AI for law firms in Europe: client files under European law

Why should a law firm's AI stay in Europe?

Because professional secrecy covers every client file an AI model reads, and a US company that holds that file answers to US law wherever its servers are. Professional secrecy is the rule that shapes legal AI in Europe. The CCBE (2 October 2025) tells lawyers not to put client data into generative AI tools without safeguards. German law backs the lawyer's duty of secrecy (§43a(2) BRAO) with criminal law (§203 StGB). The SRA Code of Conduct (paragraph 6.3) requires solicitors to keep their clients' affairs confidential.

US law follows US companies, not servers. The CLOUD Act obliges a US provider to disclose data in its "possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States" (18 U.S.C. § 2713).

An EU region of a US cloud changes where a client file is stored, not which law the company answers to. That is exactly the kind of exposure a confidentiality duty asks a lawyer to weigh before sharing a file. The German bar says so in its own terms: under §43e(4) BRAO, a provider abroad may only see client secrets when the protection there is comparable to Germany's. The BRAK (December 2024) notes that AI providers are often based in the US, that it is not settled whether their data protection level is enough, and advises preferring providers with servers in Germany or Europe. Deutsche Telekom's announcement for Noxtua (25 March 2026) makes the same point: in light of the CLOUD Act, processing legal data in sovereign data centers run by European cloud providers is "particularly crucial".

EUrouter keeps the whole chain European. EUrouter B.V. is an Amsterdam company under Dutch law, its DPA requires all processing of personal data to take place in the European Economic Area (section 5), and every request goes to EU-hosted model providers, 15 of its 17 providers being headquartered in the EU.

Why do legal tech teams build on EUrouter?

EUrouter is the European AI API for law firms and legal tech companies that build their own tools: a European company, processing in the EEA, no prompts stored and European providers, on every plan. Each point below is in a document your DPO can read.

A European company under European law

EUrouter B.V. is registered in Amsterdam, and its terms and DPA are governed by Dutch law. The router runs on Scaleway in Amsterdam (DPA sub-processor list, updated 9 October 2026). If an authority orders EUrouter to disclose personal data, the DPA commits EUrouter to tell the customer as soon as possible and within 5 business days, to the extent the law allows, so the customer can pursue a legal remedy (section 4).

Processing only in the EEA

EUrouter's DPA states that all processing of personal data takes place exclusively within the European Economic Area, and that EUrouter transfers none outside it unless the customer authorises it in writing (section 5). The DPA also binds EUrouter and its staff to keep client information strictly confidential (section 8).

No prompts or client documents stored

By default, EUrouter applies zero retention to the content of API requests and responses, and keeps only metadata such as timestamps, model identifiers and token counts for billing (DPA, section 10). EUrouter does not train on customer data. All 17 providers in the catalog are listed with no prompt retention and no training on the providers page (27 October 2026).

EU-only on every plan

Every EUrouter request goes only to EU-hosted provider endpoints, from the Free plan up (routing docs, pricing). There is no special endpoint to remember, and when a provider is down, EUrouter falls back to another EU-hosted provider.

European providers with one setting

On 27 October 2026, 15 of EUrouter's 17 providers are headquartered in the EU, among them Mistral AI, Scaleway, OVHcloud, IONOS and Nebul. AWS Bedrock and Microsoft Foundry serve Claude and GPT models from EU regions, and both are US companies. Set eu_owned: true on a request, or once for the whole account in a routing rule, and EUrouter uses only providers headquartered in the EEA. 101 of the 154 models in the catalog have at least one such provider.

What can law firms use AI for?

Five uses fit law firm work today, as long as a lawyer checks every output before it reaches a client or a court: contract review, research summaries, drafting, document triage and transcription. Each runs on models from European labs or on open models served by European providers. The tables below read the live EUrouter catalog, show only providers headquartered in the EU, and list prices in EUR.

Contract review and clause extraction

Contract review with AI means reading a contract, finding the clauses that matter (liability caps, termination, change of control, governing law), comparing them with the firm's playbook and returning the deviations as a table a lawyer can check. Long-context chat models from European labs, such as Mistral, Aleph Alpha and Apertus, suit this work:

Finished tool: Legora, a Stockholm company (privacy policy), can be configured for full EU data residency, with hosting primarily in Sweden (Legora).

