AI in marketing: finding the right first step

How is AI used in marketing?

A helpful classification follows the system’s role in the workflow, not product names:

Assist: A system creates a first draft, summarises material, organises feedback or proposes variants. A person remains the author, reviewer and approver. Typical examples are briefing drafts, language variants or structuring research.

Detect: A system finds patterns, groups or deviations in existing data. In marketing this can be an anomaly in campaign metrics or recurring themes in customer feedback. A pattern is not yet a cause. The team must check whether the data is complete, comparable and suitable for the question.

Forecast and recommend: A system estimates a future value or prioritises a next action. This includes forecasts, scores and recommendations. They can support a decision, but they replace neither the definition of the business goal nor the review of side effects.

Act: A system changes a running process within defined limits. The closer the action is to budget, price, customer access or public communication, the more important approvals, logs, thresholds and a way to stop or reset become.

These roles can be combined in one flow. A writing assistant might simply produce a draft. An integrated system could also enrich the draft with customer data, serve variants and use the results for the next selection. The second variant is not simply “more AI”. It has different data flows, dependencies and risks. The guidance for Swiss SMEs also names varied examples such as text generation, forecasts and decision support. That is not a formal marketing taxonomy, but it shows the breadth of possible capabilities.

These levels differ in their data, risks and maturity. A writing assistant is not the same as an agent with system access. A campaign forecast is not a business decision. An automated handover is ready for routine use only when failures, duplicate records and recovery are handled.

How does marketing automation differ?

Marketing automation describes the repeatable flow: trigger, condition, action, exception, responsible person and evidence. AI can support one step, but it is not needed for every automation. A fixed rule is often easier to test and explain. AI can help with tasks that involve language, pattern recognition or uncertain inputs when its contribution improves the result and the system stays within defined limits.

Start with the process on paper. Mark the inputs, systems, personal data, decisions, waiting times and manual handovers. Do not automate an unresolved process; otherwise you scale its mistakes.

Where can AI make a useful contribution?

An application area is interesting when the task recurs, enough material is available and a better result would be recognisable. In marketing, six work fields can be distinguished:

1. Research and planning: sort material, collect questions or prepare a briefing. Source verification stays with the team.
2. Content and creation: produce drafts, variants or editing suggestions. Brand fit, facts, rights and approval remain human tasks.
3. Media and campaigns: merge signals from campaigns, flag anomalies or make limited proposals. Platform automation must be kept separate from independent impact measurement.
4. Customer interaction and personalisation: adapt content or next steps to rules and signals. Here the demands on purpose, data quality, consent, fairness and fallback paths rise.
5. Analysis and measurement: structure large amounts of observations or prepare forecasts. A plausible comment proves neither causality nor additional revenue.
6. Marketing operations: support recurring handovers, classifications and quality checks. The system needs a responsible person, error handling and a documented manual alternative.

This list is a map, not a recommendation to implement everything. The SATW guidance for Swiss SMEs recommends a needs analysis, small pilot projects and clear rules and responsibilities. Suitability can only be judged for the specific case.

Which use cases should be assessed first?

A good first use case is frequent enough, clearly bounded and reversible. It has a responsible person, available test cases and measurable value. Examples are structuring a briefing, classifying internal requests, summarising a reporting finding or preparing a proposal.

Weak first use cases access sensitive data directly, publish without review, control large budgets or promise an impact that cannot be measured without a control group. Rank every candidate by benefit, data risk, harm on error, explainability, integration effort and reversibility.

The most important decision comes before the AI

A robust use case starts with a sentence without a technology term: “We need better, faster or more reliable results for …”. Only then does the question follow whether rules, classic automation, better data, process changes or AI can help. Writing support inside an existing tool is organisationally different from a system that joins several data sources and triggers customer actions.

Seven questions suffice for a first check:

• Which concrete task or decision should improve?
• Who is responsible for the outcome and for harm?
• What is today’s baseline performance, including time, quality and errors?
• Which data is used, and may it be used for this purpose?
• Which errors are tolerable, which are not?
• What does a person check before publication or action?
• Under what condition is the use stopped?

