Question
Clearly sourced
Expert knowledge for digital decisions
Which Tasks Are Suitable for AI – and Which Are Not?
Short answer
Tasks that are suitable are those where an assessment is sufficient and a mistake can be corrected easily: sorting, summarizing, designing, extracting. Tasks that are unsuitable are those where accuracy is essential and a mistake is costly: calculating, making legal judgments, making final decisions.
The Crucial Question
Not "Can AI do this?", but "What is the cost of a mistake, and will someone notice it?" A wrongly sorted inbox can be corrected in ten seconds. A wrongly calculated invoice may never be noticed.
Well-Suited
| Task | Why |
|---|---|
| Classifying and forwarding texts | Errors are immediately visible and inexpensive |
| Extracting documents (OCR and extraction) | Results can be verified against the original |
| Summarizing with source reference | Statements are verifiable |
| Writing draft responses | A person approves them |
| Searching in own documents | The source is verifiable |
Poorly Suited
| Task | Why |
|---|---|
| Calculating | Language models do not calculate; they estimate. This belongs in code. |
| Legal or medical assessment | Errors can cause significant damage |
| Final decisions about people | Article 22 GDPR, and it is unfair |
| Tasks with a single correct answer and strict verification | A rule is cheaper and more reliable |
The Most Common Misconception
Using AI for something that can be handled with a simple rule. A regular expression pattern that recognizes an invoice number is faster, cheaper, and one hundred percent reliable. Using a language model for this is more expensive and less effective.
Rule of Thumb: If you can write the rule in one sentence, use the rule.
Key facts
- Guiding Question
- What is the cost of a mistake, and will it be noticed?
- Rule of Thumb
- What can be formulated as a rule does not belong in the model
Sources
All external claims are backed by traceable sources.-
01
Datenschutz-Grundverordnung (Verordnung (EU) 2016/679) Amt für Veröffentlichungen der EU