The creative work is human. AI reads the data.

One rule decides where the line falls. Use AI when it gives a real gain and a low risk. Do not use AI when the gain is small and the risk falls on you.

AI slop is making the internet worse

The web is filling up with text nobody wrote and pictures nobody made. It is cheap to produce, and that is the whole problem: nothing limits how much of it can exist. Search results get thinner. Readers learn to distrust a page before they have read a line of it. Work that took a person a week competes with a page that took nine seconds and knows nothing. And all of it costs real electricity, drawn from a grid that is still mostly not clean.

We are not adding to that pile, and the boundary is mechanical rather than a matter of taste. Nothing our models generate leaves here as finished work. Most of the AI behind our business logic runs on local models, on our own hardware in Austria and Finland. We publish no emissions figure for it yet, because we cannot yet measure it well enough to stand behind a number. You can follow this work on our sustainability page.

Where we use AI, and where we never do

Data, volume and repetition go to script-based automation, and to AI where that is not enough. Judgement, taste and authorship stay with a person. Someone doing the same check for four hours starts to miss things, and a machine does not miss things that way. A better AI model will not move this line next year, because the line was never drawn at what the model can do.

Handed to AI: the data work

• Invoice and report data. AI reads line items, dates and VAT rates out of PDF files. AI collects the monthly numbers with deterministic tools. A person reads and explains the numbers.
• Contract review. AI reads every clause. AI does not get tired on page forty. AI flags what changed since the last version.
• Sorting at volume. AI sorts keywords into topics. AI routes mail to the right person.
• Paperwork into tasks. Our internal AI agents can write it into our task management system. An email pasted to our secure system becomes a Company, a Contact, a Project and a Task, descriptions included, in one step. This used to take a lot of clicks.
• Competitor reading. AI pulls competitor ads from the public ad libraries. AI writes a short brief from them. A person decides what any of it means.
• Proofread and pre-launch checks. AI finds typos and dead links. AI finds an old date in a heading. AI finds a tracking tag that stopped firing.
• Task order. AI is good at ranking the work list. It can pick the task with the best return, not just the nearest due date.
• Idea testing and our own docs. AI argues against a first plan late at night. AI keeps our runbooks correct as our systems change.
• A second check overnight. A local AI model reads your task and project history each night, against our own documentation. It speaks up only when it finds a real, easy fix for your visibility that nobody had time to see.
• Variants of something a person already made. A person writes the headline. AI then produces thirty variants of it to test against, and the person's version often wins. This runs after the creative work, never in place of it.

Kept with a person: the creative work

• The idea. AI is good at the average of what already exists. It is no good at the idea nobody has had yet. A person finds that idea.
• Authorship. There is always a human behind the content we publish. This means the copy, a campaign's main image, the video and the design. A person writes, shoots, animates, edits and grades it.
• Craft decisions. The cut, the tone, the frame, the word that carries the line. These are made by someone with the taste to know why, and they are the part you are paying for.
• Responsibility. One dedicated specialist owns your account. This person answers for what goes out. "The AI did it" is not an answer we give.
• Decisions that bind you. This means quotes, money out and anything we publish live. A person checks each one before it moves.
• Reviews and proof. We do not use AI-written reviews. We do not use a fake customer voice or a person who does not exist. Proof must be real, or it is fraud.
• Rights and ownership. This means licences, copyright, rights clearance and the signature at the bottom of this page. You own the content you buy from us.
• Bidding, watched by a person. We use automated bidding. A person sets its limits. A person checks the conversion data it acts on.
• The relationship. A person writes to you when a project goes wrong. The specialist you book is the specialist who comes to the meeting.

Automation first, and a model only where one is needed

A job can be automated without using an AI model. We ask three questions, in this order:

• 1) Can we automate it safely? If yes, we do. A person doing the same task by hand makes more mistakes than a machine.
• 2) Can we write it as plain code? If yes, we write plain code. A script that checks two exports gives the same answer every time, and we can test it. An AI model asked to do the same task is right most of the time, not every time. On a daily job, that is a fault.
• 3) Does the task need real judgement? Only then do we use an AI model, and only after our own code fetches, cleans and structures the data first. We give the model only the part that needs a decision.

An AI model doing a whole job at once can fail at any step. Clean input and one question make it right more often, and cost less to run. From the outside this looks dull and good: automation that works the same way on Tuesday as it did on Monday.

The ad platforms already run on AI

No person picks who sees your ad. Not you, not us, and not a person at Google or Meta. A machine-learning system decides every single ad view, and has done for years:

• Broad match decides which searches your keyword can answer.
• Smart Bidding sets the price of each auction, thousands of times a day.
• Performance Max and Advantage+ choose the audience, the placement, and which of your assets to show.

You give the goal and the data. The bidding algorithms decide the rest. No setting turns this off, at any budget. So the real question was never whether AI touches your advertising.

Wrong data is dangerous, missing data is not

The bidding algorithms act on your conversion data. Two things can go wrong with that data, and they are not equally bad.

Missing data cannot mislead the algorithm. A visitor who declines consent is simply not measured. This gap is visible, and the campaign does not hide it.

Wrong data is the dangerous kind, because the algorithm believes it. A conversion can fire twice, or on a page nobody reaches. Either one teaches the algorithm to pay for the wrong people.

The algorithms do not pause, and they do not flag the error. They optimise hard toward a false goal and spend the whole budget getting there. This looks like a bidding problem. It is really a data problem underneath.

