Skip to main content
eData Edge
November 16, 2026

Your eDiscovery Playbook Should Account for Employees’ AI Agents

eData Edge: Navigating the Everchanging World of eDiscovery

Introduction

On October 21, 2026, Anthropic introduced Claude Duet, joining the other major AI platforms in offering an always-on personal agent. Claude Duet, OpenAI’s Dots, Meta’s Muse, and Google’s Gemini Spark have all entered the scene within months of one another, demonstrating the market’s appetite for AI that helps individuals manage their lives.

Unlike a chatbot, which waits for a question, an always-on personal agent keeps working when the user is not. It runs in the provider’s cloud and connects to the user’s email, calendar, messaging, payment, and other accounts. It checks those accounts throughout the day, remembers what it learns, and acts on the user’s behalf, usually asking approval before it sends a message or makes a payment. A morning brief might flag a contract that came back with redlines, a past-due vendor invoice, and a school pickup that conflicts with a late client call, with draft replies already prepared.

So far, most of the public debate has been about agents acting as customers, meaning whether retailers and banks will let a customer’s agent shop or transact for the customer. Walmart opened its platform to AI shopping agents the same day Duet launched, while Bank of America and Amazon have publicly resisted.

Less attention has gone to what employees’ AI agents mean for the security of company documents and data. The value of these tools is in the fact that they work across an individual’s whole life. That benefit, though, is also the risk to employers, with real consequences for how organizations govern, preserve, and collect data. Four features drive that risk: the agent draws on many accounts at once, works in the background, remembers what it learns, and acts rather than only answers.

A Separate Work Device May No Longer Be Enough

The line between personal and work technology is not new. For years, the standard answer to intermingling has been separation: a work phone, a work laptop, or a work login. Personal AI agents may make that separation insufficient, for two reasons.

First, intermingling is the point. Duet, for example, sends users daily briefs and keeps tabs on their connected apps. For executives, lawyers, and others whose work and personal lives constantly overlap, and who are often a company’s key custodians, an assistant that knows the full calendar and helps juggle every obligation is the appeal. For many, the value comes from seeing both sides of a person’s life at once and helping to manage them together.

Second, these agents reach a person’s information through accounts, not through a device. Duet inherits the connected apps from a user’s Claude account, which can include thousands of connectors to email, messaging, financial, and other platforms. The agents also may be able to work through their own web browser, and Muse can use login credentials a person shares with it, with password-manager integration on the way for both Muse and Duet. And because they largely run in the cloud, it does not matter which phone or laptop the user is holding. If the same agent is granted access to both a personal account and a work account, a separate device does not keep that data separate.

Why an Agent Poses More Risks Than a Personal Device

An employee who answers a work email from a personal phone still creates a work record in a work system. The work email remains separate from the other actions and applications on the phone, and a collection can target the work email account specifically. An agent adds a layer. In the same moment, it may send a work email and a personal one. Each lands in its own account’s sent folder, but the agent’s record of doing both lives in a third place, alongside everything else it did that day. That record is one more source a requesting party can seek, and because it does not separate work from personal, discovery aimed at work activity can reach personal activity too.

That record is the agent’s working memory, and it is built to persist:

  • Memory. Duet carries over the user’s memory, skills, and connected apps from their Claude account, and Meta says Muse remembers what matters to a person and can act on details that person mentioned only once. If an individual connects their work calendar to their personal AI agent, the memory of meetings and the substance of those invitations would be stored in the agent’s memory, in addition to the calendar itself. Depending on settings, that may also include any automatic notetaking activities as well. One can also imagine that prioritizing one personal task over a work task may create imperfect memories that inform the agent’s future actions and may decrease the employee’s adherence to employer policies when the agent deems them inconsistent with the individual’s personal priorities or imprecise instructions.
  • Blended outputs. A morning briefing is a new document, separate from the work and personal calendars it draws on, and it may rank a client deadline next to a pediatrician appointment. If the briefing is relevant because of the client deadline, the pediatrician appointment comes with it. Producing it means redacting personal details entry by entry or fighting over its scope, and the briefing itself may show how a person weighed work against personal obligations on a given day, which can bear on what they knew and when.
  • Approval trails. Duet asks for approval before sending messages, making payments, signing documents, or deleting information. Each approval records what the user was told, what they authorized, and when.

Consider a routine example. An employee has her agent order her lunch most workdays. One day when the employee is in a client meeting, it orders lunch for the employee and a client, then drafts the reimbursement email. That ordinary task ties a personal agent account to company business and company systems. The agent account now holds a record of a client meal and a business expense, which could surface in a later dispute over expenses or client entertainment.

AI interaction logs are already treated as potential evidence. In In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143 (SHS) (OTW) (S.D.N.Y. May 13, 2025), a magistrate judge ordered OpenAI to preserve and segregate ChatGPT output log data that would otherwise be deleted. A personal agent’s history may go further than a chat log or search history. It is an aggregated record of a person’s work and personal activity, and the aggregation is what makes it richer evidence. Activity that would otherwise have to be pieced together from separate email accounts, calendars, and apps appears together, in sequence, often alongside the user’s instructions to the agent. That can show not only what a person did but also what they were paying attention to and why.

