Closed-Loop AI for Law Firms: What Two 2026 Court Orders Actually Require

7 min read
A grey padlock tag set aside on the left, with an arrow to a pale-blue contract with four checked terms and a signature, under a magnifying glass

Firms have been asking us some version of the same question on calls: is this AI “open” or “closed”? We think the question underneath is plainer. Can the firm safely put client records into this tool?

We can’t say where any one firm picked up the phrase. It travels through articles, CLEs, colleagues and vendor pitches. Two federal court orders from March 2026 are still useful context, because each judge had to write down what acceptable AI use looks like for information produced in a lawsuit.

The short answer: “Closed-loop AI” is not a legal standard, and neither order uses the phrase. The Kansas court in Jeffries v. Harcros Chemicals limited the parties to “closed AI Tools” and defined them by safeguards. The Colorado court in Morgan v. V2X went further and required contract terms covering storage and training, third-party disclosure, deletion on request and written documentation. Both are magistrate-judge orders in single cases. They don’t bind your firm, but they show what a complete answer to “is it closed?” has to cover.

Where does “closed loop” come from?

Not from the courts. The Kansas order uses “open loop” exactly once, in quotation marks, to summarize what the defendants were asking for: they sought “to prevent parties from uploading even non-confidential documents into public or ‘open loop’ generative artificial intelligence tools (collectively, ‘open AI Tools’).” The court’s own terms are “open AI Tools” and “closed AI Tools.” The phrase “closed loop” does not appear in either order.

Colorado went a step further and turned a label down. The plaintiff, who represents himself, proposed that AI tools “operate within a secure, closed-circuit environment.” The court found that language “seems to address cybersecurity concerns about unauthorized access rather than the distinct risks associated with today’s widely available AI platforms.”

Neither judge left “closed” undefined. Kansas tied it to specific safeguards, and Colorado replaced it with contract terms that say what has to be true of the information after it goes in.

What did the Kansas order in Jeffries v. Harcros require?

A locked folder inside a small dashed boundary, with arrows pushing it out to a larger boundary that also holds six plain documents, and a gavel at the corner

Jeffries v. Harcros Chemicals is a set of consolidated putative class actions brought by people who live within three miles of a chemical facility in Kansas City, Kansas. On March 25, 2026, Magistrate Judge Angel D. Mitchell granted the defendants’ motion to amend the protective order.

The AI safeguards were not new that day. The original protective order already required a party using an AI tool with Confidential Information to:

  • give the other parties notice and a chance to object
  • use the tool in a secure environment
  • keep Confidential Information from training or improving any AI tool, except one used only for that case, and destroy any information used that way when the litigation ends
  • delete all Confidential Information at the end of the case

The court said those provisions “effectively require parties to use only closed or secure AI Tools (‘closed AI Tools’)” for Confidential Information. What the March 25 order changed was reach. The amended order extends the same AI provisions to all “Discovery Materials,” meaning everything produced in discovery, not only documents marked confidential.

Training was the heart of the defendants’ argument. An open AI tool uses what it receives to keep improving, they said, so it is “practically impossible” to claw back material later found to be privileged, or to delete it when the case ends. The plaintiffs “never directly address” that point, the court noted, before citing “the very real security risks of public AI Tools, including the inability to effectively claw back information from the AI Tool.”

Two limits matter as much as the holding. The court said the defendants’ proposal “does not foreclose a party from using any AI Tools; it only prohibits using open AI Tools while allowing the use of closed AI Tools.” And it found good cause in “the AI-specific risks identified by Defendants in this particular case”: documents from companies in a critical infrastructure industry, where disclosure may violate data privacy laws (including the GDPR, which applies to two of the defendants).

What did the Colorado order in Morgan v. V2X require?

A document whose path splits two ways: one to a gear crossed out with a slash, the other to a database with a magnifying glass beside it

Morgan v. V2X is an employment discrimination case. The defendant asked the court to add AI language to the stipulated protective order and to make the plaintiff name the AI tool he was using. On March 30, 2026, Magistrate Judge Maritza Dominguez Braswell granted the motion in part.

The court turned down both sides’ drafts. The plaintiff’s addressed AI risk “in a very limited and vague way,” and the defendant’s appeared “crafted to fit the precise bounds of Defendant’s contractual engagement with AI providers.” So the court wrote its own:

No party or authorized recipient may input, upload, or submit CONFIDENTIAL Information into any modern artificial intelligence platform, including any generative, analytical, or large language model-based tool (“AI”), unless the AI provider is contractually prohibited from: (1) storing or using inputs to train or improve its model; and (2) disclosing inputs to any third party except where such disclosure is essential to facilitating delivery of the service. Where disclosure to a third party is essential to service delivery, any such third party shall be bound by obligations no less protective than those required by this Order. In addition, the AI provider must contractually afford the party or authorized recipient the ability to remove or delete all CONFIDENTIAL information upon request. A party intending to use AI that it contends meets these requirements must retain written documentation of these contractual protections.

