The AI assistant that knows your entire company

The AI assistant works with every module the logged-in user has access to: knowledge, CRM, finance, HR, ISMS, projects, marketing, assets and more. Ask 'Who manages the customer Müller GmbH and which contracts are currently running?' and the assistant combines CRM data, contract documents and accounting in a single answer. Permission awareness is built in: whatever you cannot access, the assistant cannot see either, neither as a filter result nor as implicit background information. Everything is multi-tenant isolated on Postgres with Row-Level Security, and the tenant context travels in the prompt header, so there is no accidental cross-tenant leak. Context sources are listed transparently in the answer footer, which means you always see where each piece of information comes from and can jump to the source with a single click. For sensitive modules, extended audit logs are activated on top.

Upload PDFs, Word documents, Excel tables or Markdown files and the assistant extracts the content and answers your questions with concrete citation references. Ask, for example, 'What notice period does this supply contract have?' and you get the answer along with the page number and a verbatim quote. This works for multi-page contracts, technical specifications, audit reports or competitors' annual reports. Uploaded documents are stored multi-tenant isolated, and processing runs EU-only via Mistral and AWS Bedrock in Frankfurt eu-central-1. OCR for scanned PDFs is built in, as is table recognition with column mapping for Excel and PDF tables. The storage period is configurable per tenant: the default is 90 days with automatic deletion after expiry, and manual deletion is possible at any time. There is also a comparison function that runs the same query across multiple documents.

The assistant doesn't just answer, it acts. In the finance module, for example, you say 'Suggest me a journal entry for this document' and the assistant fills account, cost center and amount directly into the form. In CRM, 'Create a follow-up appointment in two weeks' creates the appointment and links the customer record with all relevant notes. In the HR module, 'Prepare the contract renewal' opens the matching template and fills in the personnel master data. In the ISMS module, 'Link this measure with the matching Annex A controls' brings up suggested mappings. All you do is click 'accept' or correct the suggestion. Actions are reversible, and every change is documented in the audit trail with the note 'suggested by AI assistant, confirmed by user' including time stamp and user ID.

Recurring tasks the assistant takes over completely on request: the monthly travel-cost document import from your Outlook mailbox, a weekly status report to the team with drill-down links, or a reminder for contract renewals 30 days before expiry with a prepared renewal template. You define the routine once in natural language ('Send me every Monday at 8 a.m. an overview of overdue measures in the ISMS'), the assistant suggests the configuration, and you confirm. Routines run via the workflow engine, and every run is documented and traceable in the audit trail including trigger and result. If something goes wrong, for example a missing data source or a changed permission, you get a notification with the concrete error cause and a solution suggestion. Routines are versioned, so you can roll back to a previous configuration whenever you need to.

When you correct an assistant suggestion, the system learns from it. Say you always book documents from a specific supplier to a special cost center: after three to five corrections the assistant takes that over automatically. Learning is tenant-isolated, so corrections from your tenant never influence other tenants, technically via Postgres Row-Level Security and logically via few-shot example selection. There is no training on your data in the model itself, everything stays in your tenant. Learning happens via few-shot examples that are loaded per query from your tenant, without your data influencing the model weights. Power users can view the learning entries and maintain them manually, for example to remove unwanted patterns or to mark especially valuable corrections as 'sticky'. When an employee changes, personal learning entries are anonymised or transferred to the successor.

The standard model is Mistral Large 2, EU-hosted at Mistral itself in France. For more complex tasks the assistant uses Anthropic Claude via AWS Bedrock in Frankfurt eu-central-1, also EU-only. No OpenAI, no data outflow to the USA, no training on your data, all contractually assured via DPA addenda with Mistral and AWS. The model is selected per query automatically by cost-benefit profile, so you don't have to take care of anything, but you can still choose manually for each query. For especially sensitive tenants, for example KRITIS or defence-related companies, a 'Mistral-only' mode is also available, in which exclusively European models without US involvement are used, configurable per tenant. Model performance and cost per query are shown transparently in the dashboard, with no hidden token consumption.