AI features overview

Where AI is used in the platform, how models are configured, and what the boundaries are.

AI in Swiss HR Easy is assistance, not automation. It drafts, summarises, finds and explains. It does not decide, and it does not filter people out.

Where AI is used

Feature	What it does
Application fit assessment	A short narrative summary and a 0–10 fit score against the position’s requirements.
Semantic search	Finds people and records by meaning rather than keyword.
Sentiment analysis	Multi-dimensional signals over feedback and survey text.
Document intelligence	Reasoning-based retrieval over long contracts, reglements and GAV/CCT texts.
Certificate generation	Drafts Arbeitszeugnis prose from the recorded assessment and Grundton.
Survey report	Observations over a stated data basis, above the k-anonymity threshold.
Survey assistant	Drafts survey questions from a stated intent.
Translation	Drafts the other locales of authored content for review.
Receipt OCR	Extracts merchant, date, amount and VAT from an expense receipt.
Import extraction	Reads master data out of uploaded documents during migration.
Occupation suggestion	Proposes the AVAM occupation code for a position.
AI coach	A conversational assistant for HR users working inside the product.
AI chat	Question answering over the data the asker is already permitted to see.

The AI coach

The coach is an in-product assistant for HR users: it helps compose a message, explains what a screen is for, suggests how to fill a form, and improves a prompt before it is run. It has a learning cycle that can be triggered on a schedule.

Its answers are grounded in what the asking user is permitted to see. It is not a back door around the permission model.

Model configuration

Every AI feature is a named entry in one central configuration. Call sites never hardcode a model ID — that is enforced in CI.

Models are resolved in this order:

1. Per-feature environment override

   AI_MODEL_<FEATURE_KEY> — for example AI_MODEL_SEMANTIC_SEARCH_REASONING. Swaps or rolls back one feature’s model without a deploy.

2. Legacy variable

   LLM_MODEL and CERTIFICATE_LLM_MODEL still win over the default for the features that honoured them before, so existing deployments keep their pinned models.

3. The right-sized default

Right-sizing

Models are matched to the work rather than defaulting to the largest available:

Class	Used for
Opus-class	Where reasoning quality is the product — the coach, semantic-search reasoning, certificate text, chat, document structure.
Sonnet-class	Structured summarisation, extraction, form suggestions.
Haiku-class	Translation, sentiment scoring, occupation suggestions.

Providers and data flow

Provider	Used for	Configured with
Anthropic (Claude)	Reasoning, generation, extraction	ANTHROPIC_API_KEY
Voyage AI	Embeddings for semantic search	VOYAGE_API_KEY

Content is sent to these providers when a feature runs. Which features are enabled is a per-installation decision, and each customer is the data controller for their own instance.

Auditing

Every AI operation is audited with action: "AI_OPERATION", including token usage in the metadata. There is a record of what was sent for processing, when, and by whom — which is what a data-protection enquiry actually asks for.

Turning AI off

FEATURE_AI_FEATURES=false disables the AI surfaces; FEATURE_SEMANTIC_SEARCH=false disables semantic search specifically. An installation that does not set the provider keys has no AI features regardless of the flags.

What applicants are told

Consent, opt-out, no cut-off scores, and the transparency notice.