You have 120 days, and the evidence is in three systems.

One

The clock started in March, and it is operational, not legal.

On 10 March 2026 the Department of Justice issued its first department-wide Corporate Enforcement and Voluntary Self-Disclosure Policy, applying across all 93 US Attorneys’ Offices. Most commentary read it as a leniency framework. For anyone who actually runs investigations, the operative detail is narrower and harder: the safe-harbour window for self-reporting runs 120 days from the point an internal report lands.

That converts a legal question into an operational one. Inside those 120 days an organisation has to establish what happened, how far it reached, who knew, and what the evidence supports — well enough to decide whether to disclose, and well enough that the decision survives being second-guessed later. The deadline does not care that HR holds the complaint, Security holds the logs, and Legal holds the privilege calls.

Most organisations cannot answer those questions in 120 days, and the reason is structural rather than a matter of effort. The same allegation gets worked three times, in three systems, by people who cannot see each other’s findings.

The deadline is not the hard part. Reconstructing one account of events from three disconnected investigations is.

Two

Federal enforcement receded. State attorneys general did not.

It is worth being precise about the direction of travel, because the common assumption is wrong in both halves. Federal activity softened through 2026: the Federal Trade Commission walked back its Rytr action in January, signalling a retreat from capability-based AI enforcement, and SEC corporate enforcement fell to a multi-decade low as the agency refocused on fraud.

State attorneys general moved the other way, and as one continuous campaign rather than isolated events. Forty-two states subpoenaed a single AI company in June over consumer protection. Fifteen states sent an evidence-preservation letter in August. Sixteen opened a formal investigation in September. Twenty-six wrote to Congress in September asking it to legislate while explicitly preserving their own enforcement authority. California served its own investigative subpoena on 1 October.

The honest reading: this is enforcement against one company, not a new general duty. But the demands in it — preserve this, produce that, show us the internal communications and the policies that governed them — are exactly the demands an organisation cannot meet when the record is scattered. An evidence-preservation letter is a test of whether you know where your evidence is.

Three

Courts have started drawing the line where your AI runs.

Three 2026 rulings matter more to this question than any regulation, because they are already binding on the parties in front of them and they are converging on the same distinction.

In Schulte v. LinkedIn (N.D. Cal., 1 July 2026) the court allowed generative AI to make final responsiveness determinations in discovery, treating it like traditional technology-assisted review under existing proportionality law, and declined to require extra validation disclosure without a specific showing of deficiency. The direction is permissive: AI review is not exotic any more.

In Jeffries v. Harcros Chemicals (D. Kan., 25 March 2026) the court amended a protective order to bar uploading discovery material — confidential or not — to open, public, self-training AI tools. It is not a blanket AI ban. The reasoning turns on retention and training-on-input, so closed tools are not caught by it. Hoffman v. University of Nebraska Medical Center (D. Neb., 2 October 2026) repeated the pattern.

Taken together: courts are writing the open-versus-closed distinction directly into protective orders. The question being asked of a platform is no longer whether it uses AI, but whether the model retains what it sees and who else can reach it. That is a question about architecture, and it is answered before the dispute starts, not during it.

The question is no longer whether you use AI on evidence. It is whether the model keeps what it saw.

Four

What a defensible record has to show.

Strip away the framework language and a record has to answer four questions, in writing, long after the people involved have moved on. Who did what. When. On what authority. And what happened to the evidence between the moment it was collected and the moment it was produced.

The fourth is where it usually breaks, and not at the dramatic moment. It breaks at the handoff — between the person who found the thing and the person who investigated it. A file copied to a share, an export emailed to outside counsel, a screenshot pasted into a ticket. Each one is a point where the chain stops being demonstrable and starts being a matter of recollection.

A useful exercise, and the one we would run with you first: take one closed case and trace it end to end. Every system it touched, every handoff, every person who had to be told something twice. Most people find the gap in under an hour, and it is almost never where they expected.

Five

Enforcement as a function, not a fire drill.

The departments that define rules are not equipped to enforce them, and that is not a criticism of those departments. Security sets controls and is measured on prevention. HR sets policy and is measured on employee relations. Compliance monitors frameworks and is measured on coverage. None of them is staffed or trained to build a case that holds.

What is missing is not another tool for one of them. It is an owner: one function, with one record, that every department can see the relevant part of and act within. Access decided by role, so the same case file can be worked by Security, Legal and HR at once without any of them seeing what they should not.

That is the operating model OrcheSight is built around. Whether you buy anything from us is a separate question from whether the model is right, and we would rather you got the model right.