Back to Blog

Care Documentation: What May AI Do, What Must the Professional Do?

7/2/2026

Anyone working in German disability support services knows the imbalance: the Federal Participation Act (BTHG) was meant to strengthen participation. In practice, it mainly multiplied documentation duties. Case notes, development reports, goal reviews. At the same time, the AI industry promises to handle all of it “at the push of a button”. Together, this creates an understandable mix of hope and distrust among managers.

After six years of my own practice as a care professional and my current work as a developer, my answer is a clear division of labour. It fits in one sentence: AI may write, the professional must judge.

What AI does well and may take over in good conscience

  • Drafting instead of a blank page: Turning bullet-point case notes into a structured report draft that follows the funding body’s template. That is legwork, not professional judgment, and it is exactly where professionals’ evenings and weekends disappear.
  • Structuring and summarizing: Condensing six months of case entries into the relevant developments, ordering them chronologically, spotting duplicates.
  • Ensuring consistency: Uniform professional language and a complete structure, whether the report is written on Monday morning or Friday evening.
  • Eliminating double data entry: Strictly speaking that is not AI but integration (APIs, n8n), often the biggest lever before any language model enters the picture.

What the professional must not hand over

  • The professional assessment: Whether a participation goal is achieved, adjusted, or dropped is a pedagogical decision. A language model can sound plausible and still be wrong, because it does not know the person.
  • Goal planning: Participation goals emerge in conversation with the people receiving support, not from text statistics. Participation is the core of the BTHG; an AI that “suggests” goals undermines exactly that.
  • Responsibility: Humans sign. Every draft must be reviewed, corrected, and professionally accounted for. The draft is raw material, not a result.
  • Quality control over oneself: Whoever waves AI drafts through unread has not automated anything. They have abolished quality assurance.

Client data in disability support services is social data and frequently health data (Art. 9 GDPR). In practice this means:

  1. No sensitive raw data to US cloud services. Copying case notes unchanged into an arbitrary chatbot is, as a rule, a reportable data protection breach.
  2. Self-hosting or EU processing with a data processing agreement. Language models can run on your own infrastructure today. The data never leaves the building.
  3. Pseudonymization as the default: Names and identifying details do not belong in the prompt if it can be avoided.
  4. Transparency and a deletion policy: If someone asks where which data sits and for how long, you must be able to answer, including clients and funding bodies.

How to recognize serious providers

Distrust anyone promising “90% time savings” or claiming the AI writes the report “ready to send”. Serious providers draw the line themselves: draft yes, judgment no. They explain where the data is processed. And they build in an approval step instead of optimizing it away.

What is realistic: pure writing time per report drops from hours to the time a thorough review takes. That is not a revolution. It is reclaimed hours per week and professional that flow back into working with people. Which is what the BTHG was made for.

The use case Care Documentation with AI Assistance shows what this looks like in practice. Whether it fits your organization is something we can settle in a free 30-minute call.