Your AI Chats Might Be Evidence: Privilege and Privacy Risks of Using AI in Investigations

At IDX we conduct a variety of investigations: business email compromise (BEC), ransomware, insider threats, web server compromises and more. Many of these are conducted under a tri-party agreement between IDX, the client and counsel. This helps ensure the investigation is conducted at the direction of counsel and protected by attorney-client privilege.

With the rapid growth of AI, many companies are now using AI in all kinds of ways, including internal investigations. One key thing to be aware of: if AI is used during an investigation, the chats and the AI’s responses may be discoverable in court.

This could include HR using AI to review personnel files, or an internal IT team member pasting logs or emails into an AI for analysis. If this happened before counsel was engaged, or outside of counsel’s direction, it may all be discoverable later.

Court Rulings

In February 2026, the Southern District of New York decided United States v. Heppner. A senior executive who knew he was under investigation for securities fraud used a consumer AI tool (Claude) on his own to analyze the investigation which included his potential exposure and possible defense arguments. He later shared those outputs with his lawyers. When the FBI seized the documents, he argued they were privileged.

The court disagreed, for a few reasons:

• An AI is not an attorney. No attorney-client relationship existed with the tool.

• No expectation of confidentiality. The platform’s privacy terms allowed it to collect, retain and potentially disclose user inputs and outputs, including to government authorities. Sharing information with a third party under those terms isn’t confidential, and confidentiality is essential to privilege.

• Sharing it with your lawyer later doesn’t fix it. Privilege can’t be created after the fact.

• Not work product. The materials weren’t prepared by or at the direction of counsel.

The court also found that putting counsel’s advice into the AI tool could also waive privilege over that advice. This means that copying your lawyer’s email into a chatbot to summarize it may cost you the protection on the original advice.

It has been noted that the outcome could be different if counsel directs the use of AI, or if the company uses an enterprise platform with negotiated confidentiality terms.

Data privacy concerns

Beyond privilege, using AI during an investigation raises data privacy issues. During an investigation, sensitive data could be supplied to the AI. Some examples of this are employee names, HR records, emails, log files, screenshots, client names and case details. Much of that contains personally identifiable information (PII).

Depending on the platform and its settings, that data could be:

• Used to train the model. This creates the risk that it could resurface to outside parties.

• Retained in the provider’s logs. Even if training is turned off, many providers keep inputs for abuse monitoring, and that retention is outside your control.

• Subject to legal holds. In 2025, a court order in the New York Times v. OpenAI litigation required OpenAI to preserve user chat logs, including ones users had deleted. If a provider is ordered to preserve data, this could include your previously “delete” chats.

• Exposed through sharing features. Shared chat links have shown up in public search results before.

How to use AI in an investigation

AI can be a powerful tool in investigations. The key is how and when it’s used. Below are some practical steps that can be taken if you plan to use AI in an investigation:

• Engage counsel before using AI on anything related to a legal matter. Ensure that AI use is done at the direction of counsel.

• Address AI in the engagement. If AI will be used, cover it in the tri-party agreement or the engagement letter.

• Use approved enterprise tools. Make sure your enterprise tools has training disabled, defined retention and contractual confidentiality terms. Avoid having employees use consumer accounts.

• Minimize what goes in. Redact or sanitize names and PII where you can.

• Document how AI was used. Document which tool, what model, what data, and for what purpose.

• Train HR, IT and security teams. These are the teams that are often the first ones to handle an incident, before counsel is involved.

• Use a local model. Local models keep data off third-party servers, since the model is hosted on your own systems. They should still be used at the direction of counsel, and any prompts or outputs saved locally are still part of the case data

The safest approach is to treat cloud AI platforms as a third party, and to involve counsel before using them in connection with a legal matter.

Shadow AI Assessment

Do you know which AI tools your employees are already using, and what they’ve pasted into them? Most organizations don’t find out until an investigation or a discovery request forces the question. IDX’s Shadow AI Assessment uses forensic artifact collection to identify AI tools in use across your environment and classify them as approved or not approved, giving your CISO and General Counsel a defensible picture of AI exposure before it becomes evidence. Contact us at response@idx.us to learn more.

Mari DeGrazia is a Director of Incident Response at IDX. She is an experienced digital forensics and incident response (DFIR) professional with over 15 years in the industry who has led multiple large-scale forensics and incident response investigations including ransomware, insider threats and advanced persistent threats. She is also a Principal SANS instructor and author of FOR563: Applied AI for Digital Forensics and Incident Response.