Not into a public consumer chatbot. Client documents belong in an AI tool your firm or company controls: business terms that exclude training on your data, hosting where your client and data protection obligations allow, access limited by matter, and logs. Whether a given use affects privilege is a question for your own counsel. This is not legal advice.
#Key takeaways
- The risk is not AI itself but where the text goes, who can read it and what the provider's terms say.
- A US federal court has held that a defendant's conversations with a public AI chatbot were not privileged.
- DIFC Courts' guidance warns about confidentiality when using public models that may store or reuse inputs.
- Lawyers will use AI anyway; an approved tool with clear rules is safer than a ban that pushes use to personal accounts.
- Write down, per tool, where data is stored and processed, who can access it, and how long it is kept.
#What is actually happening in legal teams?
An associate pastes a draft settlement letter into a chatbot to tighten the wording. A commercial manager uploads a counterparty contract to ask what it means. An in-house lawyer summarises a board paper on a personal phone. None of them think of it as disclosure. All of them are sending client or company information to a third party under terms nobody in the firm reviewed.
This is common at work generally: 78% of AI users bring their own AI tools rather than company ones (Microsoft and LinkedIn Work Trend Index, may 2024).
#Does using AI waive legal privilege?
There is no single answer, and it depends on the jurisdiction and the facts. One signal: in United States v. Heppner, a federal court in new york held that a criminal defendant's conversations with a public AI chatbot were not protected by attorney-client privilege or work product (Orrick, march 2026).
Commentators read that ruling narrowly, and it does not settle how courts elsewhere will treat a tool used under a lawyer's direction on a controlled platform. It does show that the provider's terms and the user's role matter. Ask your own counsel how privilege applies to the tools your team uses.
#What do courts in the Gulf say about AI and confidentiality?
The DIFC Courts' guidance on large language models warns that public AI models may store or reuse input data, which can breach confidentiality, and asks practitioners to keep clients informed about AI use in submissions (DIFC Courts, december 2023). Personal data inside legal documents is also covered by data protection law, such as the UAE's federal decree-law no. 45 of 2021 (UAE government portal).
#What should a legal AI use policy cover?
- Approved tools, by task: drafting, summarising, research, translation.
- What may never go into any tool: for example sealed material, privileged advice to a named client, or data a client contract excludes.
- For each approved tool: where data is stored and processed, whether it is used for training, retention, and who at the provider can access it.
- Access by matter, so people only reach documents they already could.
- Client consent and engagement letter wording where your professional rules or client terms require it.
- Logging of who used which tool on which matter, and a named owner who reviews the log.
#When is an enterprise chatbot plan enough?
For many teams, a business or enterprise plan of a mainstream assistant, configured with no training on your data, single sign-on and retention settings, covers drafting and summarising non-sensitive material. Check where it processes data before relying on it. A private assistant on your own documents fits when matters are sensitive, client terms restrict processing location, or work must stay separated by matter.
We compared these options in is Copilot or ChatGPT enterprise enough and a private ChatGPT on your own documents.
#How does Stennir help legal teams use AI safely?
Stennir's AI consulting audit maps which legal workflows touch which data, writes the tool policy and the hosting decision for each, and names who approves AI output. AI implementation builds the private assistant or workflow on your accounts, hosted where your rules require, with matter-level access and every action logged. AI training teaches lawyers and staff the policy on their own documents.
We are
Stennir decides where legal data is stored and processed before anything is built, and leaves you owning the accounts, the logs and the off switch.
We aren't
Stennir is not a law firm and gives no opinion on privilege or professional conduct; your counsel and regulator rules decide those.
#What should a general counsel or managing partner do next?
Ask your team which AI tools they used on client or company documents last month, without blame. The answer is your starting inventory. See how compliance and legal teams use AI on AI for compliance, take the free readiness assessment or book a 30-minute call.
#Common questions
Is ChatGPT enterprise safe for client documents?
It can be part of a safe setup, but safety comes from the configuration and the terms, not the brand. Confirm that your data is excluded from training, where it is processed, retention and admin access. Then decide which matters may use it. Your counsel confirms whether that meets your professional and client obligations.
Where does our data stay if Stennir builds a legal assistant?
Where your rules require. Stennir writes the hosting decision into the scope before building: which country, which cloud and which model can run there, and where logs sit. The accounts are in your name, access follows your matters, and you hold the off switch.
Does a private legal assistant work in Arabic?
Yes. Stennir builds and tests assistants in Arabic and English on your own documents, including bilingual contracts and Arabic correspondence. Legal Arabic terms are tested against your own glossary before rollout. Other languages are available on request.
What could go wrong with an approved legal AI tool?
Access can be too broad, so someone sees a matter they should not; people can paste documents into an unapproved tool anyway; and outputs can be wrong. Matter-level permissions, logging, a short policy people understand and a named person approving output cover most of it.
