Skip to content

AI Assistant (add-on)

Applies to: Both (Self-Hosted & SaaS) · License: requires the AI Assistant add-on (feature flag ai) · Permission: ai.use

The AI Assistant is a read-only, privacy-first helper for operators. It summarises activity, explains failures and drift, and answers product questions — but it never makes changes and nothing is stored: prompts and responses are ephemeral.

Where to find it

  • A dedicated AI Assistant page, opened with the AI assistant button in the top bar.
  • Inline AI buttons on the Activity log, the reconciliation drift report and the employee detail page.
  • A global ⌘K command palette (fire an AI action or jump to any page, with a "Recent" group).
  • Settings → AI Assistant — shown only when the add-on is licensed.

Bring your own key (BYOK)

The assistant routes through litellm and works with the provider you choose:

  • OpenAI, Anthropic (Claude), or Google (Gemini) — an admin pastes the API key in Settings → AI Assistant (stored masked, like the SMTP password) and picks the active provider + model.
  • Ollama (self-hosted) — point at a base URL; no key required. The base-URL field is hidden in managed/SaaS mode.

Output is produced in the user's UI language (English or German).

What it can do (all read-only)

  • Activity log summaries — an on-demand plain-language summary of recent console activity.
  • "Explain this failure" — per-workflow explanation of why a step failed and what to check.
  • Reconciliation drift explainer — explain the whole drift report or a single drift category.
  • In-app help assistant — RAG over the product documentation, with source citations and multi-turn follow-ups.
  • "Explain recent activity" — a summary of what happened to a specific employee, on their profile.

Access control & privacy

  • Gated by the ai add-on licence and the ai.use permission (granted to Operator and above; viewers are denied). The Settings → AI Assistant area is hidden entirely when the add-on isn't licensed.
  • A usage meter shows per-provider request counts only — never prompt or response content.
  • Nothing the assistant sees or produces is persisted.

Full capability list also appears in FEATURES.md §19.