AI in marketing: where it delivers measurable results and where it does not

What we mean by AI in marketing

When we say AI in marketing, we mean mainly generative AI tools: assistants that write and summarise text, image generators, transcription and analysis tools, and automations that use such models. It is not one technology but a set of tools with different strengths and limits.

A few useful terms:

• Large language model (LLM): the system that generates text from patterns learned in data. It does not “know” facts the way a person does; it predicts plausible continuations.
• Prompt: the instruction you give the model. Its quality directly affects the quality of the result.
• Hallucination: an answer that sounds convincing but is false or invented (figures, quotations, sources).
• Automation: a workflow that runs by itself according to a rule. AI can be one step in it, such as classifying a message, but the workflow itself remains a rule that needs human oversight.

What separates this from hype is simple: we look for hours saved, lower costs or better results that can be verified, not for an impression.

Where AI brings clear gains

In our own work, and in what we see with the teams we train, gains appear where the task is structured, repetitive and easy for a person to check.

Research and first drafts

An AI assistant can gather, in a minute, the questions customers tend to ask, the possible angles on a topic or an outline for an article. As a starting point it is valuable. As a finished product, rarely.

Adapting content between channels

From one good article you can quickly produce a series of social posts, a short video script, and an email subject line and preview text. The reformatting work that many people put off becomes much easier, provided you reread each version.

Analysis and summaries

Reviews, comments, call transcripts, long reports: AI can group them by theme and highlight recurring points. The result is a good working hypothesis, not established fact, so check it against a handful of real examples.

Images and variations for testing

For mock-ups, ad variations and draft material, speed goes up. For final brand imagery, check rights, consistency and authenticity carefully.

Simple automations

Sorting incoming messages, drafting replies to frequent questions, tagging contacts: these are tasks where an occasional error can be caught by a person, so the risk is manageable.

Which tasks suit AI, and what control each needs

The table is an indicative guide, not a fixed rule. How much AI helps depends on the tool, on how you use it and on the quality of the material you start from.

Where AI does not help, or misleads

Strategy. A model can produce a nicely worded marketing plan, but it does not know your margin, your team’s capacity, your local competition or what has worked before. Treat it as a sparring partner that asks questions, not as a decision-maker.

Brand voice. Left unedited, AI text tends to sound the same everywhere. Differentiation comes from your examples, your way of explaining things and what only you know about your customers.

Accuracy. Models can invent figures, study names and quotations with surprising confidence. Any number, and any statement about law, prices or products, should be checked against the primary source.

Content at scale. The temptation to publish hundreds of generated pages usually leads to mediocrity. Google’s guidance says, in essence, that content is judged on how helpful it is rather than how it was produced, but content created at scale mainly to manipulate rankings is treated as spam. Volume without value is not a strategy.

Relationships. Customers can often tell when replies are automated, especially in delicate situations. An unhappy customer wants a person who listens, not a well-packaged text.

How to measure whether AI is really helping

Without measurement, “it feels faster” is the only evidence, and first impressions can mislead. A simple test, in five steps:

1. Pick one concrete task, for example writing the monthly newsletter.
2. Record the baseline: how long it takes now, what it costs and the quality of the result, judged against a simple rubric: accuracy, tone and usefulness.
3. Run a pilot of four to six weeks with AI in the workflow, with the same person and the same criteria.
4. Compare: time saved, number of corrections and, where relevant, results (opens, clicks, enquiries).
5. Include the real cost: subscriptions, time spent checking and correcting, team training.

If the time saving disappears in editing, the process is not yet good. If the time drops and quality holds, you have a real gain. Business outcomes such as enquiries and sales belong in your monthly report, as the guide to marketing KPIs explains.

A healthy process in five steps

1. A person sets the direction: audience, objective, message and limits (what is never said).
2. AI produces variations from clear instructions and your own material, such as examples of good copy and a tone guide.
3. A person selects and edits: chooses, cuts, adds their own experience and checks the facts.
4. Publishing follows an approval rule: who signs off, and for which types of content.
5. Results are measured and fed back into the instructions: what worked becomes an example, what failed becomes a prohibition.

Repeating this cycle is worth more than a “perfect” prompt. Teams that get stable results usually have written instructions, good examples to hand and one person responsible for each stage.

Data, GDPR and internal rules

The most common problem is not technical but a matter of discipline: people pasting customer lists, contracts or internal data into a public tool. Decide in advance what may and may not go into an AI assistant.

• Do not enter customers’ or employees’ personal data into tools you have not assessed.
• Check each tool’s privacy settings and terms of use, including whether your data is used for training.
• Keep a short register of the tools used and their purpose.
• Say clearly when material has been generated or altered by AI where failing to do so would mislead the public, for example realistic images of people who do not exist.

This section is general information, not legal advice. GDPR (Regulation (EU) 2016/679) applies to personal data, and the EU AI Act (Regulation (EU) 2024/1689, as amended in 2026) applies in stages: transparency duties under Article 50 apply from 2 August 2026, while most high-risk obligations have been postponed to December 2027. Check your own situation and the current timetable with a lawyer or your data protection officer. For a model set of internal rules, see our guide to an AI policy for a marketing team.

A hypothetical example: an online shop with a team of three

Picture a small online shop with one marketing specialist, one support employee and one product administrator. It is not a real case, only an illustration.

First attempt: “let’s generate descriptions for every product”. The result is clean text, but identical in structure and sometimes wrong in the details. The fix is to give the AI the real specification sheet and a tone guide, let it produce a draft, and have a person check the specifications. Time falls and quality holds.

Second use: a monthly summary of support messages. The AI groups the questions by theme, and the team discovers that a single piece of information is missing from the delivery page. The page is corrected and the number of questions drops. That is the kind of gain you can show in a report.

Third: draft replies to frequent questions. A person approves the drafts before they are sent, especially for complaints. The rule is simple: AI drafts, a person sends.

Common mistakes

• Publishing generated text directly, with no edit and no fact-check.
• Vague prompts (“write an article about marketing”) and results to match.
• Using AI for every task, including those that need judgement and a relationship.
• No data rule: confidential information in unsuitable tools.
• No measurement. You do not know whether you saved time or simply moved the work into proofreading.
• Expecting a magic button. Results come from a process, not a subscription.
• Ignoring the team. If people are not trained, they use the tool by trial and error and results vary from day to day.

How we approach it

We use AI in our own work for research, drafts, adaptations and summaries, and we keep decisions, checking and brand voice with people. When we work with a team, we start by mapping how they work today, then choose two or three tasks with clear potential, test them and only then expand. The mapping can begin with the digital maturity scanner, and the next steps connect naturally to a 90-day marketing plan.

If you want your team to work this way, our AI and marketing training starts from your company’s own cases rather than generic slides. We apply the same principles in social media, email marketing and the improvement of existing pages. In conversion rate optimisation we use AI mainly to analyse hypotheses, not to deliver the final verdict.

In short

• AI helps most with repetitive, structured tasks that are easy to check.
• Strategy, brand voice and facts remain the job of people.
• Measure with a pilot: baseline, period, comparison, real cost.
• A repeatable process (direction, variations, editing, approval, measurement) beats the perfect prompt.
• Set rules for data, transparency and compliance in advance.