ChamberOne

Retention and deletion of AI-generated content

By the ChamberOne team 5 minute read

Your file retention policy probably says what happens to letters, emails, notes and drafts. It probably does not say what happens to the question a paralegal typed into an AI tool last Tuesday, or to the summary it produced. Those items now exist in several places: the tool, the matter, a chat history, perhaps a spreadsheet. If nobody has decided what they are, they will be kept by accident or deleted by accident, and either can cause trouble. This article gives compliance leads a way to decide.

Key points

  • Map prompts, outputs, accepted text and logs onto categories your file policy already has.
  • Decide and record whether each kind belongs on the file, checking your regulator's guidance and your insurer's expectations.
  • Find every place AI content sits, including provider-side storage, exports and backups.
  • Ask vendors whether you can delete and export by matter, and whether content is used for training.
  • Give the rule an owner and a review date, and make sure legal holds cover AI content.

Start by asking what the item is

AI-related material falls into a few different kinds, and they do not all deserve the same treatment. A prompt is something a person wrote, often containing client facts. An output is a draft produced by a tool. An accepted output is text that someone reviewed and adopted as part of the firm's work. A log entry records that a call was made, by whom and when.

Think of the closest existing category. A prompt that contains client facts and instructions resembles an internal note or an instruction to a junior. An output that was never used resembles a discarded draft. Accepted text resembles any other work product. A log resembles a system or access record. Mapping new items onto old categories is usually more productive than writing a separate policy for 'AI'.

Are prompts and outputs client records?

There is no single answer that fits every jurisdiction and every kind of work, and this article cannot give one. Rules on what belongs on a client file, who owns it and how long it must be kept differ between countries and between regulators. Check the guidance from your regulator or bar, and ask your insurer what they expect to see if a claim arises.

What you can do is make a reasoned internal decision and write it down. Many firms land on a simple split: accepted outputs and the material needed to show how they were checked form part of the file, while discarded drafts are treated like other working drafts. Prompts that contain client confidential information are confidential whether or not they are kept, so where they live matters as much as how long.

Mapping where the material actually sits

Before you set a rule, find out where the content lives. In most firms it is more scattered than expected.

  • Inside the practice management system, attached to a matter.
  • In the history of a standalone chat tool, tied to a personal login.
  • On the AI provider's side, subject to that provider's own retention terms.
  • In exported documents saved on laptops, shared drives or email.
  • In backups of any of the above, which may outlive the deletion you thought you had done.

Questions to put to a vendor

A rule that says 'delete after six years' means little if you cannot delete in the tool. Put these questions to every vendor and record the answers on file.

  • Does the provider keep prompts and outputs after the request is complete, and for how long?
  • Is any of our content used to train or improve models? Where is that stated in the contract?
  • Can we delete a single matter's AI content on request, and does deletion reach backups?
  • Can we export prompts and outputs in a readable format before deleting?
  • Is there a log of who made each request, and can we keep it longer than the content itself?
  • What happens to our data when we end the contract?

Drafting the rule

A usable rule is short. Suppose a five-lawyer litigation firm decides as follows: accepted outputs are saved to the matter and follow the matter's retention period; discarded outputs and chat histories in approved tools are cleared on a regular schedule; prompts must not include more client information than the task needs; and personal AI accounts are not to be used for client work. That is four sentences, and each one can be checked.

Add a named owner and a review date. Retention rules for new technology tend to be wrong in the first draft, and the first review usually finds a location nobody listed.

Do
  • Save accepted outputs to the matter and follow its retention period
  • Clear discarded outputs and chat histories on a regular schedule
  • Include only the client information the task needs
  • Name an owner and a review date
Don't
  • Use personal AI accounts for client work
  • Put more client information in a prompt than the task needs
A short retention rule for AI content

Deletion is harder than it sounds

Three problems recur. First, legal holds: if a dispute or investigation is reasonably anticipated, ordinary deletion should stop for the relevant material, and that includes AI content. Make sure whoever manages holds knows that prompts and outputs exist. Second, client requests: a client may ask for their file or for deletion of their data, and your answer should cover AI content in the same way as everything else. Third, audit records: you may want to keep a record that a request was made and by whom, even after the content has gone, because that record helps show how work was supervised.

Keep a short log of deletion runs, saying what was cleared, when and by whom. If you ever need to explain why something no longer exists, the log is your answer.

A worked example of the decision

Suppose a firm's litigation team uses an AI tool to summarise witness statements. Each summary is read by the associate and, if it is used in a chronology, saved to the matter. The associate's chat history in the tool is cleared at the end of the month. The provider says it keeps requests for a short period to detect abuse, and does not use customer content for training.

The compliance lead records those facts, notes the provider's retention term as something to review at each contract renewal, and confirms that the saved chronology is the document that matters. If the case is later challenged, the firm can show the accepted text, who reviewed it and when. It does not need every discarded summary, and it has a reason for not having them.

Where AI does not solve the problem

No tool can tell you what your regulator expects of a client file, and no vendor can decide your retention periods for you. AI features also make it easy to create large volumes of low-value text quickly, which increases the amount you must classify. A cautious approach is to save very little by default, and to save deliberately when something is adopted.

Brief your people. The most common failure is not a bad policy but a paralegal who does not know the policy applies to chat tools. A ten-minute walk-through at induction, with a one-page summary, does more than a long document nobody opens.

This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.

Read next

Retention of AI-generated content in law firms | ChamberOne