Practical guide · AI

AI in SMEs: Start with useful cases instead of tool chaos

Start with a clearly defined process, not the latest model.

Using AI in production rarely starts with a large company-wide rollout. A better first step is a clearly bounded use case with understood benefits, data, quality criteria and risks.

1. The wrong opening question: “Where can we use AI?”

That question quickly produces a list of impressive demos. A more economically useful question is:

Which recurring knowledge, text or data tasks currently take substantial effort – and which parts could be supported in a controlled way?

This keeps the discussion focused on the business process rather than the tool.

2. A good first AI use case has clear boundaries

Good initial candidates are tasks where:

  • The input and desired output can be described precisely.
  • Enough examples or reference data are available.
  • Errors can be recognised or checked.
  • An incorrect result cannot cause uncontrollable damage.
  • The task occurs often enough to generate value.

Typical candidates include structuring, classification, drafting, summarising and supporting standardised communication.

3. Data and permissions come before the model

Before integration, clarify:

  • Which data is sent to the system?
  • Does it include personal or confidential information?
  • Who may start the process and view its outputs?
  • Where are inputs and outputs stored?
  • What data can be used for logs or quality assurance?

A powerful AI model does not replace these organisational decisions.

4. Quality must be measurable before launch

“Looks good” is not sufficient quality control. Production use cases need prepared test cases. Depending on the task, relevant criteria include:

  • Correct classification.
  • Completeness of required fields.
  • Compliance with the expected format.
  • Share of results needing manual correction.
  • Processing time before and after adoption.
  • Cost per transaction.
Practical approach: Prepare 30–100 real test cases, define desired results and compare prompts and models against them before going live.

5. Human-in-the-loop is not a failure

For many B2B applications, allowing AI to make autonomous decisions is not economically sensible. AI can prepare a draft or recommendation while a person reviews the relevant cases. Even partial automation can save substantial time while keeping risks manageable.

6. From a chat window to a business process

The real benefit often appears only after integration:

  1. Retrieve data from an existing system.
  2. Perform the AI task according to fixed rules.
  3. Validate the output.
  4. Flag uncertain cases for manual review.
  5. Pass structured results to a CRM, database or subsequent workflow.
  6. Monitor costs and errors.

AI is then no longer an isolated experiment, but a controlled component of an established workflow.

7. A simple prioritisation matrix

Rate potential use cases from 1 to 5 in each category:

  • Benefit: How much time, quality or speed could improve?
  • Frequency: How often does the task occur?
  • Data quality: Is sufficiently good input available?
  • Controllability: Can errors be detected reliably?
  • Risk: How problematic would an incorrect output be?

High benefit and frequency, good data and controllability, and manageable risk make a strong starting point.

Conclusion

The best AI starting point is not the largest possible project. It is a limited process with demonstrable benefits and controllable errors. From there, the solution can expand step by step.

Discuss an AI use case