What’s at Stake

For Individuals

  • Personal details become part of the dispute. This is already a familiar problem with text messages and messaging apps, where work and personal conversations share a thread. Those messages are expensive to collect and review and regularly lead to disputes over scope. Agents compound the problem. A text thread mixes two kinds of conversation, but an agent may mix a person’s email, calendars, finances, and everything it remembers about them. Once an agent account is shown to hold relevant work material, the personal information intermingled with it is no longer easy to wall off, and fights over what must be searched and produced reach a person’s whole life, not just their work.
  • Privilege. In United States v. Heppner, No. 25-cr-00503 (JSR), 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026), Judge Rakoff held that 31 documents a defendant generated with a consumer AI tool were neither privileged nor work product. The court also found no reasonable expectation of confidentiality given the tool’s privacy policy and held that by feeding his lawyers’ advice into the tool, the defendant waived privilege over that advice too. That reasoning turned in significant part on the tool’s consumer terms, and an enterprise deployment with different confidentiality and data-use commitments may be analyzed differently. A personal agent, though, will almost always be a consumer product, which the employee signs up for on the provider’s consumer terms. A personal agent connected to an employee’s email may read and summarize messages, including messages from the employee’s own counsel, as part of its ordinary work to, for example, triage the individual’s inbox. Whether that exposure waives privilege is untested, but Heppner suggests the risk is real.

For Companies

  • Business records the company may not be able to reach. A party must produce documents in its “possession, custody, or control.” If an employee’s personal agent drafted and sent business emails or logged approvals for business purchases, the company may have the emails themselves if they went out from a work account. It will not have the agent’s instructions, drafts, memory, or approval logs, which sit in the employee’s account with the AI provider. Employee handbooks, BYOD policies, and similar agreements often give the company a right to business data on personal devices and accounts, and depending on how a court defines control, those terms may establish it. The same terms can cut the other way, giving an adversary a basis to argue the company has control and must preserve and produce. Exercising that right still depends on the employee’s cooperation, which is hardest to get from a departing or adverse employee. Without it, the company may end up litigating whether it has control at all, or seeking the data from the AI provider, which may decline to produce content in response to a civil subpoena. Each step adds cost and delay to what would otherwise be a routine collection.
  • Agents that act can destroy evidence. An agent told to clean up an inbox or reorganize files may do so on a work account in the middle of a litigation hold. Duet asks for approval before deleting anything, but a custodian approving a routine cleanup may not connect it to the hold. Users can also edit what agents remember: Those changes can erase relevant information without anyone at the company knowing the source existed. The provider’s own retention settings may also purge task history and memory on a schedule, with no action by anyone.
  • Confidential information leaves company control. Consumer AI tools may retain queries, use them to train models, and disclose data to third parties. An employee who connects work email to a personal agent moves company and client information into that environment. The same is true of material produced under a protective order. If designated documents sit in an inbox connected to a personal agent, they may have been disclosed to a third party the order does not authorize.

What to Do Now

For Organizations

  • Decide whether employees may connect work accounts to personal AI agents. This is a threshold policy choice about work email, calendars, and files.
  • Consider enforcing that decision through technical controls. In Microsoft 365, for example, administrators can restrict or turn off users’ ability to consent to third-party applications accessing organizational data.
  • Consider offering an enterprise-provisioned agent. Providers are beginning to offer agents that an employer controls, such as OpenAI’s specialist agents provisioned with their own identity and credentials and a forthcoming preview of Spark in Google Workspace for business customers. A sanctioned option reduces the pull toward personal tools.
  • Confirm existing policies are consistent with that decision. Acceptable-use, BYOD, AI-use, and records-retention policies may each need to address agents and connectors.
  • Add AI agents to litigation hold notices and custodian questionnaires.
  • Prevent agents from interfering with preserved data. Hold instructions should reach automated actions, such as directing custodians to pause agent tasks like inbox cleanup or file reorganization on accounts subject to a hold, leave agent memory and task history intact, and check the agent’s retention settings.

For Litigators and Discovery Teams

  • Ask about agents in custodian interviews. Find out whether key custodians use personal agents and what those agents are connected to.
  • Treat agent data as a potential source. Memory, task histories, and approval logs may belong in document requests and ESI protocols.
  • Expect scope and privacy disputes. Proposing targeted search protocols, rather than wholesale collection, will help address the intermingling problem.
  • Confirm current capabilities before making representations. Duet, for example, is an early research preview, and retention and export features may change.

Personal AI agents offer real value, and they are not going away. Organizations that treat them as a data source to plan for now will be better positioned than those that discover them mid-litigation.

© Arnold & Porter Kaye Scholer LLP 2026 All Rights Reserved. This Blog post is intended to be a general summary of the law and does not constitute legal advice. You should consult with counsel to determine applicable legal requirements in a specific fact situation.