Research summaries

AI research summaries condense judgments, statutes, a long brief or a set of documents the lawyer supplies. The known failure is invented case law, which is why the CCBE tells lawyers to verify AI output and the Dutch bar to check citations, case law and facts by hand. Give the model the sources and ask it to quote them, instead of asking it to recall case law from memory: that is retrieval-augmented generation, covered in the RAG chatbot guide.

Finished tools: CoCounsel Essentials from Thomson Reuters hosts in the US and processes its OpenAI GPT and Google Gemini requests in the US, with region-based hosting listed for the UK, Australia and Canada but not the EU (Thomson Reuters). Beck-Noxtua, built by the Berlin-founded Noxtua with the German publisher C.H.BECK on beck-online, moved into Deutsche Telekom's AI factory in Munich in March 2026 (Deutsche Telekom).

Drafting

AI drafting produces first versions of letters, memos, clauses and summaries from the firm's own templates and the facts of the matter, which a lawyer then edits and signs off. The European models in the contract review table suit drafting too. Mistral Large 4 accepts about one million tokens of context (EUrouter catalog, 27 October 2026), so a long contract and its schedules fit in one request.

Finished tools: Harvey, headquartered in the US (privacy policy), hosts on Microsoft Azure and offers processing in the EU and Switzerland (Harvey). Microsoft Copilot is an EU Data Boundary service for EU customers, with Anthropic models currently excluded from that boundary (Microsoft). Ordalie, made in Paris, hosts in France on S3NS (Thales) and processes model requests in France or the EU without retaining them (Ordalie).

Document triage and search

Document triage sorts and searches large document sets, such as a data room in due diligence or a disclosure set in litigation. An embedding model turns each document into a vector, a reranker puts the best matches first, and a chat model answers with citations to the documents it used. Each of these embedding models is served by at least one provider headquartered in the EU:

Rerank models are listed on the model catalog. The RAG chatbot guide walks through the whole pipeline with code.

Transcription of meetings and dictation

AI transcription turns client meetings, dictation and hearing recordings into text that a chat model can then summarise. A recorded client meeting is covered by professional secrecy just like a document in the file. Each of these speech-to-text models is served by at least one provider headquartered in the EU:

Can lawyers use ChatGPT or Claude on client files?

Not through a consumer chat account. The CCBE tells lawyers not to enter client data into generative AI tools unless safeguards are in place, the SRA names a staff member using online AI such as ChatGPT on a client's case as a confidentiality threat, and the NOvA says to put no confidential or client data into public AI models.

The safeguards the CCBE lists are a contract that binds the provider to confidentiality or zero data retention, a data processing agreement that limits use to the firm's purposes, or a system that runs locally or in a secured environment the firm controls, with processing in the EU/EEA where appropriate. An API under those terms is a different risk from a chat account, but OpenAI and Anthropic are US companies, and Claude and GPT models in EU regions run on AWS and Microsoft, also US companies. For client files, models from European labs on European providers take the US-law question off the table.

In France, the CNB guide (September 2024) tells lawyers never to send data covered by professional secrecy to generative AI: no client names, no confidential information, no case strategy. For French firms, that means anonymising before any model sees the text, wherever the model runs.

What do European bar rules say about AI and client data?

European bars agree on three points: keep client data out of public AI tools, check every output, and know which company processes the data and where. Professional secrecy sits on top of GDPR, and in some countries it is backed by criminal law.

Sources: CCBE guide, BRAK KI-Leitfaden, §43e BRAO, NOvA recommendations (announced 16 December 2025), SRA Code of Conduct, SRA Risk Outlook and CNB guide.

For a firm sending client files to a US company, these rules leave work to do: the firm has to show that the protection is comparable despite US law, as §43e(4) BRAO asks, and map every subcontractor, as the NOvA asks. A European provider under a DPA that keeps processing in the EEA answers both questions in writing.

EUrouter's DPA binds EUrouter to confidentiality, applies zero retention to request and response content by default, and lists every sub-processor and model provider with its location. EUrouter has not published a secrecy undertaking in the form §43e(3) BRAO describes, with its warning about the criminal consequences of a breach. Whether EUrouter's DPA meets that requirement for your use is for your firm to assess, and a DPA review is the place to raise it.

How do GDPR and the EU AI Act apply to a law firm's AI?

