In 2024 we wrote about generative AI as a revolution that was only just arriving. Two years later we have enough projects behind us to say something less spectacular but more useful: AI shortens specific stages of the work, not the work itself — and it is very easy to move the time you saved into the revision stage if tasks are split badly between the model and the person.
This article replaces three of our older posts on AI in design and content. It is about what actually works in a design and content team, where AI fails, what a process with a model inside looks like, and which questions — about rights, quality and cost — a company should answer before it starts relying on it.
Key takeaways
Generative AI works best where a task has a clear criterion of correctness and a person can verify it quickly: variants, summaries, drafts, test data.
It works worst where understanding the company's context matters: design decisions, brand voice, factual claims without a source.
The design process has not changed its stages. What changed is how many variants you can look at in each of them and how quickly the first prototype appears.
Before a company builds a process on AI, it needs three decisions: who verifies the output, who owns what was created, and what happens to the data sent to the model.
Generative AI and language models — what we are actually talking about
The word "AI" in design and content covers two different things that are worth separating, because they are judged differently.
Large language models (LLMs) — tools such as ChatGPT, Claude or Gemini — work on text: they understand instructions, summarise, transform, write first drafts, and answer questions based on the documents they are given. Their strength is language and context; their weakness is facts they were not given in the prompt.
Generative AI for media — image, video and voice models — produces graphics, layout variants, visualisations and voice-overs. Excellent for sketches and reference material, still unreliable where brand consistency across hundreds of assets matters.
In practice teams use both at once, but the business value is easier to calculate for the first group: it is language models that automate processes which today cost people hours. One of our automations — an assistant working on HR documentation — cut a payroll department's monthly work from more than two days to three hours. It did not write a single piece of marketing copy.
Where AI genuinely shortens a designer's and a copywriter's work
Tasks with a clear criterion of correctness that a person can verify faster than they could do them:
Variants and alternatives
Ten versions of a headline, three section layouts, five name proposals — to choose from, not to publish. AI does not decide; it widens the field of choice.
Summaries and synthesis
A summary of user interviews, workshop notes, long documentation. The person reads the summary and goes back to the source wherever something does not add up.
First drafts
A wireframe from a description, an article draft from an outline, an email mock-up. A draft to start from instead of a blank page.
Test data and placeholder content
Realistic names, product descriptions, comments in a prototype — instead of "Lorem ipsum", which hides layout problems.
Transformations
The same text in three lengths, a working translation, a change of register from formal to casual. Mechanical work that is easy to check.
Competitor analysis
Collecting and organising what other companies publish — as material for your own conclusions, not as the conclusions.
Where AI fails — and generates revisions instead of savings
The common denominator: tasks where the company's context matters and the model was not given it.
Design decisions
The model will propose any layout, but it does not know that your customers are 60-year-olds on tablets, or that the sales team has spent a year fighting for one field in a form. The decision about what matters is made by a person.
Brand tone and voice
Without concrete examples of your own texts, the model writes like everyone else. The result is content that is correct and bland — easy to recognise and easy to ignore.
Factual claims
Numbers, dates, names, quotes. The model can invent them with complete confidence. Every fact published under your brand needs a source, and the model is not one.
Consistency at scale
A hundred product descriptions in one style, fifty graphics in one visual identity — here supervision often costs more than generation.
Content without a purpose
The most common abuse: publishing texts because they can be generated, not because anyone is looking for them. Google and readers spot it faster than you would think.
Has the design process changed?
No. The stages are the same as a decade ago — in the so-called double diamond version: discover (researching users and context), define (reducing the research to a clearly named problem), develop (ideas and prototypes) and deliver (implementation and collecting feedback). The process is iterative; the most expensive mistake is still a good solution to a badly framed problem.
Two things have changed. First, the number of variants you can look at in each stage before the team commits to something — previously limited by the designer's time, today limited by the time available to evaluate them. Second, the distance from an idea to the first clickable prototype — from days to hours, which moves user testing to an earlier, cheaper moment.
What has not changed: the problem-definition stage requires talking to people, not to a model. Every one of our experiences with projects that "went wrong" comes down to this stage — not to who drew the mock-ups.
A design process with AI inside — how it looks at our studio
The same double diamond, with the model's place marked at each stage:
Discover
Interviews and workshops are run by people. AI transcribes, organises and summarises — and the designer reads the summary with the source open next to it, because the most interesting things get lost in summaries.
Define
Framing the problem and the personas stays with the team. Here the model is a tool for asking questions ("what have we not considered?"), not for giving answers.
Develop
This is where AI gives the most: layout variants, screen sketches from a description, placeholder copy, quick prototypes. The team chooses, combines and corrects — and that is design work, not generation.
Deliver
The design goes into code — at our studio the same team designs and builds, so the mock-up does not drift away from the product. AI helps with testing: it generates scenarios, data and variants for A/B tests.
Want to find out which process in your company is worth automating?
We start with the one process that hurts the most and measure the effect before proposing the next. Describe where your team loses the most time.
Creating content with AI — rules that keep your site from filling up with noise
The biggest risk of generative AI in marketing is not technical. It is the temptation to publish more because it is cheaper. The site fills up with correct, interchangeable texts that answer no specific question — and the whole domain loses in search as a result, including the pages that were good.
The rules we stick to for our own content and our clients':
- Demand first, text second. We write for questions people actually ask (you can see them in Search Console), not for topics that can be generated.
- The model writes the draft, the person writes the text. The draft is an outline, a structure, headline variants. The published version goes through someone who knows the company and answers for every sentence.
- Every number has a source. If it cannot be verified, it comes out of the text — even if it sounds good.
- Experience cannot be generated. A real implementation, a real number from a project, a real mistake — these are the only fragments competitors cannot copy and readers remember.
- We do not re-date content. Changing the date on a text without changing its content does not make it current. We update when the facts have changed.
Three decisions before a company builds a process on AI
The technical implementation is the easiest part. These three questions are worth settling in writing:
Who verifies the output and who is accountable for it — The model bears no responsibility — the person who publishes or deploys does. Without a named owner of the verification stage, the "saving" turns into bugs in production.
Who owns what was created — The legal status of AI-generated content is not settled — it depends on the jurisdiction and on how much creative input the person contributed. For material that is meant to be your asset (a logo, key texts, code), it is worth consulting a lawyer before it becomes the foundation of the brand.
What happens to the data sent to the model — Customer interviews, HR documentation, production data — check whether the model provider uses them for training, where they are stored, and whether a data processing agreement is needed. For implementations on your own documentation we resolve this at the audit stage, not after the fact.
What it costs and how to measure it
The cost of the tools is usually the smallest item. The real costs are the team's time spent learning, changing the process and verifying results — which is why an "AI everywhere at once" rollout almost always ends at the pilot stage. The reverse order works better: one process where you can count the time before and after, a measurable effect, and only then the next one.
Good measures are not "the number of pieces of content generated" but: time from brief to publication, number of revision rounds, the cost of an hour of work saved on a task, and for content — whether anyone reads it and whether it leads to enquiries. If after three months none of these numbers has moved, the tool is not working, no matter how good it looks in a demo.
How we approach AI implementations in terms of process and pricing is described on our AI implementations for business page. When it is worth building your own tools instead of using off-the-shelf ones is the subject of why ready-made applications block business growth.
Frequently asked questions
Will AI replace the designer or the copywriter?+
Which design tasks are best handed to AI?+
Does AI-generated content hurt SEO?+
Who owns content and graphics generated by AI?+
Where should a company start with AI?+
Have a project in mind?
We’ll help you design and deliver it — from strategy to a finished solution.