If a team cannot answer these questions, it is too early for automatic action. Manual assistance can still be a suitable learning step, provided no inadmissible data is entered and the outputs are reviewed.

Which data and decision boundaries are needed?

Record the data source, purpose, legal basis, retention, recipients, model provider and deletion path. Check whether personal or confidential information is really necessary. Pseudonymisation and hashing do not remove every duty. The FDPIC notes that Swiss data protection law also applies to AI-assisted processing; a concrete assessment remains case by case.

Then define the system’s scope. May the system only create drafts, propose priorities, write data or send external messages? Which steps need human approval? Which inputs are forbidden? When does the workflow stop automatically?

What people must still decide

AI can generate options, but it holds no mandate of its own for corporate strategy. People must keep at least five kinds of decisions.

Goal and priority: A model does not know which trade-off is acceptable for the company. More leads can mean worse leads; faster production can weaken brand quality.

Evidence: Fluent text can be wrong. A correlation can sound like an explanation. Specialists must check sources, measurement method and uncertainty.

Rights and relationships: Customer data, internal documents, images, voices and third-party texts must not enter a system or a campaign arbitrarily. Contracts and terms of use belong in the review.

Approval and responsibility: Who may see a draft, change a budget or approach a customer? These rights must be narrower than the technical capability of the tool.

Exception and stop: A team needs a way to halt an action, correct faulty outputs and return to a manual flow.

Human-in-the-loop therefore does not mean clicking “OK” briefly at the end. It means that a qualified person has the necessary information, real decision-making power and enough time for the review, in good time.

How are tools compared?

Compare tools against the use case, not against a feature list. Check:

1. data access and tenant separation;
2. model provider and further data processors;
3. logging and export;
4. roles, approvals and technical limits;
5. integrations and failure behaviour;
6. testability, versioning and rollback;
7. price at realistic volume;
8. exit and portability.

A tool can impress in a demo and be unsuitable in operation. Test it with synthetic and adversarial cases before real data or automatic actions follow.

When is an implementation partner needed?

A partner pays off when marketing knowledge, technical setup and adoption in the team have to come together. Agree systems, deliverables, acceptance and support per project. That way it is clear from the start what your team can actually do better after the introduction.

From idea to a controlled test

A low-risk entry can take four steps. First the team describes the task, the users, the data and the baseline. Then it limits the pilot: one channel, one segment, one content type or one internal process. In the third step, quality, time, errors and necessary rework are measured. Finally an explicit decision follows: stop, change, test again or put it into routine use.

A pilot should inform a decision about adopting the process, rather than merely show that the tool works. Record the baseline values, test cases, quality criteria, time budget, data-protection review, extent of human review, error classes and stop criteria. Start in parallel operation: the system produces a result, but the existing process stays authoritative and is not changed automatically.

Each pilot should have a short protocol: tool version, permitted data, test cases, evaluation criteria, known limits, responsible people, approvals and incidents. This record shows what was actually tested. This documentation is not bureaucracy for its own sake. It makes learning and a later decision possible.

Compare quality, processing time, correction effort and the relevant business metric. Time saved alone does not establish an overall benefit; compare quality and total effort as well. One positive case is not a scaling decision. Document which cases deliberately stay outside the pilot.

This sequence prevents two common confusions. An impressive demonstrator is not yet reliable operation. And saved drafting time is not business value if the review takes longer, risks rise or the marketing goal stays unchanged.

What AI in marketing does not prove

Using a well-known model proves no strategic maturity. A large number of generated variants proves no additional impact. A score does not prove a sale. A recommendation from the advertising platform is not an independent assessment. And a provider demonstrating generative text has not thereby shown competence in data integration, measurement, security or organisational change.

That is why a company should keep three levels apart: output quality describes whether the immediate result is usable. Process performance describes whether work becomes faster, more reliable or cheaper overall. Business impact describes whether a relevant goal is reached better. Each level needs its own comparison and can turn out differently.