When AI prompts become evidence, the discovery question nobody is asking
AI tools used in M&A diligence generate prompt logs, retrieval traces, and model output history. None of that is privileged by default. The discoverability of that record is the question the corporate bar has not yet faced.
A first year associate runs a contract through an AI document review tool and asks it to identify change of control provisions. The tool returns a list. The associate sends the list to the partner. The deal closes. Three years later, the deal goes south, the buyer sues the seller, and discovery opens.
The associate's prompt history is on the vendor's server. The model's response history is on the vendor's server. The vendor's terms of service say they retain that data for ninety days, or for the contract term, or until the customer requests deletion. The discovery request from the buyer's counsel asks for "all documents and electronic records relating to diligence of the change of control provisions, including any artificial intelligence tools used and the outputs of those tools."
The question of whether that material is discoverable, privileged, work product, or something else is being decided in real time across the country. The honest answer in 2026 is that the answer depends on which jurisdiction and which judge, and that the corporate bar is mostly not yet thinking about this.
Why this is different from ordinary tool discovery
Lawyers have always used tools whose outputs were technically discoverable. Westlaw research history, Bloomberg terminal searches, paralegal markup of contracts. Most of those have settled into established privilege patterns. The lawyer's mental impressions, recorded in research notes or marked-up documents, are protected as opinion work product. The underlying tool outputs (a Westlaw search results list) are not particularly interesting evidence and rarely get fought over.
AI tools are different in two ways. First, the prompt itself is often an explicit articulation of the lawyer's strategy or concern. "Identify the customer contracts most likely to have assignment issues if the deal closes by Q3" is a clear statement of the lawyer's theory of the case before it even becomes a case. That kind of crystallized thinking is what work product is designed to protect, and it is exactly what a prompt looks like.
Second, the AI output is treated by many lawyers as a finished work product to attach to a memo. A list of flagged contracts with model-generated annotations is qualitatively closer to an associate's diligence memo than to a Westlaw search result. Courts have not consistently grappled with how to characterize this output.
The prompt is the lawyer's theory of the case before it becomes a case. That is exactly what work product is designed to protect, and exactly what a prompt looks like.
What vendors are actually doing with the data
Most of the major legal AI vendors offer some version of "your data is not used to train our models." That is not the same as "your data is not retained at all." Several major vendors retain prompt and output logs for a defined period for service quality, debugging, and audit purposes. Some offer enterprise plans with zero retention or short retention. Reading the actual data processing addendum before procurement, not the marketing page, matters.
The retention period determines the litigation exposure. A ninety day retention with a deletion-on-request workflow means most matter-level discovery requests will come in after the data is gone. A two year retention with a comprehensive log means the request will land squarely on a live record.
What corporate counsel should do now
Three concrete actions, in order of how immediately they pay off.
Build prompt hygiene into associate training. The prompt is going to live somewhere. Treat it like a memo. Do not write strategy or candid opinions into prompts that you would not want a future plaintiff to read. The model does not need your editorial commentary to do its work. "Identify change of control provisions in this contract" is a fine prompt. "Identify any change of control provisions that the buyer's team might miss, particularly the ones that look enforceable in a sympathetic forum" is the prompt that becomes Exhibit B.
Negotiate retention in vendor contracts. Most enterprise vendors will agree to short retention or zero retention if asked. The default contract terms are vendor friendly. The negotiation is small and the outcome matters.
Treat the vendor's logs as a litigation hold target. When a matter goes into hold status, the hold notice should explicitly reach AI tool prompt and output logs, and the firm's records management policy should have a path to preserve or delete those logs in alignment with the hold and with the firm's broader retention schedule.
The corporate bar will figure this out, mostly through litigation that surfaces unfortunate prompt history. The associates and partners who think about it before the unfortunate first case is theirs will be better off.
I built a small tool that flags prompt-text risk patterns. It is at /tools/privilege-prompt-auditor/. It does not save what you paste and it is not legal advice. It is a fast way to look at a prompt you are about to run and see whether it reads like work product worth protecting or like a future deposition exhibit.
Walter Allison is a corporate attorney in Denver. He writes here about M&A, private equity, and venture capital structure.
Follow on LinkedIn ·
Firm bio ·
More posts