Read the first term slowly. It names storing alongside training, and elsewhere the court pointed to “the real risks associated with mainstream tools that persistently collect and store data.” So a vendor that says it doesn’t train on your data has answered part of the question.

Retention is a separate one.

The court was candid about what this does in practice. The provision “will (at least for now) bar the parties from using most, if not all, mainstream low-to-no-cost AI to process Confidential Information,” and enterprise-tier accounts that qualify “may be available only through organizational procurement processes.” Its summary for the plaintiff was blunter: no uploading Confidential Information “into any mainstream AI tool like standard ChatGPT, Claude, Gemini, or similar platforms.”

What decides it is the contract. Confidential Information is not to “be entrusted to platforms that lack the contractual safeguards described above, regardless of the sophistication or apparent trustworthiness of the tool.”

Do these orders apply to a disability firm?

Not directly. Each is a discovery order by a magistrate judge, resolving a dispute between the parties in one federal case. Neither binds another court or sets a nationwide rule, and neither bans AI. Neither involves Social Security disability practice, either.

When two judges had to decide what adequate protection for confidential material looks like, they wrote it down in plain terms, and Colorado’s terms happen to be specific enough to put to a vendor. That makes them more useful than a label. Your own court’s orders and your state’s rules are what govern your firm (this article is general information, not legal advice).

What should a firm ask about an AI feature?

Colorado’s provision was written for one case, but its terms work as a checklist for any AI feature that will touch client records, whether it summarizes medical evidence or drafts a hearing brief. Ask about the AI provider behind the feature, not only the software company selling it, and ask for the answers in writing.

What Colorado requiredAsk the vendorAn incomplete answer
No storing or training on inputsIs the AI provider contractually barred from storing our inputs, and from using them to train or improve its model?We don’t train on your data, with nothing on storage
Disclosure only when essentialWhich other companies receive our inputs, and is each one needed to deliver the feature?We only work with trusted partners
Same protection downstreamAre those companies bound by terms at least as protective as the provider’s?They have their own privacy policies
Deletion on requestCan we have our inputs deleted on request, and how do we ask?We keep data as long as needed to provide the service
Written documentationCan you give us those terms in writing to keep on file?A verbal yes on a sales call

These five are specific to AI features. The broader diligence list (audit reports, encryption, data sharing, who can contact your clients, cancellation and ownership) is in our nine security questions to ask any SSD software vendor.

How does Chronicle answer these questions?

We do not use firms’ case data to train AI models. We do use aggregated, de-identified case data to synthesize industry trends for broader educational efforts, like our Know Your Numbers talk.

Chronicle has its own SOC 2 Type II report, issued to Chronicle Technology, Inc. Chronicle’s SOC 2 Type II audit covers the Security criterion. Chronicle is audited every year. Data is encrypted in transit and at rest. Firms doing due diligence can ask Chronicle for a copy of the report.

None of that, on its own, answers the five questions above. Those are about the terms behind a particular AI feature, so for the feature a firm is evaluating, we’ll explain its providers, the permitted data uses, and the retention and deletion terms, so the firm can assess them against its own requirements.

Frequently Asked Questions

Closed-loop AI isn’t a legal requirement in either order. Neither uses the term, and both define acceptable AI use by specific safeguards rather than by a label.

Does “no training” mean a vendor doesn’t keep our data?

A no-training promise says nothing about retention on its own. Ask separately how long inputs are kept and how deletion works.

Did either court ban AI?

Neither court banned AI. Kansas allowed closed AI tools, and Colorado said uses that don’t involve uploading Confidential Information are not restricted by its order.

Does a SOC 2 report answer these questions?

A SOC 2 report helps, but not with this. A SOC 2 Type II report covers a company’s own controls, while the terms an AI provider behind a feature has agreed to are a separate document to ask for.

Before your next AI demo

List the AI features your firm uses or is considering that touch client records. Send the five questions ahead of the call, so the answers come back in writing, and keep them with your vendor files.

Ask them of Chronicle’s AI features too.

Built for SSD practices

Give your team the hours back

ERE monitoring, document routing, hearing prep, and medical chronologies — all in one platform built for disability firms.

Book a Demo

Join disability practices across the country

Free training invites, workflow guides, and industry reports.