Privilege and confidentiality before using prompts

The prompts in this library are thinking aids. They do not become thinking aids until someone supplies them with real context — the actual company, the actual facts, the actual question. The moment real context enters a consumer AI assistant, the board member has acted outside the confidentiality envelope that would normally protect those facts. This note sets out what that means, and how to use the prompts without creating an incident.

1. Legal privilege does not extend to consumer AI assistants

Swiss attorney professional secrecy (Berufsgeheimnis, Art. 13 BGFA / Art. 321 StGB) attaches to communications between a client and an attorney admitted to the Swiss bar in the course of a mandate. It does not attach to a board member’s conversation with ChatGPT, Claude, Gemini, or any other consumer AI assistant. Pasting privileged communications or privileged work product into such an assistant risks compromising the protection over that material — at minimum for the text pasted, potentially more broadly depending on the facts and forum.

Enterprise deployments with contractual confidentiality, zero-retention terms, and no-training commitments materially reduce the confidentiality risk; the privilege analysis is independent of the contractual terms and turns on how a Swiss court (or a foreign court applying Swiss confidentiality rules) would characterise the disclosure.

2. Confidentiality obligations run independently of privilege

A director’s duty of confidentiality under Art. 717 OR is a separate and broader obligation than legal privilege. Sharing confidential company information with a third party — including a consumer AI provider — requires a basis. Board members should assume no basis exists for sharing material, strategic, or personal information through a consumer AI assistant.

3. Data protection adds a further layer

The revised Swiss Data Protection Act (nDSG, in force since 1 September 2023) and, for EU/EEA personal data, the GDPR impose obligations on the processing of personal data. Named individuals — executives, counterparties, employees — pasted into a prompt are personal data. The processing must be lawful, proportionate, and documented. Many consumer AI deployments do not, by default, meet these requirements.

4. Safe patterns

There are several patterns that let a board member use AI assistants meaningfully without exposing privileged or confidential material:

1. Redacted prompts. Replace names, companies, sums, dates, and other identifying details with neutral placeholders (“the company,” “the counterparty,” “a sum in the low tens of millions,” “the previous quarter”). The reasoning quality usually survives; the identifiability does not.
2. Hypothetical framing. Rephrase the question as hypothetical — "Imagine a Swiss listed company in the life-sciences sector facing…" — and keep the prompt at that level.
3. Abstracted variables. For decisions that turn on numbers, feed ratios or ranges rather than absolute figures. For decisions that turn on timing, use relative windows rather than dates.
4. Counsel-mediated use. For prompts that require real detail, route the prompt through counsel. An attorney using an AI assistant within the scope of the mandate may maintain privilege over the output, subject to the usual conditions.
5. Enterprise deployments with scoped terms. When an organisation has contracted for enterprise AI use with data-processing terms, zero retention, and restricted training, more real context may be supplied — though privilege is not restored by contract alone.

5. Patterns to avoid

1. Pasting verbatim privileged communications (attorney letters, memoranda, court filings in proceedings not yet public) into a consumer AI assistant.
2. Supplying named counterparties, executives, or internal personnel where it is not necessary to the question.
3. Using AI to draft public-facing statements during active regulatory investigations without counsel review.
4. Uploading entire board packs, internal audit reports, or investigation files to obtain summaries or analyses.

6. A working rule

Write the prompt as if it will become a public document — because, in the worst case, it will. The prompts in this library are written to produce high-quality output even on redacted, hypothetical, or abstracted input. Good prompt design and good confidentiality discipline are, for this purpose, the same thing.