A marketing team buys several AI tools. Each can create text, images or a report. After three months, however, nobody knows how much work has actually disappeared. There are new subscriptions, new tabs and an obligation to check more outputs.
I came to AI for the opposite reason: I wanted to reduce repetitive work. In the DreamROI project we worked on personalisation and data predictions, alongside text generation. This experience taught me to separate an attractive demonstration from a process that pays back economically for a company.
Three types of use, three different conditions
Generation
A model prepares text, images, an advertising variant, an outline or a summary. Deployment is quick, but costs often hide in prompting, checking and corrections. The benefit grows when it works with approved materials and a repeatable format.
Automation
A system connects steps: it takes source material, creates a draft, sends it for approval, publishes it and saves the result. Some of it need not use AI at all. Value comes from removing manual transfer and waiting.
Prediction and recommendation
A model estimates the likelihood of a response, ranks offers or selects the next step. It needs quality history, enough signals and a way to test continuously. For a small business, an ordinary rule may be cheaper and easier to understand.
How to recognise a suitable process
Before choosing a tool, rate the task with six questions:
- How often does it repeat? An hour once a year is not a good candidate for a complex integration.
- Is the input similar? A stable template automates better than a different strategic problem every time.
- Can we check the result? Without a quality criterion, we cannot recognise improvement.
- What does an error cost? A typo in an internal outline and a wrong campaign price are not the same risk.
- Do we have the required data and permissions? A broken CRM record will not be saved by a smarter model.
- Who owns it? Someone must monitor, update and be able to switch off the workflow.
The best candidate is frequent, relatively stable, easy to measure and has a low cost of error. Typical examples are preparing variants from approved material, categorising feedback or creating a first draft of a regular report.
Where AI commonly helps today
- Content: outlines, variants, format changes, interview summaries and editorial critique. Without first-hand experience it produces only generalities.
- Email marketing: drafts from approved content, segmentation and recommendations when the data is sufficient. More is covered in AI in email marketing.
- PPC: creative variants and query analysis. A person must control budget, measurement and limits.
- Customer feedback: sorting large numbers of responses and suggesting topics. Sensitive data needs an approved environment.
- Reporting: commentary on prepared data and finding deviations. A model must not replace the source table or invent missing values.
- Internal knowledge: searching approved documentation. The benefit depends on current documents and permissions.
Five costs a demo does not show
1. Preparing data and source material
A tool may be cheap, but cleaning a catalogue, unifying campaign names or anonymising documents takes time. Include this cost in the pilot.
2. Human review
The higher the cost of an error, the more experienced the reviewer must be. A senior hour may cost more than generation itself. Measure time to approval, not time to the first draft.
3. Integration and maintenance
APIs change, permissions expire and input formats move. A workflow without an owner slowly stops working or starts producing errors.
4. Supplier dependency
Price, limits and service terms can change. For an important process you need data export and a plan for an outage or product termination.
5. Reputation and privacy
One invented case study or leaked client document can outweigh months of small savings. Decide in advance which data may not be sent to the service and how content requiring expert approval is marked.
How to build a pilot that proves something
- Choose one process. For example, turning an expert article into a newsletter and two posts each week.
- Measure today’s state. Time, number of corrections, error rate and business metric.
- Set boundaries. Approved sources, prohibited claims, required review and shutdown.
- Keep a comparison. Produce some outputs with the original method or compare the same period.
- Count all time. Prompting, integration, review, corrections and maintenance.
- Decide after a limited period. Continue, adjust or end the pilot.
Years ago I estimated significant time savings for some tasks. Today I would not promise a universal percentage. Savings depend on input quality, task frequency, required accuracy and review cost. An honest pilot is stronger than an impressive estimate.
When to leave AI aside
Do not deploy it merely because competitors use it. It makes no sense for a rare task without a clear output, a process with broken data or anywhere the company cannot assign responsibility for an error. I would not delegate a strategic decision about brand positioning, price or a sensitive client to a model.
I would also fix measurement first. Without basic information about enquiries, you cannot tell whether AI is optimising the business or merely creating more content.
The first step is not a licence
Take one repeated marketing process and map its steps. For each, write the time, input, output, errors and owner. Only then look for whether a rule, integration, generative model or combination can help.
For practical text work continue with the AI text-creation workflow. For content that should work in search as well, see How to become a source for Google and AI answers. If you want to calculate a pilot for your process, get in touch via contact.