How Are AI Agents Doing Marketing Now?

AI marketing agents can now research your buyers, make your ads, launch them and pull the losers while you sleep, for a fraction of a creative team's cost. That part is real. What most owners miss is why one company's agent turns out winners while another's turns out slop, and the reason has little to do with the AI.

Here's the picture I keep coming back to. Hire the best chef in the city and drop them into an empty restaurant on their first day. There's no pantry, no set of keys, no knives, no recipe box, and nobody standing at the pass tasting plates before they go out. That chef will still cook you something, and it won't be good. An AI agent is that chef, and the model is rarely the problem. The kitchen is.

I'll come back to that kitchen in every section, because it explains most of what's happening in marketing right now. But first, the question nearly every owner asks before anything else.

Do AI-generated ads actually work?

Yes, when you use them to find winning ideas fast, and no, when you use them as a cheaper way to make the same ad you were already making. Most owners I talk to judge AI ads by the worst ones they've scrolled past, the glossy, slightly waxy spokesperson who blinks wrong. Those ads exist. They're just not what the companies getting results are running.

The reality is that ad platforms have become so good at matching an ad to the right person that a strong ad now finds its own buyer. You don't need to hand Meta a perfect audience anymore. You need to hand it a strong idea, in a form people stop scrolling for, and then another strong idea after that. AI's real advantage is that it lets you test forty ideas in the time it used to take to shoot two.

In my experience, the AI-made ads that work don't look AI-made. The video ones start from a photo of a real-looking person you have the rights to use, remixed with a new background, a different age or a different setting, so the face carries real skin and real light instead of that plastic sheen. The static ones look like a sharp designer made them, because the agent studied your brand before it touched a pixel.

And sometimes the AI version simply wins. Run the same script past the same audience, once read by a hired creator and once by an AI spokesperson, and every so often the AI read pulls more leads. Nobody can fully tell you why. You don't have to know why. You have to test it, because your guess about what your buyers will respond to is far less reliable than their behavior.

Will some drafts come out garbled? Of course. A face warps, a hand grows a sixth finger, a word gets slurred. When a draft costs a few dollars instead of a few hundred, you throw that one away and move on. Back in the kitchen, a chef who can try forty dishes this week will find your menu long before the one who can afford to try two.

How many ads should you run now that Meta's Andromeda update decides who sees them?

More than you're running today, and far more different from each other than you'd guess. Andromeda is Meta's system for the first cut of ad delivery. Every time someone opens Facebook or Instagram, it narrows tens of millions of eligible ads down to a few thousand candidates for that one person, and Meta built it to handle an explosion in the number of ads. By late 2024, more than a million advertisers were already using Meta's own generative tools to make over 15 million ads a month.

Think about what that means for you. A media buyer's old job was choosing the audience. Now the system chooses the audience, and your ad is how it figures out who to show. Your creative has become your targeting. If you hand Meta four ads a month, you've given the most powerful matching engine in advertising four guesses to work with. Meta is still one of the busiest rooms for most companies, which is why it held up well when I ranked every major marketing channel.

Volume alone won't save you, though. Thirty versions of the same ad with different background colors are still one idea, and in my experience they perform like one. What moves results is thirty ads making genuinely different arguments: a different pain, a different promise, a different person talking, a different reason to care this week.

So how many is right for you? It comes down to your budget and your cost per lead, because every ad needs enough spend to prove itself. Picture a $10M software company paying about $150 per demo request. If each new ad gets roughly three leads' worth of spend, $450, before anyone judges it, then a $15,000 monthly testing budget supports about 33 new ads a month. That company is probably running a tenth of that today.

Don't panic when an ad gets a couple of conversions and then the platform quietly stops spending on it. That usually means Meta found the handful of people that ad was right for and moved on. It isn't a failure. It's the system doing its job, and your job is to have the next ad ready.

One habit separates the companies that learn from the ones that just spend. Every AI-made ad starts as a set of written instructions: the script, the visual description, the voice, the format. Save those instructions right next to the ad's results. A few hundred ads later, you're no longer guessing which ingredients sell. It's a recipe card stapled to every plate's sales receipt, and almost nobody in your market is keeping one.

How does an AI agent make an ad from start to finish?

