Paid advertising

AI agent for advertising: what it can do, and what it can't

More and more people ask me whether they can let AI run their Google Ads, Meta or LinkedIn. The answer is yes for a large part of the work. An AI agent for advertising can review accounts, find errors, suggest changes and build ads. The decisions still need to sit with a person. Here's what an agent can realistically do today, and where the line is. And here are the questions to ask an agency before you let an agent into your accounts.
Nicolai Vittrup sitting on a sofa in the office
→ All articles

The difference between an agent and a chat

A chat answers what you ask it. An agent fetches data itself, assesses it and carries out a task. It can pull yesterday's numbers from the ad account and tell you if a campaign has spent half the month's budget in a week.

An agent is also something different from the platforms' own automation. Smart Bidding in Google Ads and Advantage+ in Meta optimise within one platform and towards the goal you give them. They know nothing about your margin, your stock or your brand. An agent can work across channels, from Facebook advertising to Google Shopping, and by your own rules.

What an AI agent for advertising can do today

At Amplify, I've led the work on building our own agent since March 2026. I've written about how we built it on my own blog. Here are the tasks it handles for us today:

  • Morning review. It reviews the ad accounts every morning and points out what needs looking at. That includes the expensive channels like LinkedIn advertising.
  • Answers on numbers. The team asks it in Slack when they need a number, an assessment or a client's history.
  • Suggested changes. It reads the search terms in Google Ads and suggests negative keywords. It tracks spend against the monthly budget and flags when a channel is spending too much or too little.
  • New ads. It builds the ads and keeps them paused until every creative has been checked, including inside the videos.
  • Reports. It prepares the monthly report, so the specialist reads it through instead of building it from scratch.
  • Measurement. It helps set up tracking, so you can see what the ads deliver.

What these tasks have in common is that they repeat. They require precision, and they need doing every time. This is where an agent is good, because it never skips an account and never gets tired of checking the same thing.

What it can't do

A language model is good at what already exists. Your business doesn't already exist in that sense. Your customers, your margins and your stance on brand aren't in any training data.

That's why there are things an agent can't handle alone:

  • Know what isn't in the data. It doesn't know that a product sells out tomorrow, or that you've agreed a campaign with a retailer, until someone tells it.
  • Choose a direction. It can tell you which campaign sells the most. Whether to go after a new market is a decision for you.
  • Feel the brand. It can check copy against your guidelines. Whether an ad feels right for your customers takes a person who knows them.
  • Be accountable. When the budget is spent, a person is responsible to you.

An agent also can't calculate correctly with numbers that are wrong. If tracking counts wrong, it draws the wrong conclusions with great confidence. At Amplify, the rule is simple: no conclusions without real data.

My experience is that the limit is rarely the model. Every time we've hit a wall, it came down to too little context, too few examples, or us using the model wrong.

Where a person must approve

The most important thing we've built into our agent is a stop. It's free to read, gather, analyse and point out problems on its own. Anything that changes something in an account must be approved by a person first.

In practice, approval is required before:

  • a new ad, a piece of copy or a creative goes live
  • a budget changes
  • a bid strategy or a campaign objective changes
  • a campaign that's selling is paused
  • anything in tracking or consent changes

The reviews have paid off. Six new ads sat idle for a whole day without a single impression, even though everything showed green in Meta. The review found the cause, the ads were rebuilt, and they were running the next morning. Elsewhere, Meta received more than four purchase signals for every real order, because two systems were sending the same purchase. That was switched off, so the platform is once again learning from real numbers. In both cases, a person approved the fix.

How to make the agent good

An agent only knows what you've told it. Most of our work with the agent is writing down how we think. When the team corrects it, the correction becomes a rule it uses next time.

If you want to use an agent in your own advertising, it needs this from the start:

  • The goal for each channel, and what a customer or a lead is worth to you
  • Your brand guidelines, including the words and images you never use
  • Seasons, campaigns and stock levels that affect the numbers
  • The history: what you've tried, and what worked
  • The limits: what it may suggest, and what it must never touch

When people get worse answers from AI than we do, the difference is almost always the context. The model is the same.

Common pitfalls

  • The platform's goal becomes your goal. If you let Google or Meta optimise for clicks or cheap leads, that's exactly what you get. Set the goal by what makes money.
  • Autopilot without a stop. An agent that can move budgets on its own can overspend in a short time if the data plays up.
  • Wrong measurement. If conversions are counted twice, the agent chases sales that never happened. Check tracking before you trust a single suggestion.
  • Too little to work with. An agent without your rules gives generic advice. It looks like the advice everyone else gets.

Ask your agency about this

If your agency uses AI in your accounts, or you're considering one that does, ask these questions:

  • Who approves before anything goes live, and can you see who it was?
  • What can the agent do without approval?
  • What data does it base its assessments on, and what does it do when data is missing?
  • How does it learn your brand, and what happens when it gets it wrong?
  • Can you see a log of what it suggested and what actually happened?
  • Who is accountable if it goes wrong?

A good agency gives a concrete answer to all of them. If the answer is just "the AI handles it", ask again.

Frequently asked questions about AI agents for advertising

What's the difference between an AI agent and a chatbot?

A chat answers what you ask it. An agent fetches data itself, assesses it and carries out a task. It can pull yesterday's numbers from the ad account and tell you when a campaign has spent half the month's budget in a week.

Can an AI agent change things in the ad account on its own?

Not without a person saying yes. With us, the agent is free to read, gather, analyse and point out problems on its own. Anything that changes something in an account must be approved first.

What's the difference between an agent and the platforms' own automation?

Smart Bidding and Advantage+ optimise within one platform and towards the goal you give them. They know nothing about your margin, your stock or your brand. An agent works across channels and by your own rules.

Why does AI give better answers to some people than others?

Almost always because of the context, rather than the model. An agent only knows what you've told it: the goal for each channel, your guidelines, seasons, history and limits. Without that, it gives generic advice.

Your next step

Start with one task that repeats, such as the morning check of your accounts or the monthly report. Write down what a person looks for today, and when they step in. That list is the agent's first set of rules.

Let the agent only make suggestions for the first month. Once the suggestions hold up, it can carry out what a person has approved.

If you'd like to see how this could look in your own accounts, get in touch. We'll look at your advertising together with you and point out the first tasks an agent can take on.

Want help with this?

→ Book a chat