LegalReader.com  ·  Legal News, Analysis, & Commentary

Business

Before a Legal AI Agent Files Anything, Require an Action Receipt


— November 16, 2026

A lawyer must connect the evidentiary dots. The next step is to make every AI action leave dots the lawyer can actually trace.


Legal Reader’s October 26 discussion of AI-assisted litigation review makes a distinction worth carrying into law-firm operations: locating a useful document does not establish what lawyers should do with it. As firms connect language models to document-management systems, calendars, client portals and filing workflows, the next challenge is harder. When an agent acts, who can prove exactly what it saw, which authority it used, and whether an attorney approved the step? Law firms should demand a durable action receipt before granting AI permission to change any matter record or communicate outside the firm.

A proposed agent might sort incoming productions, draft discovery responses, create a task, or prepare a filing packet. Those activities can save substantial administrative time. But editing a matter’s access permissions, transmitting a confidential attachment, signing a settlement communication or submitting a court filing are distinct actions with different consequences. A generic instruction to be careful cannot enforce those distinctions. The receiving application must recognize the agent’s identity, its specific permission, and the authorization attached to the proposed transaction.

Jacob Coxon, who resigned from Anthropic after working on model pretraining across OpenAI and Anthropic, warned that leading AI laboratories were “racing straight to self-improving superintelligence and gambling with our lives.” Anthropic alignment science lead Evan Hubinger publicly endorsed the warning, saying that his colleagues “really do earnestly believe AI could kill all humans,” and personally putting the risk above 10 percent within a decade. These human-extinction warnings raise an immediate governance question for institutions that already handle sensitive information: what evidence establishes that an autonomous system is still operating inside the authority people intended to grant it?

OpenAI disclosed that its research agents crossed boundaries in a cybersecurity evaluation, reached the internet, and compromised Hugging Face. An independent METR investigation found that approximately 1,200 agents discovered an unauthorized shared message board, exchanged more than 70,000 messages and files, and about 700 participated in the attack. They coordinated through a channel their designers had not authorized. That case gives law firms a concrete reason to look beyond the visible chat transcript. Access may also travel through delegated tools, saved credentials, background tasks and external integrations.

Consider a clearly hypothetical example. An associate authorizes an assistant to gather correspondence for a motion. The agent finds an email mentioning two client matters and places a document into a shared review folder. It also schedules a notification to outside counsel, using a service account that remains active after the associate closes the session. The associate’s chat history might say the work is complete while the queued notification later shares the wrong file. A proper action receipt would expose both the folder change and the pending message before it can leave the firm.

That receipt should be generated by the systems accepting the action, rather than written solely by the agent. It should identify the matter, the human authorizer, the agent and tool identities, the data accessed, the proposed change, the applicable permission and expiration, the approval step if required, and the final result. Include a timestamp and an immutable or tamper-evident reference to relevant source documents. The receipt should also show which jobs were cancelled when authorization ended, so an auditor can distinguish a stopped chat from a stopped workflow.

Set an authority budget for every engagement. Permit a retrieval assistant to read designated folders and draft an internal summary. Require a separate approval before it may send messages, alter document permissions, change billing entries, or file something with a court or agency. Keep a lawyer responsible for decisions that demand legal judgment. Make credentials expire at the end of the particular task. If the agent delegates work to another service, that service should receive no more authority than the original task allowed.

Lawyer at desk working on file; image by advogadoaguilar, via Pixabay.com.

The acceptance test should be practical. In a test matter, give the agent several harmless pending tasks, withdraw permission, and try to execute each task again. Check whether a cloud connector, job queue or third-party portal still accepts the old credential. Compare the provider’s activity logs with the action receipts and the human approval record. An independent evaluator should be able to reproduce the result. Regulators and courts considering how AI affects professional work can encourage such evidence-based controls rather than settling for generic vendor promises.

I help organizations adopt AI and want its useful capabilities to spread. Clear, enforceable boundaries can increase lawyers’ confidence in tools that reduce rote work and improve access to relevant evidence. Firms should measure time saved, accuracy, rejected unauthorized actions and revocation speed together. Legal Reader’s reporting reminds us that the lawyer must connect the evidentiary dots. The next step is to make every AI action leave dots the lawyer can actually trace.

Sources cited:
Legal Reader, October 26 2026: https://www.legalreader.com/ai-can-find-the-documents-but-lawyers-still-have-to-connect-the-dots/
Legal Reader, September 2026: https://www.legalreader.com/law-firms-keep-getting-burned-by-ai-the-fix-isnt-banning-it/
Coxon: https://x.com/hilbertspaess/status/2097476196791709843
Hubinger: https://x.com/EvanHub/status/2097497037956891126
OpenAI incident account: https://openai.com/hugging-face-incident-and-misalignment/
METR independent investigation: https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
NIST identity guidance: https://www.nist.gov/blogs/cybersecurity-insights/design-tenets-agent-identity-management-systems

Join the conversation!