What three AI shifts this week mean for your Q3 plan

I cleared my Friday morning to write something else. Then I read the week back. EU governments and Parliament cut a deal on the Digital Omnibus, Moonshot AI closed two billion dollars in fresh capital, and ServiceNow announced the most ambitious enterprise agent governance product yet shipped. Three signals, 72 hours, all of them quietly rewriting most Q3 plans.

This week’s signals

Signal 1: The EU AI Act just got its first real haircut. On 7 May the Council of the EU and the European Parliament reached a provisional agreement on the Digital Omnibus on AI. NicFab, Modulos, and ResultsSense all carry the same substance. High-risk obligations under Annex III, the rules covering AI in employment, credit, education and law enforcement, move from 2 August 2026 to 2 December 2027. Annex I systems embedded in regulated products move to 2 August 2028. The watermarking and synthetic-content disclosure rules under Article 50(2) stay put, kicking in on 2 December 2026. That date is seven months from now, and it is the most implementable piece of the framework.

Signal 2: The agent-governance vendor map matured. At Knowledge 2026 on 5 May, ServiceNow extended its AI Control Tower. CX Today’s 6 May write-up details the substance: a Veza access-graph integration, 30 enterprise connectors spanning AWS, Google Cloud, Microsoft Azure, SAP, Oracle and Workday, a real-time kill switch for off-script agents, and a Microsoft Foundry hookup that lets one control plane govern Microsoft 365 agents alongside ServiceNow’s. CX Today references a global energy company that cut threat containment time by 97% and a U.S. financial institution that eliminated 96% of dormant non-human identities. Innovation Lab in May, general availability in August. IBM’s Think 2026 announcements the same week point in the same direction.

Signal 3: Open-source models just got venture-validated at hyperscale. Also on 7 May, Moonshot AI closed roughly two billion dollars in a Meituan-led round at a 20 billion dollar valuation. SiliconANGLE and TechCrunch report the Beijing-based developer of the Kimi chatbot saw annualized revenue top 200 million dollars in April, and that since January cumulative funding has reached 4.4 billion dollars, surpassing MiniMax and Zhipu AI as China’s best-funded LLM startup. Moonshot is open-source. That detail is the whole story.

The thread connecting them

The Q1 plan most leaders wrote in February made three quiet assumptions. August 2026 was the hard EU compliance wall. The model layer was probably OpenAI or Anthropic, with Google as a hedge. Agent governance was a Q4 problem to solve later. This week shifted all three.

The August wall is now 2 December 2026 for watermarking and 2 December 2027 for the high-risk stack. That is a real reprieve, but the compression is sneaky. Most companies will discover the December watermarking deadline only by missing it, because the 2027 headline drowns out the 2026 one.

The two-vendor model layer is no longer a two-vendor layer. A $200-million-ARR open-source model maker just got valued at 20 billion dollars on the bet that enterprises in Asia and elsewhere want frontier-grade reasoning at utility-grade prices, with the source code in the box. The Q3 vendor review needs a third row.

Agent governance is no longer a Q4 problem. ServiceNow with Veza, IBM Think 2026, and Microsoft Agent 365 from a few days earlier turned “discover, observe, kill, audit” into a buyable, integrable category with a public GA date.

Segment lens

For CEOs, the next board meeting needs three lines. One: there is a December 2026 EU watermarking and disclosure obligation, here is the named owner, here is the inventory of every model that produces audio, image, video or text in our products. Two: we are revisiting model-layer concentration this quarter and stress-testing what happens if the OpenAI or Anthropic relationship gets rebid, including a serious look at one open-source provider for at least one production workload. Three: agent governance has moved from the strategy section to the budget section, with a vendor decision targeted for September. None of those lines requires panic. They require being written down before a board member notices the gap.

For Series A and B founders, the implication is shorter. Enterprise buyers just received the same news. By Q3, every procurement team will be asking about watermarking posture, model portability, and whether a control tower like ServiceNow’s can see the agent. Those answers do not have to be perfect, but they have to exist. The founders who walk into Q3 with a one-page model-portability statement, a watermarking implementation plan, and a confirmed mapping into one of the agent control planes will close faster.

One thing to do

Open the Q3 plan today. Find every line that quietly assumes August 2026 EU compliance, OpenAI and Anthropic as the only providers, and agent governance as a later-year project. Mark them. Schedule one 90-minute working session in the next two weeks to rewrite those three assumptions against the new dates and the new market. The week did the hard part already.