AI for media buyers: what works today, what is a promise

AI for media buyers is worth exactly what you can verify. Today it is genuinely good at four things on a live account: reading the numbers, comparing periods, spotting what moved, and executing one specific change you confirm. It is still a promise at the thing it gets sold on, which is running the account for you. If you are deciding whether to put AI in your workflow this quarter, this is the split between what already works, what is marketing, and how to start without putting a client's budget at risk.

The quick answer

What already works.

• Reading. Pull spend, CPL, CPA, ROAS for a period, by campaign, by ad set, by day. Fast, and correct when it is querying the API instead of guessing.
• Comparing. This week against last week, this account against that one, this placement against the other. The mechanical work of a comparison is exactly what a model is good at.
• Spotting anomalies. "Anything with CPL above US$ 20?" or "which campaign changed most since Monday?" gets you a shortlist in seconds instead of a scroll through Ads Manager.
• Executing a simple, bounded change. Pause this campaign. Set that daily budget. One action, confirmed by you, inside a limit you set.
• Drafting. Ad copy variants, the written half of a client report, the summary paragraph you rewrite every month anyway.

What is still a promise.

• Strategy on its own. Which offer, which audience, which channel deserves the next US$ 5,000. The model does not know your margins, your sales team's capacity, or the client conversation from last Thursday unless you tell it.
• Unsupervised optimization. Anything sold as "the AI optimizes your campaigns while you sleep" is asking you to trust a decision loop you cannot inspect. The platforms already run their own optimization inside the auction; a layer on top that also decides is a second driver with a second steering wheel.
• Judging creative. It can describe a creative and pattern-match against what usually performs. It cannot tell you the hook is dead in your market this month.
• Knowing what it does not know. This is the important one, and the next-to-last section is about it.

The three levels of using AI for media buyers

Almost everybody starts at level one and thinks that is the whole thing.

Level 1: paste. You export or screenshot and ask.

Example: "Here is last month by campaign. Which three cost the most per lead, and what is the spend behind each?"

It works. It also means the export is still your routine, and every follow-up question is another export. The failure mode is subtle: the model reads what is in the image, including the column you cropped, and answers confidently about a partial picture.

Level 2: connect the account, read-only. The account is exposed through a connector, and the model queries it live. This is where AI stops being a novelty. It needs MCP, the open standard that lets a model call outside tools, explained in what is MCP for marketers.

Example: "Across all connected accounts, which campaign had the cheapest lead in the last 7 days, and how much did it spend? Then break that account down by placement."

That is two questions in one line, on live data, with no export anywhere. It is also completely safe: reading changes nothing.

Level 3: execute with guardrails. Same connection, plus a small set of write tools that you switch on deliberately.

Example: "Pause the ad set with the highest CPL in account X, but show me the numbers first."

The model shows you what it found, proposes the action, waits for your yes, runs it, and the change lands in a log. That is a very different thing from an AI deciding on its own to pause something at 3am.

The real risks, and how to protect yourself

Three failure modes matter. The rest are variations of them.

It answers about data it does not have. A model asked a question it cannot answer will often produce a plausible-looking number rather than admit the gap. This is the single biggest reason to prefer a connected account over pasted screenshots: when the numbers come from the API through a defined tool, there is nothing to invent, and a well-built connector tells you the field is not available instead of improvising. Protect yourself by spot-checking. Pick one number a week and open Ads Manager.

It acts without you meaning it to. "Clean up the underperformers" is not an instruction, it is a mood. If write access is on and the confirmation step is missing, an ambiguous sentence becomes several pauses on a live account. Protect yourself with a confirmation on every action, a cap on what a single action can set, and the habit of naming the campaign you mean.

Access with no scope. Connecting every account you have ever touched, with full permissions, because it was easier than choosing. Then someone on the team asks a broad question and gets an answer spanning clients who should never appear in the same sentence. Protect yourself by connecting only what you are actively working on.

The checklist before you give any AI access to an ad account

Run this against any tool, including one your own team built. If you want the technical version of the same argument, Meta Ads MCP server compares building your own connector against using a hosted one.

• Authorization is OAuth, through the platform's own login screen. No tool should ever ask you to type your Facebook or Google password into it.
• You choose which accounts are included, and you can shrink the list later.
• Write actions are off by default, as a separate switch from reading.
• There is a cap per action, so a misunderstanding gets refused instead of executed.
• Every action is confirmed in the chat before it reaches the platform.
• Every change is logged with author, origin, and the state before and after, refused attempts included.
• You can revoke access in one click, from the tool and from Facebook or Google directly.
• The tool says "I do not have that" when a metric is not available, instead of producing a number.

Eight lines. If a vendor cannot answer all eight without hedging, stay read-only.

Where to start

Read-only, one account, one week. Pick the account you know best, so you can tell immediately when an answer is wrong.

For that week, replace one habit and only one: whatever you currently open Ads Manager to check every morning, ask it in the chat instead. Verify the answer against the interface for the first three days. By day five you will either trust it or you will have found the specific thing it gets wrong, and both outcomes are worth the week.

Then decide separately about writing. Turning it on is not the natural next step, it is its own decision. Plenty of good media buyers stay read-only forever, and that is not a half-measure. If you do turn it on, start with a cap low enough that the worst possible action is annoying rather than expensive.

What does not change

The job. You still have to know which question is worth asking, which is the same skill as knowing what belongs in a report, covered in Meta Ads reporting. You still own the offer, the creative direction and the client relationship. Ads Manager is still where the campaign lives; a connector complements it rather than replacing it.

What changes is how much of your week goes into assembling numbers instead of deciding things. That is not a small change, and it does not require believing any of the bigger claims. Start read-only on one account, and Claude for paid ads has the prompts and the guardrail settings for whenever you are ready for the next level.

If you are still choosing an assistant, see Gemini for paid ads and how to use ChatGPT for marketing.