May a law firm, tax practice or medical office use AI with confidential client data? Yes, but only under conditions, and the cloud makes them demanding: entering a protected file into an online model discloses it to a third party, which engages both data protection law and professional secrecy. Run the model on your own hardware instead, and nothing leaves the building, so the hardest question never arises.
Stand: 2026-10-13
The bind falls on exactly the professions that would gain most from AI. Lawyers, tax advisers, accountants and doctors handle document-heavy, repetitive files that language models handle well. They also all carry a duty of secrecy, and that duty grants no discount for new technology. Overlook it and the exposure is not only a data protection fine but, in several member states, a breach of professional or even criminal law.
What professional secrecy demands of AI
Across the EU, legal professional privilege and medical confidentiality are long settled. What they share is a simple test: confidential information must not be made accessible to anyone outside the privileged relationship without a legal basis. A cloud service does not need to read a file for the duty to bite. It is enough that the file has been handed to an outside party for processing.
Data protection sits alongside this. The GDPR treats an AI vendor that processes personal data on your behalf as a processor under Article 28, which requires a written data processing agreement. Where the files contain health data, criminal matters or other sensitive categories, Article 9 adds a further layer, and Article 32 requires appropriate security throughout. These duties apply whether the tool is cloud or local.
The two national examples below show how much weight the secrecy duty can carry beyond the GDPR baseline.
The cloud route and its conditions
In Germany the duty is criminal. Section 203 of the Criminal Code makes it an offence for a professional to disclose a client secret without authorisation, punishable by up to one year of imprisonment or a fine. A 2017 reform, in force since 9 November 2017, added paragraphs 3 and 4: disclosure to an outside service provider, such as a cloud or AI vendor, is permitted only where using that provider is necessary, and a provider who is not itself bound by professional secrecy must be put under a duty of confidentiality. Fail to impose that duty and the professional, as well as the provider, can be criminally liable.
For lawyers, Section 43e of the Federal Lawyers Act turns this into concrete requirements that go beyond data protection. Access is limited to what is necessary. The provider must be selected carefully and dropped at once if it breaches the terms. The contract must be in text form, must flag the criminal consequences of a breach, and must bind the provider to secrecy, with the same obligation passed down to any subcontractor. A provider abroad is only permissible where comparable secrecy protection exists, which is generally assumed within the EU. And where the service serves one specific matter, the client must consent.
Spain reaches the same place by a different route. Professional secrecy for lawyers is anchored in Article 21 of the General Statute of the Spanish Bar (Royal Decree 135/2021, in force since 1 July 2021) and ultimately in Articles 18 and 24 of the Constitution. In April 2026 the Spanish Bar issued an interpretive circular confirming that using generative AI to draft legal texts is lawful and admissible, while stressing that it creates no technological exception to the duty of confidentiality. The tool is permitted; the duty stays.
The practical catch is the same everywhere. A free account on an online model brings no text-form confidentiality commitment, often no data processing agreement, and frequently trains on whatever you type. Outside the EU the logic is identical: the American Bar Association's Formal Opinion 512 of 29 July 2024 tells US lawyers to check, before any client information goes into a tool, whether it trains on inputs, where the data is stored and who can access it, and to obtain informed client consent where needed.
Local AI: why the disclosure does not happen
Here the logic inverts. When the language model runs on a machine inside your own office, no document goes to a third party. There is no processor to disclose to, so on our reading the entire service-provider chain, from Section 203 of the German Criminal Code to Article 28 of the GDPR, simply does not arise. No text-form contract, no confidentiality undertaking from an outside vendor, no cross-border assessment, because there is no vendor and no border.
This is practical today. A local stack of an open-weight model, a vector database and an embedding model can search contracts, opinions or medical records without a single token leaving the network. We have shown what that looks like for confidential files in local RAG for private documents. The legal lever is larger than the technical one: here data sovereignty is not a convenience but the clean answer to a duty of secrecy, which is why we build local AI rather than route client data through someone else's servers.
What local AI does not solve
Local is not the same as carefree, and it would be dishonest to pretend otherwise. Three things remain.
First, security under Article 32 of the GDPR. A machine in your own office still needs access control, encryption and a permissions concept. An open server on the network is not progress over the cloud, only a different risk.
Second, professional review. A model writes fluently even when it is wrong, and responsibility for every output stays with the professional, whether the draft came from the cloud or from your own server. Local AI removes the disclosure question, not the diligence.
Third, the remaining data protection duties. Where personal data is processed, information duties, retention limits and, in some cases, a data protection impact assessment still apply.
A checklist before you deploy
Before the first client file goes into a model, settle five questions. Does the tool process data locally or in the cloud? If cloud, is there a contract that commits the provider to confidentiality in the form your jurisdiction requires, not just a data processing agreement? Is access limited to what is necessary, and is training on your inputs switched off? Does the work serve one specific matter, so that client consent is needed? And, independently, is technical security under Article 32 of the GDPR actually in place?
Answering all five cleanly with a cloud tool is a lot of contractual work. Work locally and the first four are settled by the architecture, leaving only the fifth to take seriously. For most firms and practices that is the shorter path to an AI deployment that holds up in front of a regulator and a client alike.
If you want to test whether local AI carries your confidential files, we set up a pilot project on your real documents, with nothing leaving your premises.
Frequently asked questions
Can a law firm use AI with confidential client data?
Yes, but with conditions. Putting client information into a cloud AI tool discloses it to a third party, which touches both data protection law and professional secrecy. The GDPR requires a data processing agreement under Article 28, and professional rules often demand more on top. When the model runs locally on the firm's own hardware, nothing is disclosed to anyone, so the hardest condition does not arise.
Is a GDPR data processing agreement enough to use AI on client files?
Not on its own. Article 28 of the GDPR is mandatory for any processor, but professional secrecy is a separate and often stricter duty. In Germany it is backed by criminal law under Section 203 of the Criminal Code and by Section 43e of the Federal Lawyers Act. A processing agreement satisfies data protection, not the professional duty.
Does running AI locally solve the confidentiality problem completely?
It solves the core of it, because confidential files never leave the firm and are not disclosed to any outside provider. It does not remove the rest: security of processing under Article 32 of the GDPR, access control, and the duty to review every AI output professionally. Local means safe from disclosure, not free from diligence.
Do free consumer AI accounts meet professional confidentiality duties?
Usually not. A free account typically comes without a data processing agreement, often trains on the inputs, and gives no contractual confidentiality commitment. For regulated professionals handling client or patient data, that falls short of both the GDPR and professional secrecy rules.