Where time disappears in your business.

We deploy open-source models like Llama, Qwen, or Mistral via Ollama on your own hardware or an EU server (Hetzner, Contabo, OVH). No model vendor sees your data, and there are no per-query API costs. You retain full control over versioning and updates.

Three differences: (1) Data sovereignty: your prompts and responses stay on your infrastructure, which is relevant to GDPR Art. 32 and NIS2. (2) Cost: no pay-per-token billing, instead a predictable hosting price. (3) Availability: no rate limits, no provider outages. Trade-off: ongoing model and hardware maintenance sits with your team or with us as operations (€490/month with a private model).

We connect via standard APIs: Freshdesk, Zendesk, HubSpot, Zammad, OTRS, or a self-hosted helpdesk. Incoming tickets are classified, pre-qualified, and enriched with suggested responses before being assigned to your team. The AI does not replace a support agent — it removes the first-pass reading and triage workload.

All model outputs run through a language filter that explicitly checks for Sie-Form and DACH-typical business register. The system rejects responses containing informal "du" or overly casual phrasing and requests regeneration. In production we sustain over 99% formal-language fidelity.

The workflow runs in three steps: (1) The invoice arrives by email or upload, and an OCR step (Tesseract or PaddleOCR on your infrastructure) extracts IBAN, amount, VAT ID, and invoice number. (2) The fields are validated against your chart of accounts and open purchase orders. (3) Via the BMD interface or an n8n workflow, a draft booking is created in BMD — your team approves it before it is posted. No manual typing, full GoBD-compliant audit trail. The same flow works with DATEV, sevDesk, or lexoffice.

PDF, DOCX, XLSX, image scans with OCR (via Tesseract or PaddleOCR on your infrastructure), scanned contracts, invoices, delivery notes, and structured forms. Extraction runs through RAG pipelines with a vector database (pgvector or Qdrant), so large archives (>100k documents) remain searchable.

Yes. In a sovereign setup, documents and their embeddings stay on your servers. AI models run on-prem or in EU hosting, and the vector database lives either inside your PostgreSQL or in a dedicated Qdrant container. Wherever we process personal data, we sign a data processing agreement (Art. 28 GDPR).

Public sources such as company websites, LinkedIn (via official APIs or GDPR-compliant scrapers like Apify), Northdata, Handelsregister, Crunchbase, and industry-specific directories. Each source is documented with its lawful basis (GDPR Art. 6(1)(f)). We do not buy address lists and do not work with GDPR-grey lead providers.

Initially on 30–50 historical deals: you label closed-won and closed-lost outcomes, the system extracts patterns (company size, industry, tech stack, region). After 2–3 months of production use, continuous re-training runs on your CRM data. Scoring decisions are explainable — no black box, but a transparent feature weighting.

Metabase, Apache Superset, and Grafana as open-source options; Power BI, Tableau, and Looker as commercial options. The AI layer sits between data sources (PostgreSQL, SAP, BMD, Salesforce, etc.) and the reporting tool: it consolidates, de-duplicates, and enriches data before it reaches the dashboard.

Yes, in full DACH business register. Models are trained on German-language reports and can produce weekly or monthly summaries in your company tone, from executive briefings to team newsletters. Generated texts pass through a review step (Slack approval or email sign-off) before being sent.