It does what a strategist, a copywriter, a designer and an editor used to do, in that order, in minutes instead of weeks. The agent isn't running one clever prompt. It's following a written procedure, usually called a skill, that tells it which steps to take, which tools to call and what good looks like at each stage. You give it one instruction and a website, and it works through the whole chain.

How does it make a static ad?

It starts by studying you. The agent reads your website, pulls out your colors, fonts, tone and the claims you actually make, and saves them as a simple brand file it will follow every time. Next it opens Meta's public Ad Library to see what you and your competitors are already running, so it doesn't hand you a copy of your rival's best ad.

Then it does the step most human teams skip. It goes and reads what your buyers say when you're not in the room: forum threads, reviews, community posts and complaints. Picture a $12M property management software company. The agent finds hundreds of property managers venting about 2 a.m. maintenance calls and tenants who never read the move-in packet, and those exact frustrations become the raw material for headlines. It's a chef walking the local market before writing the menu.

From there it proposes five or so concepts, each with a headline and a visual idea, and waits for you to approve or redirect. Once you do, it generates the images with whichever image model is performing best that month, since they're interchangeable, and uploads the finished ads into your ad account as a new ad set, ready for a last look before they spend a dollar.

How does it make a talking-head video ad?

The trick is never letting the video model improvise. Hand a video generator a script and ask it to make an ad, and it will invent a face, mumble words that aren't in the script and give you something you can't use. So a good agent builds the ad in pieces and fences each piece in.

First it writes the script and generates the voiceover on its own, as an audio file. Then it sends three things to the video model together: that audio, the exact script and a reference image of the spokesperson. With the words, the voice and the face all locked, the model has one job, which is to make that person deliver that recording. It renders a cheap low-resolution draft first and only upscales the drafts worth keeping.

Then comes the part that used to need a video editor. The agent transcribes the draft word by word and checks it against the script. If the spokesperson mangled your product name eleven seconds in, it finds that exact moment and patches only that word instead of regenerating the whole ad. The same word-by-word timing then drives the captions and on-screen motion graphics, so text appears exactly as each word is spoken.

Want the spokesperson holding your product? Give the agent a photo of the product as a second reference. Want one version for warehouse managers and another for CFOs? Swap the reference person and the script. The whole chain can run on a cloud server overnight, so you wake up to a folder of finished drafts instead of a production calendar.

How do you keep AI ads on brand?

You write the brand down once and make the agent read it every time. That brand file holds your colors, type, voice, the words you never use and the claims you can legally make, and it sits alongside the agent's instructions. When an ad comes back off-brand, you don't fix the ad. You fix the file, and every future ad inherits the fix.

Can you use Claude or ChatGPT as your marketing agent?

Yes, and which one you pick matters far less than what you plug into it. Claude Code, ChatGPT and the other general-purpose agents can all run the workflows above. The instructions that make them good at marketing live in reusable skill files, not inside any one product, so the same procedure runs in whichever agent your team prefers. Swapping chefs in a bare kitchen gets you the same bad meal.

The models doing the image and video work are just as swappable. When a better image model comes out next quarter, you change one setting and keep going. That's why I tell owners to stop debating which AI is smartest. The question that matters is whether you own the procedures, the data and the connections, which is the whole idea behind owning your software instead of renting someone else's black box.

Two practical details make a big difference. First, the agent doesn't have to run on anyone's laptop. It can live on a cloud server and work through the night. Second, the setup can be shared, so everyone on your team directs their own agent against the same ad accounts, with the same brand file, the same rules and the same data, without rebuilding the plumbing each time. I've written about how this changes the marketing leader's job in the marketing engineer.

Will AI agents replace my marketing team or agency?

They'll replace a lot of the tasks, not the judgment, and the companies that confuse the two get burned. Graphic design, ad copywriting, video editing, motion graphics and the research a sales rep does before a cold email can all be handled by an agent with the right setup today. Even long-form editing has fallen. An agent can strip the dead air out of a ninety-minute podcast, cut to whoever is speaking and drop in the screen shares on its own.

Some of what agents do, no human team would ever do at any price. Picture a Google Ads account matching a few thousand different search terms every month. An agent can read every one of them, ask whether the person searching is actually ready to buy what that landing page sells, and flag the mismatches before another dollar goes out. A person could do that in theory. Nobody does, because it would eat the whole week.

