Choosing an AI tool in five steps

Step 1: the task comes before the tool

The most common wasted investment starts with the question of the best tool. It makes more sense to work backwards: note down for a week which tasks actually cost you time, writing quotes, summarising minutes, checking translations, sourcing images.

Only that list tells you which category you need at all. A team switching between German and English every day gains more from a specialised translation tool than from the best chatbot in the world.

Pay attention to frequency rather than excitement: the task that comes up five times a day is worth more than the spectacular one that happens twice a year.

Step 2: settle the deal breakers first

Before comparing features, settle the points on which a tool is disqualified: where data is processed, whether a data processing agreement exists, whether your inputs are used for training, and how it connects to systems you already run.

That order saves a lot of work. It is galling to test a tool for three weeks and then be stopped by the legal department.

For companies in Germany it is also worth looking at the AI Act: depending on the purpose, transparency and labelling obligations arise, particularly for synthetic media and systems touching personal data.

Step 3: test two candidates on a real case

Limit yourself to two tools and test both on the same real task, not on an invented example. Take a real text, a real spreadsheet, a real customer case.

Judge the result by the effort still needed to make it usable. A draft you rewrite by 80 percent is no time saving, however impressive it looks at first.

Write the results down. After ten prompts the impression blurs, and the decision otherwise tips towards whichever tool you tested last.

Step 4: do the maths honestly

On top of a subscription come onboarding, review effort and often a second tool for special cases. Budget for cost per user per month plus roughly two hours of onboarding per person.

Against that stands the time saving you measured in your test. Twenty minutes a day adds up to a substantial sum over a year, but only if the time is actually used for something else.

Be careful with usage-based billing: with video and agent tools, real costs regularly exceed the first estimate.

Step 5: write the rules down before everyone starts

A tool introduced without rules creates trouble. Put on one page which data may be entered, who checks results before they go out, and how AI use is labelled.

Name someone to ask. Experience says the deciding factor is less model quality than whether somebody in the building can answer questions.

And schedule a review after three months: is the tool being used? For what? What is missing? That half hour prevents subscriptions nobody opens any more.

In short: clarify the task, check the deal breakers, test two candidates on a real case, do the maths honestly, write the rules down. Model quality decides last, not first.