# Legal tech: AI in legal practice

By Ivo Donker · 1 November 2026 — written with AI assistance (Claude Opus 5.5); facts checked against the sources listed below

Legal tech is software for legal work, from case-law research to contract review and file analysis. Since 2023 it has mostly been about generative AI. This page sets out what works in Dutch legal practice, which rules apply (the Dutch Bar, the judiciary, the AI Act, the GDPR) and how to choose a tool without putting professional secrecy at risk.

## What counts as legal tech?

Legal tech is any technology that makes legal work faster, cheaper or easier to check. The categories overlap in practice, because most vendors now run a language model underneath their product. A useful split:

| Category | What it does | Where AI makes the difference |
| --- | --- | --- |

| Legal research | Searching legislation, case law and literature | Plain-language questions, summaries with citations |

| Contract review | Finding clauses, comparing them to a standard, flagging deviations | A fast first read of large numbers of contracts |

| Document drafting | Producing drafts and standard documents | A first version from your own templates and the facts of the file |

| File and e-discovery analysis | Organising and searching large document sets | Surfacing timelines, entities and contradictions |

| Compliance and governance | Keeping rules, registers and risk assessments up to date | Linking internal processes to legal requirements |

| Practice management | Time recording, intake, planning, billing | Automatic intake and time capture from email and documents |

Concrete tools per category are listed in the directory: [AI tools for legal work](https://directory.llmnet.nl/en/ai-tools-juridisch) and [AI tools for legal services](https://directory.llmnet.nl/en/ai-tools-voor-juridische-dienstverlening). Compliance platforms have their own list: [AI governance and compliance platforms](https://directory.llmnet.nl/en/ai-governance-en-compliance-platforms).

## What already works, and where it goes wrong

Language models are good at reading, organising and writing a first draft. They are weak at exactly what legal practice depends on: a source that really exists, quoted correctly. A model can invent a ruling, an ECLI number or a statutory provision that looks convincing. So every legal use follows the same rule: **check every reference against the original source** before a document leaves the office.

Vendors now build tools for this, such as automatic citation checks. One example is the gallery Anthropic added to Claude for Legal, covered in [Claude for Legal gets a gallery of ready-made artifacts](https://nieuws.llmnet.nl/en/claude-for-legal-artifacts). A check like that helps, but it does not replace the lawyer's final responsibility. Which models handle Dutch legal texts best is covered in [models for legal texts](https://hub.llmnet.nl/en/modellen-voor-juridische-teksten).

## Rules for lawyers: the Dutch Bar recommendations

The Netherlands Bar (NOvA) has published [recommendations on AI in legal practice](https://www.advocatenorde.nl/voor-advocaten/digitalisering-ai/aanbevelingen-ai-de-advocatuur) (in Dutch). They give direction and sit alongside the existing professional rules, such as the duty of confidentiality. The core:

- **Confidentiality.** No confidential or client data in public AI services. Share only what is strictly necessary and check where a vendor stores and processes data.

- **Verification.** Check citations, case law and facts yourself, and choose tools that show their sources.

- **Informing clients.** Tell clients what the firm's AI policy is and ask for consent before AI is used on their file.

- **Knowledge.** According to the NOvA, basic knowledge of (generative) AI is essential for every lawyer.

## The Dutch judiciary: no robot judge

The Dutch judiciary has its own [AI strategy](https://www.rechtspraak.nl/organisatie-en-contact/innovatie-binnen-de-rechtspraak/ai-strategie) (in Dutch). Judicial decision-making explicitly stays human work; a "robot judge" is ruled out. The judiciary focuses on low-risk uses such as rostering, pseudonymising judgments, plain-language summaries and help with structuring large case files. High-risk AI stays out of the judicial process until the judiciary can meet the stricter requirements. A public register of where AI is used is planned.

That this is not theoretical became clear in 2024. In a dispute between neighbours over lost solar-panel yield, a sub-district judge at the Gelderland District Court used ChatGPT to estimate the lifespan of solar panels and the electricity price ([ECLI:NL:RBGEL:2024:3636](https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:RBGEL:2024:3636), 7 June 2024). The ruling started a debate on whether such an estimate counts as a fact or as expert evidence. More background in [AI in the Dutch judiciary: use and limits](https://nieuws.llmnet.nl/en/ai-in-nederlandse-rechtspraak).

