Six months of AI in the ad account – what works?

Six months on, the picture starts to clear

Over a few weeks in April and May, Google, Meta and TikTok released official MCPs for their ad accounts. Google went first, on April 28, Meta the next day, and TikTok on May 13. An MCP works as a connection that passes data from the ad platform on to a language model like Claude, ChatGPT or Gemini. There is no intelligence of its own in the connection itself; it reads data and carries out changes the platform approves, but it says nothing about what you should do with the data.

Since the MCPs launched, feeds have filled up with guides and promises that business owners will be able to handle advertising entirely on their own, without an agency's help. Six months on, the enthusiasm has cooled somewhat. In StackAdapt's global report The AI Delegation Gap, a survey of 500 marketers in the US, nine out of ten said they're happy to let AI suggest actions, but only half wanted it to decide on its own. What they wanted to control the most was the budget.

Why the promise of full automation falls short

Torkel Öhman, CTO and co-founder of the ad tool Amanda AI, discussed where the limits lie when he appeared on the podcast Digital Marknadsföring with Tony Hammarlund this summer.

One of the problems he raises is that a language model is built for language, not math. Budget and optimization require exactly that: math, where the same question should always produce the same answer. A language model guesses its way forward and can give different answers to the same question twice.

Öhman compares the uncertainty of AI in advertising to a weather forecast. A 70 percent chance of rain on Saturday is perfectly fine, but nobody wants that same uncertainty about their bank balance; there you need an exact answer. The exact same logic applies to the ad budget.

Then there's a risk that's easier to miss. Because a language model compresses data and draws conclusions from summaries, you get more and more summaries of summaries the longer a session runs. That creates a snowball effect where small errors grow and hallucinations creep in. The model also only has access to what you feed it, yet answers with the same confidence even when it's wrong.

Budget is the riskiest thing to hand over

Since the media budget is your biggest lever, real money is at stake if something goes wrong. A campaign showing a good ROAS still doesn't mean it can absorb unlimited spend, because saturation and competition set limits that an AI model rarely keeps track of.

Another important detail raised in the podcast episode is that the platforms themselves do not optimize with language models. Meta runs several purpose-built systems for different parts of the ad platform, and Google's automated bid strategies are driven by machine learning. If a language model produced better results, bidding would already rely on one. The fact that it doesn't shows where the limit lies, Torkel Öhman argues.

Other sources make similar arguments. The ad-tech company Basis, which sells automation tools itself and therefore has a vested interest here, states in an industry review that what measurably delivers results is algorithmic bidding specifically, while strategy, judgment and exceptions still require a human.

What is AI actually good at in advertising?

At the same time, it's easy to underestimate what the technology is good at. Language models excel at producing ad copy, setting up campaigns and generating image material at scale. The strength lies in creating unique text for each product and landing page instead of a generic version that fits everything half-well, and there's a real time saving there.

They're also good at reporting. A language model can replace a regular report or dashboard and answer a question like "which ads performed best over the past seven days" in plain text instead of a table. In other words, it describes what has happened.

Explaining why the numbers look the way they do is harder. For that, the model needs three things fed into it:

1. 1

   Finished calculations from a system built to compute.

2. 2

   The rules for how marketing works and should be judged.

3. 3

   Awareness of what is happening around it, such as other campaigns, seasonality and competitors.

If the AI has these three pieces, it can put the numbers into words, while the actual analysis happens elsewhere. Without them, though, there's a risk of a polished report built on the wrong basis. One example: you run a TV ad at the same time as advertising on Google, and the TV spot drives sales up. The AI model only sees the Google account, not the TV ad, so it goes looking for an explanation that isn't in the account.

Three ways in, with different levels of control

If you're going to let AI into the ad account, you currently choose between three paths, and they differ in how much control you keep and how big the risks become:

1. 1

   Connect an MCP to a language model, like Claude or ChatGPT. You control everything yourself, but you also carry the full responsibility, and there is no support line to call if something goes wrong.

2. 2

   Use the platforms' own AI assistants, like Meta AI Business Assistant and Google's Ask Advisor. They have more built-in guardrails, but only see their own platform. Meta only looks at what happens in Meta and Google only at Google, so you lack an overview across your channels.

3. 3

   Choose a dedicated service that oversees several channels at once and relies on a system built to compute and optimize, with the language model as a layer on top that explains and communicates.

How to test without losing money

You can try all of this without it affecting your campaigns. Open a language model like Claude or ChatGPT, without connecting an agent or MCP, and feed in your data by hand. Ask the same question multiple times and check whether the answers seem reasonable and consistent. A good trick is to create two identical projects, give them the same task and compare the answers; if they differ a lot, that's a sign the model is guessing.

Always have a knowledgeable person involved to check the work. Start with the low-risk stuff, like reporting, and let the model propose suggestions that you approve yourself before anything is actually carried out. Only once a task has worked reliably, time after time, should you let it run automatically, and even then it's best to keep a person reviewing everything.

What this means for you

Letting AI into the account doesn't change the fact that control must stay with a human. Agencies are likely to be around for a long time, and those who believe AI solves everything on its own will pay the price for it. Consultants are already emerging – like SEO expert Nikki Pilkington – who specialize in fixing websites broken by blind automation, and we see the same pattern in performance marketing.

The tools take work off our plate so we have time for other things, but letting fear drive decisions is no way forward either. The answer lies somewhere in the middle: let experts and the right systems handle what requires precision, use AI for what it's actually good at, and automate one step at a time. That way, results improve without gambling with the budget.

Frequently asked questions