EU hosting settles where the model runs. Real client files go out only when the contract is signed, the data flow is mapped and a synthetic-document test has passed.
Can we send our clients' contracts to an EU-hosted LLM API without breaching confidentiality? The question comes up when a legal-tech engineer builds document analysis. A lawyer, data protection officer or security lead has to approve it. This guide follows one hypothetical request through the checks: a 30-page supplier contract sent for a summary of its termination clauses.
We are Lyceum, a sovereign European AI inference provider for open-source models. Serverless means you call a model through our API and pay per token. Our serverless models run in Paris and Finland. We show what happens to a prompt on our Trust page. We use Germany as the professional-secrecy example here. Other countries have their own rules on lawyer-client secrecy, so do not read one country's rules across Europe.
Map the legal documents and data you would send
A request carries the document text, your instruction and the metadata of your account. The contract itself may hold names, addresses, salaries or health details. It also holds the client's commercial secrets, such as prices and termination terms.
At an assumed 500 tokens per page it is 15,000 input tokens. A token is the unit a model reads and bills by, usually a word or part of a word. The assumed answer is 900 tokens. The page count, tokens per page and answer length are hypothetical, not measurements.
List what each document type can contain before you pick a model. Ask three questions:
- Does it name living people? Then the GDPR applies to it.
- Does it include sensitive data, such as health details, that the GDPR protects more strongly?
- Does it hold a client's secret? Then professional secrecy applies on top.
Separate data protection from professional secrecy
Two bodies of law apply, and they ask different questions. The GDPR asks whether you may process personal data and with whom. Professional secrecy asks whether you may reveal a client's secret at all.
Under the GDPR, a controller is the party that decides why and how data is processed. It must have a lawful basis, a legal reason for processing such as a contract or consent. It may use only processors that give sufficient guarantees, meaning they can show appropriate technical and organisational protections. A processor is the party that handles data on the controller's behalf. Special categories, the sensitive data such as health data, are prohibited unless an exception applies.
For German lawyers, section 43e BRAO adds duties on top. The lawyer must select the service provider carefully. If the service is provided abroad, the protection of secrets there must be comparable to the protection in Germany. Our serverless models run in Paris and Finland, outside Germany, so for a German firm the service is provided abroad. The firm must judge whether secrets there get comparable protection. If the service directly serves one client matter (a mandate), the client must consent.
Section 203 of the German Criminal Code (StGB) makes it a crime for a lawyer to reveal a client's secret without authority. The penalty is up to one year of imprisonment or a fine. That is why the confidentiality contract in the next section matters.
Check storage, access and the full processing chain
The table below follows your example request through our serverless inference, as published on 20 October 2026.
| Stop | What we see | What we keep |
|---|---|---|
| Your app sends the request | Nothing yet, the request travels over HTTPS | Nothing |
| Our gateway checks your key | Your API key, the model and your prompt, in memory | A usage record |
| The GPU runs the model | Your prompt, in GPU memory in a European data centre | Recent prompt prefixes (the start of your prompt), until newer requests replace them |
| The answer returns | The response, as it streams back | The output token count, added to the usage record |
The usage record holds your organisation and user, the model, token counts and the time. We never log prompt text or write it to a database. Prompts and outputs are processed, not stored, and never used for training. The one short-lived exception is the memory cache covered below the table.
The GPU row matters most for a confidential contract. The start of your contract text stays in GPU memory, not on disk, until newer requests replace it. Ask every provider the same question about caches and logs.
Then map the rest of the chain. Check access by staff, retention of logs, subprocessors and deletion. A subprocessor is a company that runs part of a service for the provider. Ours are mainly in the EU and the European Economic Area (the EU plus Iceland, Liechtenstein and Norway). For any outside that area, GDPR safeguards apply, such as standard contractual clauses, the EU's model contract terms for data transfers. We tell you before we add or replace one, and you can object within 10 business days. The list is available by email, and we name no individual subprocessor here.
Match the contract to the legal workflow
The contract is where both laws meet. Our data processing agreement (DPA) is Annex 1 of our terms and applies when you accept them. Under it you are the controller and we are the processor. You can instruct us in writing, audit us with notice and choose deletion or return when you leave.
For a German lawyer, section 43e BRAO asks for more. The contract with the service provider must be in text form. It must bind the provider to confidentiality with notice of the criminal consequences of a breach. It must limit the provider's knowledge of secrets to what the contract needs. And it must state whether the provider may involve further persons.
Our DPA does not replace that assessment. Whether our terms satisfy section 43e or your professional-secrecy duties is for your firm to decide. If your legal team needs changes, send them to compliance through our Trust page.
ISO 27001 is a standard for managing information security. Our ISO 27001:2022 audit is in progress. Read that as a work in progress, not a certificate.
Test document handling before using client files
A model can fail on a legal task without leaking anything. It can omit a clause, attribute a sentence to the wrong page or invent a citation. Test for omissions, misattribution and invented citations before any real file goes out.
Use synthetic or properly anonymised documents for the first round. Then have a qualified person check whether each statement points to the right clause, what the answer omits and any invented citations. Do not expect automatic redaction, legal accuracy or protection of lawyer-client secrecy from any API, and we do not offer a legal-specific product.
A first evaluation costs little. The calculation is hypothetical and uses 40 synthetic contracts of 30 pages each. It uses the prices of a long-document model, MiniMax M3, from our models page, read on 20 October 2026. Input costs $0.40 per million tokens and output $2.00 per million.
one contract = 30 pages x 500 tokens = 15,000 input tokens, 900 output tokens
test set = 40 x 15,000 = 600,000 input tokens
40 x 900 = 36,000 output tokens
input cost = 600,000 / 1,000,000 x $0.40 = $0.24
output cost = 36,000 / 1,000,000 x $2.00 = $0.072
total = $0.312, about 31 centsCalculated example from hypothetical inputs: it uses the full input price, with no discount for cached prompt prefixes. The calculation shows that a first evaluation costs cents. It does not show whether the answers are correct, and it leaves out the real cost, qualified human review time.
When you are ready, bring your document-processing and confidentiality requirements to our compliance team through our Trust page.