Open full size

Doing this well, and lawfully, is our job:

• Conversion tracking that measures the real sale, not a page view that resembles it.
• Consent Mode, so we keep only the data we are allowed to collect.
• Audiences built from your own customer data, matched to the platform's users, so the model starts from real buyers rather than a rented guess.

Done well, this pays in both directions. Fewer ads go to people who were never going to buy, so the same budget reaches further. And the person on the other end sees an ad that fits them, not the fourth showing of something irrelevant. That is the digital marketing and measurement work we sell.

How we use your data and AI models

This work is governed by our privacy policy.

We do as much AI work as we can in-house. We run local AI models on our own hardware for this work, and your data stays on our systems.

Some tasks need a very long context window. For these tasks, we may use an outside AI service instead.

We do not send personal data outside our company.

Some tasks need to process personal data with an outside service, such as a customer list. For these tasks, we anonymise the data first, on our own systems. We send only the anonymised data out. We hold our own key to de-anonymise the result, back on our own systems.

An outside AI vendor must not use our or your data to train its models. If we cannot get this guarantee, we do not use that service.

Open full size

AI-assisted client communication

There is a middle ground between full automation and full hand-work. We use it to move an idea from our head to yours, and to lose as little of it as we can on the way.

An idea in one head is not yet a shared idea. Three paragraphs of text leave room for the wrong camera angle, the wrong tone, or the wrong product. So a video script might reach you with AI-made storyboard frames attached, and a concept with a rough mockup rather than a page of adjectives. The form barely matters: text, an image, or a short draft cut. What matters is the translation. We describe the idea to a model in terms exact enough that it understands, and that is most of the work of describing it clearly to you too. If the model understood it, you almost certainly will.

We label this as working material, and none of it survives into the finished piece. A person writes, shoots, animates, edits and grades the finished piece with real production skill. Generate and send is not a production method. You do not need us for that.

AI and copyright: who owns what you get

Fully AI-generated material does not get copyright protection because there is no human author behind it. Nobody can own it, and nobody can stop a competitor using it.

Everything we hand over carries a person's judgement and decisions, so there is something there to own, and as the buyer you own it. This holds whether or not AI touched the production, because the human authorship is what creates the right in the first place. AI is a tool inside the production, the way a camera is.

If your contract rules out generative AI in your creative material, we respect that and use deterministic tools instead. What you own is the same either way.

If we use AI-generated output in your deliverable, we label it clearly. Read more: how we label AI content.

Where this line is not clean

Two limits, stated up front.

• No per-sentence record. We do not send you a record of which words came from an AI model, because we cannot keep that record correct across a long piece of work today. On every deliverable, you can get the name of the author and the name of the person who signed it off.
• An unclear task goes to a person, not to AI. The two lists above cover most of our work, but not every task fits either one cleanly. When we cannot tell which side a task belongs on, a person does it by hand. This is slower and costs more, but it is still cheaper than letting the model try.

The same task, done two ways

The difference between the two columns is not how much AI is involved. It is which end of the task AI is pointed at.

If your audience is sensitive to AI

Some audiences will not forgive generative AI in the creative work they see, and some categories rest on the fact that a person made the thing. If that is your brand, the reaction is not something to argue your customers out of. It is a real constraint on the work, and we would rather write it into the contract than repeat the conversation on every project.

Two questions get confused here, and only one of them is yours to settle.

The automation behind our own business. Project management, mail routing, reporting pipelines, our own pre-launch checks. This is a large part of why the price is where it is, and none of it reaches your audience or your material.

Generative AI in your creative work is your call. Whether a model touches a headline, an image, a script or a cut is a decision you make, and we can work either way.

The usual arrangement

Most brands that raise this want the same shape: analysis tools yes, generation no. Spell and grammar checks, link and tracking QA, reporting, competitor research. Then the copy is written by hand, the assets are made by hand, and nothing generated reaches the delivered material.

This costs more in hours and we quote it that way, so you see the actual cost. It is not a special mode we have to invent for you. It is how the work was done before generative AI existed, and how this company worked for its first years. The craft skills it needs did not go anywhere.

What the contract can guarantee

You can have a flat rule, no generative AI anywhere in the delivered material, or a named list of the steps it may touch. Either one goes into the contract rather than into a briefing note somebody stops reading.

The clause worth asking for is the one about provenance. Delivered material can pick up an invisible watermark or a content credential from further down the production chain: a stock asset, a subcontractor, an editing tool with a generative feature switched on by default. For everything we make ourselves, we can warrant against that in writing.

Where the material comes from a supplier, the guarantee rests on that supplier being straight with us, and this is the limit worth understanding before you sign it. EU law now requires AI-generated material to be declared as such. If a stock library hands us a photorealistic image and declares it as a photograph, there is no test we can run on the file that proves otherwise. We require the declaration in writing, we buy from libraries that give it, and we drop the ones that turn out to be wrong. What we cannot do is promise to catch a supplier who lied, so we do not write that promise into a contract.

AI bought us quality instead of time

The answer surprised us too. When AI took over more of the mechanical work here, our hours did not fall. What changed is where they go. Mistakes now get caught that used to reach the client. A draft gets a sixth check-and-correction round instead of two. A result that was good enough gets the extra afternoon that makes it excellent. So the gain never arrives as a smaller invoice. It arrives as craft work we used to cut first to fit your budget. Switch the generative part off and you pay for those rounds directly, which is a fair price for a constraint your audience holds.

#ai #automation