Joel D. Bush, writing on 21 July 2026, sets out where discovery of generative AI material has landed, and the answer is less dramatic and more inconvenient than the headlines suggest.
Why This Reaches an Accounting Firm
There is no special rule. Prompts, outputs, conversation logs and the metadata around them are being handled under the ordinary framework for electronically stored information. In a 2025 decision involving OpenAI in the Southern District of New York, AI logs were evaluated under the standard relevance analysis. In an October 2025 ruling in the same matter, employee prompt spreadsheets that carried no privilege markings were held discoverable.
Not exempt because they are new. That is the whole holding, and it is the one that matters.
A CPA firm is not a party to an AI copyright case. It is, routinely, a party or a witness in something else.
Malpractice claims. An IRS examination. A partnership dispute where the firm prepared the returns. A client insolvency where somebody wants to know what the accountant knew. In all of those, the firm's working material is in scope, and the question of what was this conclusion based on has a new possible answer that did not exist three years ago.
If a staff member asked an AI tool to summarise a set of contracts and that summary informed a position taken on an engagement, the prompt and the response are part of how the work was done. A retention policy that covers email, workpapers and the document management system but says nothing about the AI tool is describing a firm that no longer exists.
The Part That Is Genuinely Hard
The difficulty is not the rule. It is where the material lives.
A lot of AI use in professional firms happens outside firm systems. A public chatbot in a browser. A personal account. A tool somebody expensed. Often in breach of a policy that exists, which is its own problem, but the records issue is separate from the discipline issue: material the firm may have to preserve is sitting somewhere the firm cannot see and cannot hold.
Bush's practical guidance is unglamorous and worth copying. Define covered material expressly, naming prompts, outputs, logs and metadata rather than relying on a general reference to electronic records. Map where it is stored, including third party platforms. Suspend automatic deletion as soon as a duty attaches. Distinguish tools the firm controls from tools it does not. Audit what is actually in use, and reissue holds as platforms change, because they change often.
That duty attaches earlier than most people assume: when a claim is identified, a complaint arrives, a subpoena lands, or litigation becomes reasonably foreseeable. Not when proceedings start.
What to Do This Quarter
Two small things, neither of which requires a policy rewrite.
Add prompts and outputs to the retention schedule explicitly, with a period, even if that period is short. A short stated period defensibly applied is a far better position than silence, because silence is what makes routine deletion look like something else after the fact.
Then find out where AI is actually being used. Not what the policy permits, what is happening. That list is the map, and no hold can cover what nobody has written down.
The Argument for Keeping It Inside
This is the strongest practical case for using tools that sit inside the firm's own boundary rather than whatever is open in a browser tab.
MetaWurks is built that way. Documents ingested into the platform are not used to train models and are not exposed to other users, role based access controls decide who can open which client's records, and audit logs record who opened what and when. The point for this purpose is not privacy. It is that the activity happens somewhere the firm can see, in a place a retention schedule can name and a hold can actually reach.
Material scattered across personal accounts cannot be preserved, cannot be produced, and cannot be shown to have been deleted in the ordinary course.
Courts did not decide that AI records are special. They decided they are normal. Normal records are the ones a firm is expected to be able to find.
Join the Conversation
If a hold landed on your firm tomorrow naming AI prompts and outputs, could you list every tool your people have used on client work this year?