Here's the mistake I watch smart owners make. They try to build one all-knowing AI marketer and hand it the keys to the ad account. In my experience that goes badly almost every time, because nothing in the setup tells it what good looks like. The version that works is humbler. Write down exactly what your best media buyer does every morning, step by step, then turn each step into its own small routine that runs on live numbers.

The same thinking applies to real people. Agents can now run a pool of human creators through the creator platform's own systems, sending each one a brief and a script variation, paying per video plus a bonus when an ad performs, then adding captions and motion graphics to the footage that comes back. The agent takes over the coordinator's job, and the faces and voices stay human.

So your team doesn't disappear. Its job moves from making every ad to deciding what good looks like, setting the bar and saying no. If you've watched an AI project stall inside your company, that shift is usually where it broke down, and I wrote about why AI projects fail and what to change.

Where should a person check the work before it goes live?

At the pass, after the machines have already done the first round of checking. Before you see a single draft, a second agent should review it the way a tough creative director would, watching the video frame by frame, checking that the captions line up, scoring it against a written standard and sending it back until it clears the bar you set. You only see what passed.

If you sell in a regulated space like health tech or financial services, add one more checker. Give an agent the list of claims you can't make and words you can't use, have it read every ad against that list, and send anything that fails back for changes. Your compliance person ends up reviewing a short stack instead of a pile.

Train that creative director agent on the fundamentals of advertising, the classic books and the best working ad minds, and on your own library of ads you admire. Then keep one human gate before anything spends money. That's where taste gets added, and taste is the one ingredient you can't buy from a model.

What does an AI agent need before it can run your marketing?

It needs a stocked kitchen: your data in one place, permission to act in your ad accounts, the right tools, a library of good work, and the pass we just covered. Skip any one of them and the most capable agent in the world will hand you confident, polished nonsense. Here's what each piece looks like inside a real company.

Why does your marketing data need to live in one place?

Because an agent that can't see the whole story will make up the missing parts. Point an agent straight at your ad platforms, your analytics and your CRM, and it drags huge piles of raw data into its working memory, starts guessing, trips over the platforms' limits on requests and often builds its own messy little database on the side, full of duplicates.

The fix is the pantry. A data pipeline pulls from every source you have, your ad accounts, web analytics, CRM, product usage and payment processor, into one data warehouse that keeps syncing in the background. A data warehouse is a single database built to hold all of a company's records in one place so they can be questioned together. The agent asks its questions of the warehouse as often as it likes and never has to hammer the platforms themselves.

The payoff goes well beyond cleaner reports. Picture two ads with the same $150 cost per lead. One brings in customers who stay three years. The other brings in customers who cancel after four months. Your ad dashboard says they're equal. Your warehouse, which can follow a customer from the first ad they saw to their last payment, says one of them is worth about nine times the other, and now the agent knows where to put the money.

Can an AI agent get your ad account banned?

Yes, and when it happens, it's almost always because the agent was pulling data straight from the ad platform. Agents are relentless. Ask one to analyze an account and it might fire off hundreds or thousands of requests in seconds, which breaks the platform's terms. Picture a new hire who calls your bank's fraud line a thousand times in a minute. Nobody at the bank cares how good his intentions were. They freeze the account.

The rule I set is simple. The agent gets permission to write to your ad platforms, meaning upload new ads, pause losers and move budget, through each platform's official connection for advertisers. Every read comes from your warehouse. Setting up that connection is quick, and a good agent can walk you through it step by step.

What tools does the agent need?

It needs a toolbox it can reach on its own: an image model, a video model, a voice generator, transcription, access to several language models and a source of business contact data if it's doing outbound. Each one is a paid account with a key you hand to the agent. With those in place, the agent stops describing work and starts doing it.

Why does it need a library of ads you admire?

Because nobody cooks well without recipes. Copywriters have always kept swipe files, folders of ads they admired, and a few dozen used to be plenty. An agent can study thousands, so give it thousands: ads that stopped you mid-scroll, competitors' long-running winners, landing pages you envy. Keep adding to it every week, and make it one shared library so every agent on your team draws from the same well.

Notice what never made the list: the AI model. Every owner I talk to is shopping for a better chef. The companies pulling ahead stopped shopping months ago and spent that time stocking the kitchen, and now the same agent you could buy this afternoon is cooking circles around everyone else's.