Safe AI use for lawyers and notaries requires a verification layer, not cautious prompts

You must now demonstrate that your firm handles confidential data safely when AI tools are deployed. This is no longer optional: it is a professional obligation that sits alongside existing duties of confidentiality, competence and transparency towards clients.

The prompt is an analysis of 5 August 2026 of how lawyers and notaries should verify AI output and data handling before deployment, which argues that safe AI use depends on verifiable data flows and contracts rather than careful prompting alone. The Council of Bars and Law Societies of Europe published a technical guide in March 2026 that builds on earlier guidance and makes explicit how firms should assess tools before case work. In our assessment, this marks a shift from treating AI safety as a matter of tool choice to treating it as a matter of configuration, contracts and demonstrable control—and it now binds your professional obligations whether you work in Europe or elsewhere.

What does professional secrecy require when you use AI?

Your core duties—confidentiality, competence, independence and transparency—do not pause when you open an AI tool. The CCBE guidance is explicit: do not enter personal, confidential or client-related data into a generative AI interface unless appropriate technical and organisational safeguards are in place. This is not a suggestion about best practice. It is a restatement of existing professional obligations in the context of AI.

Comparable guidance from Singapore's Ministry of Law and the Law Society of Singapore reaches the same conclusion: client names, case details and other sensitive data have no place on public platforms. With enterprise tools, data retention and model training must be under contractual and technical control. The advisory on public AI tools goes further, discouraging the upload of privileged or confidential data altogether, and calling for anonymisation and redaction where any such data might otherwise be exposed.

Which failure modes must you guard against?

• Uncontrolled data exposure — uploading confidential client information to public or insufficiently secured AI platforms without contractual safeguards.
• Unverified output — using AI-generated content in legal work without checking its accuracy, completeness or relevance to the case at hand.
• Opaque data flows — deploying AI tools without visibility into where data goes, how long it is retained, or whether it is used to train the provider's models.
• Absent audit trails — using AI without logging which data were processed, which model was used, and how output was reviewed before use.
• Inadequate provider due diligence — selecting tools without assessing the provider's cybersecurity measures, privacy controls and contractual terms on data protection.

What concrete controls must you be able to demonstrate?

1. Classify your data before deployment — identify which information is confidential, privileged or client-related, and determine whether it can safely be processed by the tool you propose to use.
2. Verify the provider's safeguards — check the tool's security measures, data retention policy, model training clauses and compliance with relevant data protection standards before use.
3. Review and verify output — do not treat AI output as final; check it for accuracy, completeness and relevance to the specific matter before it enters your work product or is shared with a client.
4. Maintain an audit trail — log which data were processed, which model was used, when output was verified and by whom, so that you can account for your decisions afterwards.
5. Document your approach — record your policy on AI use, access controls, and the steps you take to verify output, so that your firm can demonstrate compliance with professional obligations.

How does verifiable execution differ from tool choice?

The guidelines place their greatest emphasis on demonstrability. They point to policy, access management, logging, output verification and audit trails as useful components of verifiable execution. This is a deliberate shift: the question is no longer which AI tool you choose, but whether you can show how you chose it, how you controlled the data that entered it, and how you verified what came out.

Multi-model verification—checking output against more than one model or approach—makes the control steps visible instead of presenting a single answer as final. This does not eliminate errors. It makes control possible and gives insight into what happened, so that your professional judgement remains yours. An audit trail that records which data were processed and how output was verified aligns directly with the emphasis in the guidelines on verifiability.

What role does technology play in meeting these obligations?

Technology can carry the mechanics of control: anonymisation before submission, secure processing environments, logging of data flows, multi-model verification, and audit trails that show what was done and when. What technology cannot carry is your professional judgement. The guidelines are clear that AI can support legal work, but only within an explicitly designed and verifiable confidentiality and verification approach. The final decision—whether to use the output, whether it is accurate, whether it serves the client's interests—remains with you. Tools that make these control steps visible and auditable help you meet your obligations. They do not replace your responsibility to exercise them.