What the AI does day-to-day

The AI in LEGALinhouse is not a free-standing chatbot but a layer that is active in every module — always with the context of the matter you are working on. Typical tasks:

  • Research: "Which BGH decisions on written-form requirements in lease agreements are relevant here?" — the AI returns verified case references with a brief summary.
  • Draft generation: briefs, dunning notices, statements, terminations — as a DIN 5008-compliant letter, always as a draft for review.
  • Contract analysis: AI-assisted metadata extraction from PDF contracts, with short summary.
  • Facts structuring: turning unstructured facts into an ordered form, with timeline and parties.
  • Deadline recognition: incoming letters are scanned for deadlines, suggestion lands in the inbox.
  • Classification: incoming documents and correspondence assigned to the right matter and process type.

21 specialized German legal assistants + international profiles

A generic AI gives generic answers. LEGALinhouse routes every query to a matching legal assistant — a system prompt with its own vocabulary, standard literature and retrieval configuration. For German law that's 21 assistants: among them labor law, tenancy law, corporate law, contract law, IT law, data protection law, tax law, criminal law, business and tax criminal law, administrative law.

On top of that, further assistant profiles cover international jurisdictions, for over 140 profiles in total. Whoever asks an international question (e.g. a contract under US or UK law) gets a assistant that at least knows the fundamentals of the relevant jurisdiction — for substantive questions this is a guide, not a substitute for local counsel.

What this looks like in practice

You open the chat in a matter about lease termination. LEGALinhouse detects: German tenancy law. The tenancy assistant is loaded, the relevant BGB provisions are pulled into context, and the answer cites only decisions that exist in the case law database. Switch into the next matter (e.g. data protection) and the data protection assistant takes over automatically.

Streaming responses in real time

Anyone who has waited on an AI response knows: three seconds for a two-sentence question is bearable, but 30 seconds for a longer research query kills the workflow. LEGALinhouse therefore delivers streaming responses in real time — you see the answer being written, you can interrupt early if you notice the AI is going wrong, and you can read along while it types.

Sources that actually exist

Hallucinated case references are the most common failure mode of legal AI. LEGALinhouse prevents this at two levels: a generation-time constraint (the AI can only cite from a retrieved list of real court decisions, with structured court / date / reference number from the graph), and post-hoc validation (every citation in the output is checked against the case law database and flagged if needed).

Details are under Concept · AI Architecture. Practical effect: if a case reference appears in a response, it exists — or it is explicitly marked as unverified.

Case directives: what your organisation has checked binds the AI too

A wrong date from an earlier AI text, a legal position from an older letter your organisation no longer holds: whatever had once made it into a case kept coming back with every new AI answer. The AI reads the case — and nothing in it said that someone had already checked. Every case now has a „Vorgaben" (directives) tab for that:

  • Checked fact: what has been checked in your organisation takes precedence over documents, letters and earlier AI answers. Where a source differs, the AI follows the directive and points out the discrepancy. Mandatory law takes precedence over every directive.
  • Do not use — always with what applies instead: a bare prohibition leaves a gap the AI would fill by itself. With a search term — say the wrong date or docket number — every occurrence in the case's AI answers is marked [UNBEDINGT PRÜFEN], and a marked draft can only be released after review.
  • Mark as superseded: directly on the letter or AI result. From then on the AI reads the source with [ÜBERHOLT]: it no longer adopts the legal reasoning, while addresses, dates and docket numbers from it remain usable.

The directives reach the AI case summary, briefing, AI task, AI chat, the AI letter draft and the summary of an Anliegen — pseudonymized like the rest of the case content. Anyone with write access to the case may create them; every entry records author, date and reason and is deactivated rather than deleted.

What does not work yet

Quick help, AI email draft, know-how, simulation and dictation polish do not read the directives yet. Documents cannot yet be marked as superseded. A free-text directive without a search term works through the instruction to the AI, not through an automatic mark — the tab states for each entry which of the two applies.

Strategy: options for action played through before your organisation decides

Before your organisation responds to a letter from the other side, an authority or an incident, it pays to look at the paths that are open. The AI task „Strategie (Handlungsoptionen)“ (strategy — options for action) plays them through for a question on the case:

  • Four options — do nothing, restrained, proportionate, escalating — each with the likely and the most dangerous reaction of the other side, plus immediate measures, early-warning signals and a pre-mortem.
  • Calculated, not guessed: first-instance court fees, the settlement range, which steps announced by the other side are credible, a decision matrix with a shortlist and the assumptions the recommendation depends on most are worked out by LEGALinhouse from the AI's assessments without further AI. All values are marked as assessments and count as unverified until a person confirms them.
  • Your frame: the kind of situation, the amount in dispute, your organisation's goal and the risk level by which the options are ranked.

The result is an internal draft for your legal department, not legal advice. Your organisation makes the decision. External collaborators cannot use this task type and do not see strategy entries.

Human-in-the-loop

AI output never leaves LEGALinhouse without human approval. Letters go out only after an employee has approved them. Contracts are only filed once the extracted metadata has been reviewed. Deadlines are suggested, not set automatically. This architecture is not a brake but a protection — it keeps legal responsibility with the human, where it belongs, and makes LEGALinhouse RDG-compliant: productivity tool, not legal service.

Long answers no longer break off

Long analyses used to stop mid-sentence. With "Continue generating" the generation can be resumed until the thought is finished. And a result you do not accept goes back to the AI with your review note: the original question, the previous result and your remarks produce a corrected new version — the old version stays traceable as "revised", so the correction path is documented rather than overwritten.

Going deeper

Anyone who wants to understand the architecture behind the AI — knowledge graph, routing, citation protection, 3-phase pseudonymization — finds it under Concept · AI Architecture. GDPR and the EU AI Act in detail under Security & Compliance.