## The AI Act for legal applications

The EU AI Act names the administration of justice as a high-risk area (Annex III, point 8). That covers AI systems intended to help a judicial authority, or someone acting on its behalf, research and interpret facts and law and apply the law to a concrete case, and similar use in alternative dispute resolution. A research tool a law firm uses for its own work is not automatically covered; the intended user and purpose decide the classification.

The Digital Omnibus, published in the Official Journal of the EU on 24 July 2026, moved several dates (source: [Future of Privacy Forum](https://fpf.org/blog/the-ai-act-implementation-timeline-what-changes-under-the-ai-omnibus/)):

| Provision | Applies from |
| --- | --- | --- |

| Prohibited practices (Article 5) and AI literacy (Article 4) | 2 February 2025 |

| Rules for general-purpose AI models (GPAI) and penalties | 2 August 2025 |

| Transparency obligations (Article 50) | 2 August 2026; Article 50(2) (synthetic content) from 2 December 2026 |

| High-risk systems in Annex III, including the administration of justice | 2 December 2027 (was 2 August 2026) |

| High-risk systems in Annex I products | 2 August 2028 (was 2 August 2027) |

The omnibus also rewords the AI literacy duty in Article 4. To understand the regulation from the ground up, read [the EU AI Act explained](https://nieuws.llmnet.nl/en/eu-ai-act-uitleg) first, then [what companies must actually do in 2026](https://nieuws.llmnet.nl/en/ai-act-in-de-praktijk-wat-bedrijven-in-2026-concreet-moeten-doen). The [AI Act category checker](https://consultancy.llmnet.nl/en/tool-eu-ai-act-categorie-checker) shows quickly which risk class an application falls into.

## GDPR and professional secrecy

Legal files almost always contain personal data, often special-category or criminal data too. Any AI service that handles them needs a data processing agreement, a clear answer on where the data is stored, and a firm commitment that the vendor does not use it to train models. The [GDPR privacy checklist](https://gids.llmnet.nl/en/avg-privacy-checklist) and [AI models and privacy](https://hub.llmnet.nl/en/ai-modellen-en-privacy-avg-compliance) go through these points. For liability when AI systems fail, the revised Product Liability Directive matters: see [liability for failing AI](https://nieuws.llmnet.nl/en/aansprakelijkheid-bij-falende-ai-de-impact-van-de-herziene-pld).

## Choosing and introducing a tool

These questions separate the usable tools from the risky ones in practice:

- **Where is the data?** Preferably in the EU, with a data processing agreement and no training on your data.

- **Does the tool show sources?** An answer without a verifiable reference is only a starting point for legal work.

- **Is there a log?** Who asked what, and which answer came back; see [audit logging and compliance](https://api.llmnet.nl/en/audit-logging-en-compliance).

- **Can you leave?** Exportable data and no lock-in to a single vendor.

- **Who checks?** Record that a lawyer reviews every AI result before it goes to a client or a court.

For highly confidential files, an open model on your own hardware can be an alternative: the data then never leaves the office. How that works is covered on [gids.llmnet.nl](https://gids.llmnet.nl/en/); mind the licence terms, see [commercial usage rights of open models](https://hub.llmnet.nl/en/commerciele-gebruiksrechten-open-modellen). For copyright questions around AI output: [AI, intellectual property and copyright](https://consultancy.llmnet.nl/en/ai-intellectueel-eigendom-auteursrecht).

## Working in legal tech

With the AI Act, demand is growing for people who connect law and technology. The profile of such a role is in [the AI compliance officer](https://vacatures.llmnet.nl/en/ai-compliance-officer-rol).

## Sources

- NOvA, [Aanbevelingen AI in de advocatuur](https://www.advocatenorde.nl/voor-advocaten/digitalisering-ai/aanbevelingen-ai-de-advocatuur) (Dutch)

- De Rechtspraak, [AI-strategie](https://www.rechtspraak.nl/organisatie-en-contact/innovatie-binnen-de-rechtspraak/ai-strategie) (Dutch)

- Future of Privacy Forum, [The AI Act implementation timeline under the AI Omnibus](https://fpf.org/blog/the-ai-act-implementation-timeline-what-changes-under-the-ai-omnibus/) (28 July 2026)

- Gelderland District Court, [ECLI:NL:RBGEL:2024:3636](https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:RBGEL:2024:3636)

This page is general information, not legal advice.