehand. How I draw that line is in AI and Privacy: What the AI Gets to See.
The guide: six steps to a finished, paused campaign
- Pick the objective. Awareness, traffic, leads or sales. The objective determines which Advantage+ features are available to you, and whether detailed targeting kicks in automatically. Which stage of your funnel the campaign serves is covered in Marketing Funnel Examples: Stages in Practice.
- Mine the Ad Library. Collect ten to twenty active ads from your market and have the AI sort them by promise, audience address and format. The output is a list of real phrasings, not a persona fantasy.
- Put the offer and the audience in writing. The five points above in a short brief. This brief is everything the AI knows about your business, and therefore the ceiling on quality.
- Build copy and creative. Three to five variations with different promises, each intended as its own ad, plus the decision, made now, on which metric counts and how much delivery each variation has to get before you judge it. If that metric is ROAS, work out your break-even ROAS first, as shown in ROAS Formula: Calculation and Worked Examples. You generate the imagery yourself or leave it to Meta's image generation, knowing that you cannot evaluate Meta's variations individually.
- Set up the campaign. In the ad account under your business portfolio, set structure, naming and audience limits. A naming scheme that doubles as tracking saves you from reconstructing later which ad was which hook. How those names travel as tags on the link into your website stats is covered in UTM Parameters: Examples and Common Mistakes.
- Leave it paused, check, then approve. Spelling, promise, landing page, budget type, run time. And the rules: Meta's advertising standards prohibit, among other things, an ad asserting or implying personal attributes of the person it addresses, and special ad categories such as housing, employment, credit and political or social issues carry extra restrictions. Every ad is reviewed before delivery.
Budget: the mechanics, not the number
Nobody who does not know your offer and your margin can tell you the right number. You should still understand the mechanics, otherwise you are buying an automation that distributes differently than you think.
A daily budget is, per Meta, an average amount per day, not a fixed cap for each individual day. A lifetime budget is the amount for the entire run and has a fixed spending cap. If you use the Advantage+ campaign budget, you set one campaign-level budget that is distributed continuously and in real time to the best-performing ad sets. Meta explicitly points out that this budget is not split evenly and that results therefore have to be read at campaign level, not per ad set. If you want guard rails, you set minimum and maximum spending limits per ad set, or you assign budgets at ad set level and buy yourself control instead of automation.
Where a human has to decide
In my shop, Susi does all of this up to the approval
All of that groundwork is handled by Susi, my AI employee for Meta ads. I give her an objective, community, webinar or retargeting, and get a complete campaign back: Ad Library research on the competition, an audience map built from real audience words, copy following a fixed four-step structure, imagery as HTML statics plus an AI image layer, and the campaign set up in my Meta account. If I'm promoting a Skool community specifically, the matching workflow for that is in Running Meta Ads to a Skool Community.
The decisive difference sits in one word: every campaign the Meta Ads connector creates starts out paused, enforced technically. Not as an agreement, not as good intentions, but as a state. Susi spends no money of her own accord. I am the one who goes live, and only after reading the ad. Evaluation follows fixed rules rather than mood in the same way: at least three full days or 3,000 to 4,000 impressions as the test window, then a verdict from three options, kill, scale by 20 per cent, or let it run. The kill threshold is set before launch, not after. Susi has worked like this since 17 July 2026; her report folder holds 44 documented evaluations and research pieces for it, as of 19 September 2026.
Meta's advertising rules follow the same principle with her: the copy neither asserts nor implies personal attributes, it promises nothing the offer does not keep, and targeting as well as special ad categories are checked against Meta's current rules. Labelling an image that came out of an AI is a house rule of ours, not a function Meta takes off our hands. Those are not courtesies, they are the condition for an ad account staying alive.
Frequently asked questions
Can an AI create Meta ads entirely on its own?
The build yes, the decision no. An AI can research, build audience and copy, and set up the campaign in the ad account. Whether it goes live and with what budget is a business decision. In my setup that is enforced technically: campaigns come into existence paused, and going live is a separate move that I make.
Do I need an outside AI when Meta has AI built in?
For anything outside English, Portuguese and Spanish, almost certainly yes. Meta's help centre ties text generation to primary text input in those three languages, and reporting runs per ad, not per text variation. Variations you write yourself and set up as separate ads can be read individually, provided each one got enough delivery.
Does Meta switch the AI features on automatically?
Partly. Some Advantage+ creative enhancements are on by default per Meta and can be turned off at any time. Detailed targeting has no off switch once you optimise for link clicks, conversions, leads or value: your interest selections then work as suggestions, not limits.
How do I find out what my competitors are running?
Through the Meta Ad Library. Per Meta it contains all active ads across Meta technologies, is accessible without an ad account, and can be searched by term, name or page. For ad copy it is the most honest source: you read what is actually running instead of guessing at what might work.
What does it cost to create Meta ads with AI?
Two separate items: the tool, meaning your AI subscription or a ready-made employee template, and the Meta ad budget itself, which you set and approve. The AI side is the predictable part, the ad budget the variable one, the same split as between agency fee and ad budget. What the tool side costs in my setup is in What an AI Employee Really Costs.
Where to go from here
Start with the Ad Library, not with Ads Manager. Twenty active ads from your market, sorted by promise by an AI, are a better foundation than any template pack. If you want to turn that into a standing role rather than a one-off exercise, the workflow is in Creating an AI Employee: How to Start. Susi herself is available as a ready-made template in my community AI Employee Framework, along with the rest of the workforce. And if you get stuck on the test design, meaning which metric decides and how much delivery a variation needs: that is exactly the question to post there or bring into one of the calls, rather than solving it on your own.