GDPR applies to every client document that names a person, and the EU AI Act mostly does not classify a law firm's own AI tools as high-risk. The two laws set different duties:

• The AI provider is a processor. Under GDPR Article 28, the firm needs a data processing agreement with its AI vendor, and the vendor needs the same terms with each sub-processor, such as the model host.
• Transfers need a basis. Sending personal data outside the EEA needs a basis under Chapter V of GDPR. For the US, the EU Court of Justice has struck down that basis twice, in 2015 (Schrems I, C-362/14) and 2020 (Schrems II, C-311/18), both times over US surveillance law. Processing the model call in the EEA with a European provider keeps that hop out of Chapter V.
• Client files hold sensitive data. Health data in an injury claim is a special category (Article 9) and criminal records fall under Article 10. Article 35 requires a data protection impact assessment for large-scale processing of these categories.
• High-risk means courts and arbitration. The AI Act lists as high-risk AI used by or on behalf of a judicial authority to research and interpret facts and law, or used in a similar way in alternative dispute resolution (Annex III, point 8(a)). After the Digital Omnibus on AI, Regulation (EU) 2026/1744 (in force since 27 July 2026), these obligations apply from 2 December 2027 (ERTICO, 7 August 2026; AI Act timeline).
• Client-facing chatbots must say they are AI. Since 2 August 2026, Article 50 requires AI systems that interact directly with people to be built so that people know they are dealing with an AI (Article 50).

The GDPR text is on EUR-Lex.

How do you build a legal AI tool on EUrouter?

EUrouter is OpenAI-compatible: point the OpenAI SDK at https://api.eurouter.info/api/v1, pick a model, and add a provider object that keeps the request on EU-owned providers with zero data retention. This example asks Mistral Large 3, served by Mistral AI and Nebul, to review one clause:

import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["EUROUTER_API_KEY"],
    base_url="https://api.eurouter.info/api/v1",
)

clause = (
    "The Supplier's total liability under this Agreement shall not exceed "
    "the fees paid in the three months preceding the claim."
)

response = client.chat.completions.create(
    model="mistral-large-3",
    messages=[
        {
            "role": "system",
            "content": "You review contract clauses for a European law firm. "
            "List the risks for the customer, quote the words that create each risk, "
            "and say when you are unsure.",
        },
        {"role": "user", "content": clause},
    ],
    extra_body={
        "provider": {
            "eu_owned": True,
            "zdr": True,
            "data_collection": "deny",
        }
    },
)

print(response.choices[0].message.content)

The x-provider-slug response header shows which provider served each request, so your audit trail names the company that processed the clause. To make these settings the default for every key, save them in a routing rule: a request can never loosen a rule's retention limit. To search a whole document set instead of one clause, follow the RAG chatbot guide. The quickstart covers API keys.

What to check before going live

Before an AI tool touches a client file, a law firm should be able to answer each of these with a document, not an assumption. Where EUrouter has an answer, it is in the item.

1. Classify the data. List which matters and documents the tool will see, mark special category and criminal data, and decide what never goes to a model.
2. Read your bar's rules. Use the table above as a starting point and check whether the client must consent, as for single-mandate services in Germany.
3. Map every hop. Write down the app, the router, the model provider, logging and monitoring, with the country and the owner of each, as the NOvA recommends. EUrouter's DPA lists each sub-processor and model provider with its location, and the x-provider-slug header names the provider of each request.
4. Sign the contracts. Put a data processing agreement in place with each processor. EUrouter's DPA covers Article 28 GDPR, confidentiality and EEA-only processing, with a signed version on the DPA page. Where your bar requires a separate secrecy undertaking, as §43e(3) BRAO does, check whether the DPA meets it and raise it in a DPA review.
5. Turn off retention and training. On EUrouter, set zdr: true and data_collection: "deny", and eu_owned: true to keep requests with providers headquartered in the EEA. Save them in a routing rule so no request can loosen them.
6. Keep traces in the EU. EUrouter's Broadcast sends request and response content to the observability tool you choose. Leave it off for client data, or send it only to a destination hosted in the EU.
7. Ground answers in your own sources. Retrieve from the firm's documents and make every citation checkable, instead of letting the model recall case law.
8. Require human review. A lawyer checks every citation, fact and figure before anything reaches a client or a court.
9. Run a DPIA. Record the risks, the safeguards and the decision, and review them when the tool or the provider changes. EUrouter's DPA commits EUrouter to assist with a data protection impact assessment on request (section 6).
10. Train the people who use it. The CCBE treats knowing the risks of the technology as part of professional competence.

Start building your firm's AI in Europe

Create a free account, point your client at https://api.eurouter.info/api/v1, and your next request runs in Europe with a European company. Create a free account, compare plans, talk to us about a DPA review for your firm, or follow the RAG chatbot guide to search your firm's own documents.