The deadline you built your roadmap around just moved - again
If your AI governance program has a target date attached to it, there is a good chance that date no longer exists. Compliance and privacy teams spent the first half of 2026 building programs against two anchor dates: the EU AI Act's high-risk obligations, originally due in August 2026, and the Colorado AI Act, originally effective mid-2026. Both are gone. Budget requests, vendor reviews, and board updates got built around them anyway, because a statutory date is the easiest thing to put in a slide.
The EU AI Act's Digital Omnibus (Regulation (EU) 2026/1744), published in the Official Journal on July 24, 2026 and in force three days later, pushed standalone Annex III high-risk-system obligations to December 2, 2027, and pushed product-embedded (Annex I) AI systems further still, to August 2, 2028. Colorado moved on a similar track: Governor Polis signed SB 26-189 on May 14, 2026, replacing the prior AI Act with a narrower automated-decision-making-technology framework and pushing deployer and attorney-general rulemaking obligations to January 1, 2027 - the second delay for that law.
What actually took effect while the headline deadlines slipped
The delays are real, but they are not a blanket postponement. Two things happened on schedule, and they matter more to a compliance program than the dates that moved. General-purpose AI provider obligations under the EU AI Act - the rules aimed at the model builders, not just the deployers - have applied since August 2, 2025, and were untouched by the Omnibus. Article 50 transparency duties and the AI Office's enforcement powers over those general-purpose AI providers took effect as scheduled on August 2, 2026.
The part of the law that reaches the companies building and shipping models is live. The part that reaches most enterprise deployers is the part that moved. That distinction gets lost in "AI Act delayed" headlines, and it is the reason a governance program built only around the deployer deadline is now missing the part of the calendar that did not change.
The enforcement that never had a deadline: state attorneys general
While the statutory clocks reset, a different kind of enforcement kept running on no clock at all. A 42-state coalition led by New York Attorney General Letitia James served OpenAI with a sweeping investigative subpoena on June 12, 2026, covering advertising practices, user data handling, and internal company policy on model behavior. On August 3, 2026, a 15-state coalition led by Iowa Attorney General Brenna Bird sent a pre-litigation evidence-preservation letter tied to a data breach, demanding the company preserve internal records. A separate, formal investigation covering 16 states and led by Montana's attorney general opened in September 2026 over the same incident.
None of these actions cite an AI-specific statute, because in most of the states involved, none exists yet. They are built on existing consumer-protection and data-practices authority - the kind every enterprise already operates under, AI or not. That is the detail worth sitting with: the subpoena and the preservation letter both function as document-production demands first and AI-policy demands second. They ask a company to locate, preserve, and hand over communications, internal policy documents, and decision records - across departments that do not normally share a case file. Connecticut Attorney General William Tong put the broader dynamic plainly on September 16, 2026: federal action is lacking, so "states have to lead." For a compliance program, the practical takeaway is that the absence of an AI-specific law is not the same as the absence of exposure, and the exposure shows up first as a production request, not a fine.
Prosecutors are already asking the governance question
DOJ's own evaluation criteria tell a similar story. The Evaluation of Corporate Compliance Programs - the document prosecutors use to decide whether a company's compliance program earns credit during an investigation - was updated September 23, 2024, and it already asks two direct questions: what is the company's approach to governance regarding the use of new technologies such as AI in its commercial business and in its compliance program, and how is accountability over AI use monitored and enforced. Those questions do not wait for an AI-specific statute to attach to them; they are evaluation criteria today.
It is worth separating this from DOJ's newer Corporate Enforcement and Voluntary Self-Disclosure Policy, announced in March 2026 - that policy is silent on AI. The governance expectation lives in the older compliance-program evaluation document, not the new self-disclosure incentive structure. Conflating the two overstates what changed in March and understates what was already sitting in the evaluation criteria before it. Either way, the underlying prosecutorial question does not ask whether a company had an AI policy on file. It asks whether the company can show who decided to use AI where, who was accountable for the outcome, and whether that accountability was actually monitored - the kind of record a policy document alone does not produce.
Courts are drawing their own line - open AI versus closed AI
Federal courts are not waiting either, and they are drawing a specific line. In Schulte v. LinkedIn (N.D. Cal., order dated June 30, 2026), the court accepted the use of generative AI to make final document-responsiveness determinations in discovery, treating the workflow the way it has long treated technology-assisted review - an evidentiary tool subject to ordinary proportionality and reasonableness standards, not a new special-scrutiny regime. In Jeffries v. Harcros Chemicals (D. Kan.), the court went the other direction on infrastructure: it amended a protective order to bar uploading any discovery material, confidential or not, to open, public AI tools, while permitting closed, secure AI tools for confidential material only where the vendor is contractually bound to no training on inputs, no onward disclosure, and delete-on-request.
These are individual orders in individual cases, not a codified standard - no rule requires this distinction yet. A proposed Federal Rule of Evidence on machine-generated evidence, Rule 707, was due for a vote in May 2026 and did not get one: the Advisory Committee on Evidence Rules declined to act, sent it back for further study alongside a related deepfake-authenticity question, and the Department of Justice is on record opposing the rule outright. The open-versus-closed distinction is happening case by case, at the discretion of individual judges, well ahead of any rule that would generalize it.
The gap between confidence and readiness
Enterprise compliance leaders mostly believe they are covered. In Schellman's 2026 State of AI Governance report, surveying more than 500 US enterprise leaders, 74 percent said their organization could pass an AI compliance audit today. Only 27 percent said their programs were actually fully mature. Just 44 percent had documented incident response procedures specific to AI issues, and only 36 percent said their boards regularly discuss third-party AI risk.
That gap, between the audit answer and the maturity answer, is the same gap the state AG letters, the DOJ evaluation criteria, and the protective-order disputes are all probing from different directions - not whether a policy document exists, but whether there is a record behind it. An incident-response plan with no documented incidents. A governance policy with no board discussion behind it. A discovery workflow with no way to show which tool touched which document, under what constraint. Seventy-four percent confidence and 27 percent maturity are not contradictory numbers - they describe two different artifacts. One is a policy binder. The other is a record that survives someone asking to see it.
What to check before the next subpoena, not the next deadline
None of the four fronts above - state attorneys general, DOJ's evaluation criteria, federal court protective orders, or your own board - are waiting on a statutory effective date to ask what your organization's AI governance record actually looks like. The live question is less "are we ready for the EU AI Act" and more "if a subpoena, a protective-order dispute, or a DOJ evaluation landed this quarter, could we produce the record." That is a records question before it is a policy question - and it is the one a compliance calendar, on its own, does not answer.
That is the infrastructure question underneath the Schulte and Jeffries split: closed, controlled AI that keeps evidence inside the environment it was collected in is being treated differently than AI running in the open. OrcheSight's AI runs on the customer's own infrastructure, including air-gapped environments - evidence never leaves the environment it was collected in, and every triage and redaction decision it makes lands in the case record. That is an infrastructure answer to an evidence question, not a compliance-deadline answer to a compliance-deadline question.
The deadlines will keep moving - that is what deadlines set by ongoing political negotiation do. The question worth asking this quarter is not when the next one lands. It is whether your organization could produce its AI governance record today, to whoever asks first.