AI in marketing: how we use it every day, and where we don’t

What AI does for us, and what a human does

We use AI for the work that used to take the most time and the least judgement. The decisions that cost a client money, we make ourselves. This is how the work is split:

The last point is deliberate. We let AI suggest, but not press buttons that change something already running in a client’s account. A mistake there costs real money before anyone notices.

Ad production: where AI saves the most time

Motion ads with Norwegian voiceover are the clearest example. We make them without a designer, without filming and without After Effects, using six tools:

• Claude Code writes the script, plans the scenes and codes the animation.
• Remotion turns the code into video, both 9:16 and 1:1 from the same file.
• ElevenLabs creates the voice, music at the exact length, and afterwards checks that the voice said what is in the script (speech to text).
• ffmpeg adjusts the audio level and pulls stills for review.

What AI doesn’t do is decide what the ad should say. That is where results are decided. And several of the rules we follow exist because AI got it wrong the first time:

• The older multilingual voice model drifted into Danish on Norwegian text. We now use eleven_v4 with language code “no”.
• Lots of short sentences made the voice sound choppy. We now write few, long sentences.
• We make 3 to 6 versions of the voice and check each with speech to text, but always listen ourselves before anything is published.
• Drawn mockups were rejected. Real screenshots and product images as the base work better.

The full recipe, with the templates we use, is in how we make motion ads with AI. We also make statics and UGC in-house, but the UGC videos have real people in front of the camera. More on creative production.

Analysis: AI sorts, we decide

Every week a script pulls the search terms from the Google Ads accounts we manage. AI sorts them and suggests negative keywords, placements to exclude and ads that underperform. The suggestions come to us, not straight into the account. An example of what the review finds: in the move-out cleaning account, the search “billig flyttevask oslo” (cheap move-out cleaning Oslo) cost NOK 298 in June and July without a single lead. It is now a negative keyword.

The numbers in the tables on arcads.no, like the conversion rates in the glossary, are pulled the same way, straight from the Google Ads API, and read by a human before they are published.

Google’s own AI campaigns, and when they fool you

Much of “AI in marketing” already happens inside the ad account, whether you want it or not. Three things are now run by Google’s machine learning:

• Performance Max chooses where ads appear: Search, Shopping, YouTube, Gmail and the Display network.
• Broad match lets Google show the ad on searches it thinks mean the same as your keyword.
• Smart bidding sets the bid per auction based on how likely Google thinks a conversion is.

All three can work well. A Performance Max campaign set up for retargeting delivered leads at NOK 116 for move-out cleaning in Oslo, against NOK 211 on search, and NOK 310 against NOK 467 for the moving company (9 June–6 Sept 2026). But the same campaign also showed ads in mobile games and in children’s videos on YouTube. Nobody there books move-out cleaning. That is why we check the placement report regularly and exclude what doesn’t belong.

Broad match has the same weakness: Google interprets the search, and the interpretation can be wrong. It is the search terms report that shows what you actually pay for, not the keyword list.

What we don’t use AI for

• Fake customers. We don’t make AI images or AI video of “happy customers”. The people in our ads are real, and we never show a person who isn’t the one speaking.
• Fake reviews and quotes. The quotes we use are real and named, like “We’re getting more leads than ever!! Amazing” from Elliot at Badstugutta.
• Made-up numbers and promises. AI will happily write “10x more customers” if you let it. Every claim in an ad must be on the client’s own website or in our numbers, and we fact-check before anything is published.
• Before and after images for clinics. Neither real nor generated.
• Decisions about money. Budgets, bids, pausing and deleting we do ourselves, after reading the numbers.

The reason is simple: an ad that promises something the business can’t deliver brings expensive clicks and disappointed customers. And an AI-generated “customer” who gets exposed costs more trust than it ever won.

Should you use AI in your own marketing?

Yes, for drafts and groundwork, with one rule: nothing goes out until a human has read it, heard it or seen it. Start with whatever takes the most time for you today. For many, that is ad copy and images in several formats. Then go through the campaigns Google already runs with AI, and check placements and search terms.