Which processes are suitable?
A process is a promising candidate when similar inputs recur, information from documents or systems needs to be combined, and outputs can be validated against rules. Preparatory work is often suitable, especially where people spend much of their time reading, transferring, sorting or requesting information.
Common suitable tasks
- Classify emails and attachments by topic, urgency or document type.
- Extract important details from orders, invoices, data sheets or forms.
- Combine information from multiple sources into a structured work record.
- Identify missing information and prepare focused follow-up questions.
- Prepare drafts, summaries or classification suggestions for approval.
- Derive deadlines and reminders from confirmed document data.
When is AI not the right first step?
Unclear or constantly changing workflows should be clarified organisationally first. Decisions with serious legal, financial or safety consequences must not be left to AI alone. If input data is inaccessible or no one takes responsibility for exceptions, the result is not a reliable operational process.
| Criterion | Favourable starting point | Critical starting point |
|---|---|---|
| Repeatability | Similar cases occur regularly. | Every case is entirely individual. |
| Verifiability | Mandatory fields and rules are known. | There is no way to determine whether a result is correct. |
| Risk | AI prepares; a person approves. | An error would have immediate serious consequences. |
| Data | Sources are accessible and unambiguous. | Data is missing, contradictory or cannot legally be processed. |
How do you start without a major project?
- Choose one bottleneck.Rather than “automate sales”, structure incoming technical inquiries from email.
- Define success criteria.Which fields must be recognised? When must the system ask for clarification? Which errors are unacceptable?
- Collect real test cases.Test normal cases, incomplete inputs and rare exceptions separately.
- Introduce an approval point.An accountable person checks the result before it is used further.
- Measure the benefit.During testing, measure processing time, rework, ongoing AI costs and error rates.
How AI differs from traditional automation
Traditional automation makes sense when every step is clearly defined. AI adds value where emails or documents vary in wording. In practice, a combination often works best: AI interprets the content, fixed application rules check the data, and a person resolves uncertainty.
Links to actual projects
The projects described on pauderer.de combine AI with rule-based checks, local applications or human control points. The project pages give actual status and avoid invented savings claims.
Open project portfolio →Frequently asked questions
How quickly can a first prototype be built?
That depends on data access, interfaces and exceptions. A well-defined workflow can be assessed much faster than an entire company process. Any general timeline without analysing the workflow would be unreliable.
Does a company need to replace its existing systems?
Often not. An agent can complement existing mailboxes, databases, calendars or line-of-business systems through available interfaces. Where no secure interface exists, an alternative handover must be designed.
Is full autonomy the goal?
No. The appropriate level of autonomy depends on risk, verifiability and process maturity. In B2B workflows, controlled preparation is often more valuable than unsupervised full automation.