Three dates moved in June. Here is what each one actually binds.

If your compliance calendar for AI in healthcare was written before June 2026, at least three of its dates are now wrong. None of the changes made the substantive obligations lighter. All of them changed when, and in one case whether, those obligations arrive. This piece lays out the record: what was signed, what was stayed, what was postponed, and what stayed exactly where it was.

Colorado HB 26-1139: a human-review floor, effective January 1, 2027

Governor Polis signed HB 26-1139 on June 2. The law reaches AI systems used for utilization review by carriers, pharmacy benefit managers, private utilization-review organizations, and managed-care entities, and it does two structural things.

First, it requires that determinations be based on an individual's clinical history and circumstances rather than group data alone. Second, it requires that any medical-necessity denial receive human review by a licensed clinician or physician. Separately, it bars carriers and Medicaid/CHP+ payers from paying for psychotherapy delivered directly by an AI system. The requirements apply on and after January 1, 2027.

The design is worth noticing. The statute does not attempt to regulate model architecture, training data, or accuracy thresholds. It regulates the decision: an AI system may participate in utilization review, but an adverse determination must pass through a named, licensed human before it binds a patient. For payer-facing and market-access teams, the operational question between now and January is which utilization-review vendors in the Colorado book of business currently issue denials with no clinician in the loop, because that workflow becomes non-compliant on a fixed date.

One correction to propagate: some trackers reported a May signing. The official legislative record shows the governor signed June 2, with the signed-act document dated June 3. Use June 2.

Colorado SB 24-205: the June 30 date that never arrived

The other Colorado story is the one that did not happen. The Colorado AI Act, SB 24-205, was the first comprehensive state AI law in the country and was scheduled to take effect June 30, 2026. It would have reached high-risk AI systems in healthcare. The date passed without effect, and the sequence that got there matters.

In April, xAI sued to enjoin the law before its effective date. On April 24 the Department of Justice intervened, its first intervention in a challenge to a state AI law. On April 27 the US District Court for the District of Colorado stayed enforcement by minute order. In May, Colorado's legislature passed SB 26-189, repealing and reenacting the AI Act as a narrower automated-decision-making framework, with substantive duties beginning January 1, 2027.

So the answer to the question every tracker was asking in the spring, what happens on June 30, is: nothing. The date was superseded before it arrived. Anyone whose vendor questionnaires, impact assessments, or contract clauses referenced SB 24-205 compliance should re-anchor to SB 26-189, a narrower statute, and to January 1, 2027, the same date HB 26-1139 lands. Colorado's two health-relevant AI regimes now take effect together.

The EU Digital Omnibus: later dates, same obligations

The European Commission's AI policy page was updated on June 16 to communicate the timeline agreed in the May 7 political agreement on the Digital Omnibus. Precision matters here: the June event was the Commission communicating agreed dates. The binding instrument, Regulation (EU) 2026/1744, came afterward, signed July 8, published in the Official Journal July 24, in force July 27.

The dates that moved: product-embedded high-risk AI under Annex I, the category that includes medical devices and in vitro diagnostics, shifted from August 2, 2027 to August 2, 2028. Standalone high-risk systems under Annex III shifted from August 2, 2026 to December 2, 2027. The co-legislators also replaced the Commission's original conditional-trigger mechanism, under which the dates would have depended on readiness milestones, with fixed calendar dates. A date you can put on a calendar is a real improvement over a date contingent on standards bodies.

Two things did not move, and they are the ones to brief upward accurately. The substantive high-risk obligations themselves are unchanged; the delay buys preparation time, not a lighter regime. And Article 50 transparency duties remained enforceable from August 2, 2026, which means the first binding AI Act obligations touching many health-adjacent systems arrive this summer regardless of the headline delay.

FDA: the null results

The FDA record for June is a set of negatives, each worth stating plainly rather than padding over.

The Office of Prescription Drug Promotion posted multiple untitled letters in June, part of ten untitled letters and no warning letters across Q2. A June 22 letter to Pinnacle Biologics challenged an unsupported "proven" efficacy characterization and a quantitative benefit claim for Photofrin. A June 24 letter to Lundbeck challenged a "100% migraine freedom" claim for Vyepti. None of the June letters cite AI-generated or AI-assisted promotional content. OPDP's theory of enforcement remains what it has always been: net impression and claim support. AI-created promotion is not yet an explicit citation basis, which is not the same as saying it is safe ground.

FDA's January 2025 draft guidance on AI to support regulatory decision-making for drugs and biologics remained the operative document through June. The comment period closed in April 2025, the joint FDA-EMA good-AI-practice principles were released in January 2026, and secondary trackers expected the final guidance in Q2 2026. Q2 ended June 30 without it. Sponsors continued operating against a draft.

What this month establishes

Read together, June's regulatory record has one consistent shape: legislators and regulators are converging on the decision point, not the technology. Colorado's new law asks who reviewed the denial. Colorado's rewritten law narrowed to automated decision-making. The EU kept its transparency duties on schedule while giving device makers more time on the heavy obligations. Nobody moved to regulate what an AI assistant tells a patient about a prescription drug, and no enforcement letter has yet cited AI-generated promotion.

That gap is the standing fact of this space. The rules arriving in 2027 and 2028 govern AI inside institutions: payers, sponsors, device makers. What AI assistants say to the public about regulated products remains governed by nothing except the engines' own retrieval choices, which change without notice and without review. Until a regulator reaches that surface, the only discipline available is measurement: document what the answers say, score it against the label, and keep the dated record. That is the work, and this month confirmed no one else is assigned to it.

This is one of three deep dives on the month. The full June record, including the vendor visibility index and the engine changes this piece sets aside, is in the June 2026 roundup.