In May 2025, U.S. Magistrate Judge Ona T. Wang, of the federal court for the Southern District of New York, issued an order that most governance setups had not anticipated. In the lawsuit brought by The New York Times against OpenAI — over the use of copyrighted work to train ChatGPT — she ordered the company to preserve user conversation logs, including the ones users themselves had deleted. The ruling reached over 400 million users worldwide. According to the judge, that data could serve as evidence, and it could not be allowed to disappear.
OpenAI had promised users that a deleted conversation would be permanently removed within 30 days. The court order suspended that promise indefinitely. It was not an isolated case: in prior years, litigation over data preservation had already touched files, emails and cloud documents. What was new was that it reached AI conversations, at a moment when most companies still treated an AI chat as just a tool, not a business record.
Judge Wang's preservation order surfaced a dilemma most companies had never faced: deciding whether to keep AI conversations is usually the vendor's call, not the company's. When the company does not decide, it carries both risks at once — retaining conversation data forever without legal authorization (a regulatory risk), or deleting data right when a court order demands preservation (a litigation risk).
The chat is already a business record
Judge Wang's order is not only OpenAI's problem. It is a warning for every company that uses AI. According to the law firm Nelson Mullins' analysis of the case, the ruling shows that vendor privacy commitments can conflict with discovery orders, and that can expose enterprise customers to data-disclosure demands they do not control. If the company sells to someone, signs a contract with an authority, responds to an investigation — and someone later questions what was done — the AI conversation can become evidence.
That changes the status of the conversation. It is no longer just a personal record of the employee who chatted with the AI. It is a record the company may be responsible for keeping or discarding, depending on the law and what a court asks for. But nobody was warned before AI use scaled up. Most departments adopted the tool thinking about productivity gains. Nobody in the company thought: "what if someone sues us and asks for these conversations?"
In practice, what happened was: conversations sat on the vendor's server. If the vendor deleted them, they were gone. If the vendor retained them, nobody at the company knew for how long, or why. By the time the legal risk arrived, the company realized it had no control.
Why generic policies fall short
Many companies wrote a data retention policy last year. The policy says something like "retain for 30 days" or "retain for 1 year." The problem is that the policy gets written once and then sits on a document server. When the company is sued over a specific matter, a judge can ask: "preserve everything about the tax transaction between October and November," or "preserve every sales-team conversation with at-risk clients." A company cannot do that if retention is just one flat default — "everything for 30 days."
According to Spinach AI's corporate retention guide, a real policy needs three layers: define what each data type is, define how long each type is kept in each context, and leave room to preserve data outside the normal schedule when a court order requires it. To do that, though, the company's AI tool needs to allow configuration by data type, department and risk level — not just a single "keep everything" or "delete everything" switch.
Most companies today do not have that. They either keep conversations indefinitely (exposure risk, regulatory risk when the law says to delete) or delete everything within days (risk of being unable to comply when a court order arrives). Neither option is comfortable.
What has to be in place
A company that wants to be ready for the reality after the May 2025 ruling needs five verifiable mechanisms, not a promise.
Retention configurable by context and data type. The company defines it: a personal conversation disappears in 30 days. A client conversation lasts 6 months. A conversation about a regulatory decision lasts for the length of the process, plus a year. When a department does something sensitive, the trail stays longer. All of it automatic: the company does not need a waiting list of "delete this one."
Automatic deletion. When the term expires, the conversation and its trail are destroyed without manual work. If there is no preservation order, the date arrives and the conversation is gone. That matters legally: the company can prove it complied with retention rules, not that it forgot to delete something.
Each company's data kept separate. If the company uses a shared platform with other organizations, company A's conversations are not visible to anyone at company B, or to job candidates, or to consultants. Logical isolation enforced inside the database itself, through policies the platform runs — not something a human operator has to remember.
Access matching each person's role. A junior analyst sees the conversations tied to the cases they work on. A director sees an aggregated summary of patterns and risks, not the conversations themselves. An auditor sees everything when they open an investigation, and that access gets logged. There is no "someone always has access to everything." Every access is a recorded, auditable action.
Audit trail. Who requested that a conversation be preserved, on what date, and why? Who accessed a specific client's conversation? How long was it kept before being deleted? During an audit or investigation, the company can answer: "here is every action, on what date, by whom."
That is how Skyller was designed: retention configurable by context, and automatic deletion once the term expires.
From improvisation to a plan
Every large company has one or two people who found a clever way to use AI, and now that trick never leaves their hands. The analyst who found the perfect way to respond to a tax notice. The person who built a script for reviewing contract documents. Nobody else knows how to do it.
When a company puts AI on a governed platform, those conversations and scripts stop being one person's property. They can be made available to others on the team — within their role and permissions — and reused. Someone creates; the whole company moves forward. And the company does not start from zero: it has access to ready-made process and policy templates.
There is also an effect on budget and visibility. When everyone buys their own AI tool, some people overspend and others barely use it. When retention is scattered across several vendors, nobody knows how much data is stored where. A governed AI platform centralizes all of it: shared credits, visible consumption, visible risk. If a preservation order arrives, the company knows where its data is and can comply.
Five questions for your next meeting
Before writing another AI policy, it is worth answering these questions with IT, legal and operations leadership:
-
If the company gets sued over a department's practices between January and March, can we preserve — and produce — every AI conversation from that period? If the answer depends on searching several vendors, or remembering who used which tool, the problem is not policy. It is architecture.
-
How much AI conversation data is being retained on platforms the company does not even know exist? If you cannot answer, start there: an inventory of the tools in use.
-
When the law changes — GDPR changes, the EU AI Act changes, a new regulator asks for something — can we adjust retention in a day, or does it take weeks of vendor tickets? If it takes weeks, the company sits in constant non-compliance risk for those days.
-
Is there a documented trail of who accessed which conversation, on what date, for what reason? If not, the company cannot demonstrate security in an audit.
-
Can a director or auditor get the conversation they need quickly, or does it involve a special request, a database dump, a call with the vendor? If it is slow, when an urgent court order arrives, the company will be scrambling to respond, instead of already being ready.
