Reference letters. Clear. Fair. Compliant.
Product · AI assistance

How we use AI — and where humans stay in charge

ZeugnisPilot uses ZeugnisPilotAI as a writing and review tool. The model formulates and checks. Assessment, responsibility and final release stay with the employer.

  • ZeugnisPilotAI on a frontier model
  • Hosting in the EU
  • No model training on your data
Concrete and honest

What the AI does for you

Formulates sentences

From your keywords on tasks, strengths and conduct, ZeugnisPilotAI formulates a coherent, linguistically clean continuous text.

Applies tone

Formality, length, goodwill and task detail are controlled via four sliders. The AI adjusts word choice and sentence lengths accordingly.

Checks for codes

In the compliance check the AI identifies typical Swiss code patterns and proposes substantiable plain-text alternatives.

Structures Kompass output

With the "Kompass" standard, the AI converts your inputs into the structured JSON schema with 5 performance fields, contribution cards and suitability profile.

Deliberate boundaries

What the AI does not do

No own assessment

The AI does not give the person a grade. It translates your assessment into words. If you give grade 4, the AI formulates as "often exceeded requirements" — not "clearly exceeded".

No data invention

If you do not name a project, the AI invents none. It then formulates neutrally at the role level — you add concrete examples in the editor.

No HR responsibility

The AI is a writing and review tool. The professional, legal and entrepreneurial responsibility for the issued reference lies with the employer.

No personality analysis

The AI does not analyse personality. It reformulates observable behaviour that you as employer have documented.

Three protection layers

Data protection and AI training

  • No training use: your inputs are used neither for ZeugnisPilotAI nor for external model training (contractually guaranteed).
  • Anonymisation in the prompt: we send only placeholders ([FIRSTNAME], [LASTNAME]) to the model. Real names are inserted server-side after generation — the AI never sees them.
  • EU processing: ZeugnisPilotAI is accessed via the EU region. Swiss data hosting for the frontend and backend layer (europe-west6 Zurich). We list the model provider used transparently in our sub-processor table.
Clear allocation

Who is liable?

The issuing company bears the HR responsibility for every reference. ZeugnisPilot is a writing tool that converts your own assessments into a linguistically clean text. Before publication, every draft passes through the compliance check and editor — final release is always a human act.

This separation of responsibility is documented both in our T&C § 3 and in the data protection notes.

FAQ

Frequently asked questions about the AI

Which AI is used?
ZeugnisPilotAI — a finely tuned language AI based on a current frontier model with explicit focus on reliability and refusal behaviour for risky requests. Set up in a GDPR-compliant way, processing in the EU. We list the model provider used transparently in our sub-processor table.
Are my inputs used for AI training?
No. Your inputs are used neither for training our AI nor for external model training (contractually guaranteed). We also pass the model only placeholder names ([FIRSTNAME], [LASTNAME]); real personal names are inserted server-side after generation.
Who is liable if the AI writes something wrong?
HR responsibility stays with the employer. That is precisely why every draft must pass through compliance check and editor — you see the text before it leaves the building. We are a writing tool, not the decision-maker.
What happens if the AI rejects a request?
We detect refusals automatically (pattern matching on typical refusal phrases) and give you a clear error message with a suggestion to adapt the input. Common cause: too unspecific keywords or too sweeping negative statements.
Can we adapt the prompt?
On the Enterprise plan, yes — you can extend the system prompt for your tenants (e.g. with industry-specific vocabulary or a fixed closing paragraph). For Pro and Starter, the prompt is fixed and best-practice optimised.
How long does a generation take?
Typically 8–15 seconds for the draft. Compliance check additionally 6–12 seconds. In total: less than half a minute of active AI time per reference.
What about hallucinations?
We minimise this through two measures: (1) Structured prompting — the AI receives a clear separation between client inputs and what it may itself formulate. (2) Compliance check as a second instance, which checks consistency with the inputs. Even so: every draft must pass through your eyes before it becomes a PDF.

AI with responsibility — try it yourself

You see every generated sentence before it leaves the building. Compliance check, editor, manual release — the AI is a tool, not a decision-maker.

Sign up nowView demo