A lawyer’s work largely involves reading, searching, and writing—tasks AI excels at, provided the data stays with you. Below, where it actually helps and where we consciously step back.
Where AI helps#
- Documents — document intelligence: upload a scan/PDF → summary, extracted fields, Q&A with citations. No more manual page-by-page contract reviews.
- Case file research — Concierge RAG on your database: ask about meaning, get an answer with source references, not the entire archive.
- Client service — initial qualification and answers to common questions, with escalation to a lawyer when needed.
- Synthesis — research agent: gathers public sources, verifies, and compiles a note with citations—under supervision.
Where AI steps back#
| Task | AI’s Role |
|---|---|
| Case file summarization | helps (with citations) |
| Database search | helps (semantically) |
| Initial client service | helps (with escalation) |
| Legal advice | no—human only |
| Procedural decision | no—human only |
| Representation | no—human only |
The rule is strict: AI prepares materials and reduces workload, but responsibility and decisions remain with the lawyer. Every answer is citable for verification, and if the match is weak, the system says “I don’t know” and escalates instead of fabricating.
Confidentiality is the foundation, not an add-on#
Case files are sensitive, so data residency is a deliberate choice. The data path is simple and self-contained: the document stays with you → embeddings are computed locally with the BGE-M3 model → semantic search runs on your server → only a masked prompt—without names, case numbers, or personal data—goes to the cloud (if at all). Masking PII is a step before every call to an external model, not an option to switch on. This approach is GDPR-compliant from the first line, not bolted on later.
Metrics: what to realistically expect#
Concrete figures depend on document quality, so we give them as honest ranges, not promises. What we see most often:
- Field extraction (parties, amounts, dates, deadlines) from contracts and filings: typically 90–98% accuracy on clean, printed PDFs; dropping to ~80–90% with scans, photos, and handwritten notes—where lower-confidence OCR comes into play. Critical fields (amounts, statutory deadlines, service dates) are always verified by a human.
- Summarization with source citation cuts the initial review of a long document by an order of magnitude—from reading page by page to a few minutes to get your bearings and jump to the fragments that genuinely need attention. The lawyer approves the result rather than trusting it blindly.
- Confidence as a threshold: fields and answers below a set confidence threshold go to human review instead of moving on as "done." This deliberately shifts the system toward caution.
These ranges are a starting point for a pilot, not a guarantee—we measure the real values on your documents in the first weeks.
GDPR and the AI Act in a law firm#
Case file data is personal data, often of a special category (health, criminal matters, financial situation). Two consequences:
- Processing agreement. If inference goes through an external API, a data processing agreement (Art. 28 GDPR) with the provider is required. Self-hosting and local embeddings eliminate this requirement at the source—the data never leaves your infrastructure, so there is no one to entrust it to.
- AI Act: limited, not high risk. An assistant in a law firm only prepares materials—advice, procedural decisions, and representation stay with the lawyer. Such a system is, as a rule, a limited-risk system, not a high-risk one under the AI Act. It is, however, subject to the transparency requirement (Art. 50): a client contacting the assistant must know they are talking to AI. We cover deadlines and obligations in more detail in the article on AI Act and GDPR obligations in 2026.
It is always worth confirming the classification of a specific deployment with a lawyer—context (e.g. systems supporting decisions by public authorities) can change it.
Try it live#
The core of document handling is summarization while preserving meaning. Paste a fragment and see the result from our sandbox (playground: PII masked, zero retention):
FAQ#
Can AI provide legal advice?#
No. AI prepares materials—summarizes documents, searches case files, answers general questions with citations—but legal advice, procedural decisions, and representation remain the lawyer’s responsibility. The system is designed to escalate, not pretend to have competencies it lacks.
Are client files secure?#
Yes—it’s a deployment requirement. Sensitive documents are processed locally, embeddings are computed on your hardware, and PII is masked before anything is sent to the cloud. Sensitive data never leaves your infrastructure.
Where to start in a law firm?#
With one narrow, highly repetitive process—e.g., summarizing and extracting fields from a single contract type. It’s easy to measure ROI and safely delegate a small, well-defined scope to AI.
How much does it cost and how do you calculate the return?#
The cost depends on scale (number of documents, whether hardware for local embeddings is needed), and we quote it as a range after a short assessment, not upfront. The starting point is usually a small pilot on a single process, where the return shows up in the time saved per document. The ROI calculator helps: you enter the number of documents and the manual handling time, and you get an indicative range of savings—without inflating it.
How is this different from an assistant for an in-house legal team?#
A law firm serves many clients and external matters, so the emphasis falls on case-file confidentiality, extraction from procedural documents, and initial client service. An assistant for an in-house legal team works inside a single company—closer to contracts, policies, and questions from the business. The technical core (RAG on your own database, PII masking, human in the loop) is shared; the context and data scope differ.