The choice of language model is not a cosmetic detail: it determines whether the solution fits your type of work. For long-context files, such as a 500-page deed or a complete case file, we generally use Claude because of its 200K-token context: the model reads an entire file in one go and delivers a structured summary without us having to chunk it artificially. For broad contract review flows with fast iteration we use GPT-4o, which is somewhat faster on shorter documents. And for firms with the strictest privacy requirements, such as trust offices, boutique M&A firms with government clients, or lawyers handling state-secrets cases, we deploy Llama 3.1 fully on-premise or in a private VPC. No outbound API connections, auditable model weights.
The use cases we build most often, grouped: document review and clause extraction for due diligence, RFP analysis and contract renewal cycles. Case file summarisation for litigation teams taking over a file or preparing a pleading. Semantic search across case law: not just keyword search on rechtspraak.nl or EUR-Lex, but conceptual search, such as "rulings in which a renewal clause in a lease was interpreted in the tenant's favour under ROZ conditions". Client enquiry triage via a chatbot front end on your website that handles first-line questions and gives the lawyer a structured summary ahead of the intake conversation. Legislation monitoring that scans the Staatscourant, the EU Official Journal and case law for relevance per client or file, with no more manual alerts.
We also build compliance mapping for client portfolios (GDPR, Wwft, AI Act, CSRD), KYC/Wwft screening for client onboarding, and, for firms that themselves advise clients on AI compliance, AI bias audit tooling that the firm uses to assess clients' applications. Time-recording automation is an underrated win: an AI that categorises email traffic, calendar entries and document edits against open files significantly reduces the administrative burden per lawyer, especially in firms under billable-hours pressure.
Whatever the use case, the approach is the same: a RAG layer that unlocks your own knowledge base, an EU-residency deployment of the model, an audit log that stores every model call with its input and output for review, and a human-in-the-loop design in which a lawyer always has the final check before output goes to a client.