AI Has Stopped Being a Capability Question

If you follow AI headlines, Q1 2026 felt like watching a pressure system build.

Model releases this quarter shared an unmistakable focus: reasoning, planning, and the ability to complete multi-step work. The regulatory map shifted (again) in two directions at once.

Open-source agents went from hobbyist curiosity to enterprise conversation over the span of a long weekend. And a very public standoff between one of the leading AI labs and the Pentagon reminded everyone that the trajectory of this technology isn't purely a technical question.

For insurance, none of this was noise. Each of these stories has a direct line to how you will compete over the next two to three years.This briefing cuts through the volume and tells you what actually matters - and why.

Story 1: The Pentagon Fired Its AI Vendor

The most significant AI story of the quarter wasn't a technical breakthrough. It was a contractual collapse.

In July 2025, Anthropic - the company behind the Claude AI model - became the first frontier AI lab to have its model approved for use on the U.S. military's classified networks under a contract worth up to $200 million. The model was reportedly being used to ingest and analyze large volumes of intelligence and targeting data.

By February 2026, that relationship had collapsed in a very public way.

Anthropic had built two restrictions into its Pentagon contract: no use of Claude for mass domestic surveillance of Americans, and no use of Claude to power fully autonomous weapons systems - meaning AI can't make targeting decisions without human sign-off. The Trump administration countered with a demand for "All Lawful Use" language - meaning the government would not allow a private corporation to dictate the constitutional or tactical boundaries of military technology.

When Anthropic refused to sign by the deadline, the response was swift, historic, and public. Defense Secretary Pete Hegseth designated the company a "supply-chain risk to national security" - a label previously reserved for foreign adversaries - and all federal agencies and private defense contractors were ordered to immediately cease using Anthropic's technology.

Within hours, OpenAI announced it had struck its own deal with the Pentagon, positioning itself as a safe alternative while still agreeing to the "All Lawful Use" requirement. As of March 9, 2026, Anthropic has filed a major lawsuit against the Trump administration. The story will surely continue.

What this means for insurance

The policy and geopolitical implications of this story are real, but I want to draw your attention to what sits underneath it: frontier AI is now trusted enough - and performant enough - to process some of the most sensitive data on earth.

The U.S. military does not award classified-network access to science projects. The fact that two frontier AI models have been approved for intelligence analysis on classified systems - and that the contract collapsed over use policy, not capability - tells you something important about where the technology stands.

If AI can meet the Pentagon's bar for handling classified intelligence, the question of whether it can handle your underwriting complexity is no longer a capability question. It's an implementation question.

Story 2: The Regulatory Map Got Redrawn - Again

Regulation moved on two fronts this quarter.

In Europe, the EU AI Act is no longer theoretical. Enforcement began in February 2025, and full compliance requirements for high-risk AI systems are due by August 2, 2026. The European Commission has also published a "Digital Omnibus" proposal that would ease some AI Act obligations and delay certain high-risk compliance deadlines, though those amendments still require parliamentary approval.

In the United States, the Trump administration signed an Executive Order on December 11, 2025, "Ensuring a National Policy Framework for Artificial Intelligence." The core thrust: consolidate AI oversight at the federal level and preempt the growing patchwork of state laws.

The EO doesn't create new binding regulations on its own - it directs agencies to identify and challenge state laws deemed too burdensome - but it signals a clear preference for an innovation-first, lightly regulated federal framework. Colorado, California, and others with stricter AI laws are now in the crosshairs.

What this means for insurance

Insurance is explicitly named in multiple frameworks as a "high-risk" AI application. The EU AI Act's August 2026 deadline matters. If you write or underwrite business in Europe and use AI in the process, compliance work should already be underway.

In the U.S., the direction is less restrictive federally, but state regulators - NAIC, California DOI, New York DFS - have been moving independently. The NAIC's ongoing AI working group has been actively building guidance frameworks on insurer AI use, and I will keep you updated here.

The net effect: there is no single rulebook yet. Navigating the patchwork is itself a competitive challenge. Organizations that have already invested in auditable, explainable AI systems will find compliance far easier to manage than those scrambling to retrofit accountability after the fact.

Story 3: The Way AI Is Built Got More Interesting

The biggest grassroots AI story of Q1 2026 didn't come from OpenAI or Google. It came from an Austrian independent developer named Peter Steinberger, who published an open-source project called Clawdbot in November 2025 - later renamed (after a trademark complaint from Anthropic) to OpenClaw.

OpenClaw is an autonomous AI agent that runs locally on your own hardware, connects to any large language model, and integrates with tools like Slack, iMessage, and Telegram. It can execute code, browse the web, manage your calendar, and work on tasks in the background while you sleep. It accumulated over 175,000 GitHub stars - one of the fastest growth rates in open-source history. Nvidia CEO Jensen Huang called it one of the most important software releases in recent memory.

In February, OpenAI acquired OpenClaw, and its creator was subsequently hired. Nvidia has announced plans for an enterprise version called NemoClaw. Security researchers at Cisco have already flagged data exfiltration risks in third-party skill plugins, a reminder that viral does not equal enterprise-ready. China has also banned OpenClaw from government offices, citing national security concerns while underscoring that the governance questions around autonomous agents are now global.

What this means for insurance

OpenClaw matters not as a product you should rush to deploy, but as a proof of concept for where enterprise AI is heading: persistent, autonomous agents that work on multi-step tasks across multiple systems, without constant human prompting.

Imagine an agent that identifies a deteriorating risk segment in your portfolio, recommends an appetite adjustment, reprices the affected book, and begins implementing the change - without anyone having spoken to it. That future is closer than most insurance executives realize, and the organizations already building structured, system-level AI will be the ones ready to integrate agentic capabilities when the enterprise-grade versions arrive.

Story 4: Reasoning Is Now the Baseline

February 2026 was one of the most active model release months in AI's short history.OpenAI launched GPT-5.3-Codex on February 5 and GPT-5.4 on March 5. Google rolled out Gemini 3.1 Pro and upgraded Gemini 3 Deep Think. Anthropic introduced Claude Sonnet 4.6 and Opus 4.6.The unifying theme across all of these releases is a shift toward what the industry calls "thinking models." The model race is no longer about who can build a better chatbot. It is about stronger reasoning, better planning, better tool use, and more reliable performance on multi-step tasks. The frontier is shifting from answering questions to completing work.

What this means for insurance

This is the story I want insurance leaders to pay close attention to - because it connects directly to how value is created in this industry.Insurance value does not come from a single prompt. It comes from chains of judgment: assessing risk quality, determining the right price, identifying portfolio concentration, and deciding what to write and what to decline. These require context, documentation, and control - and that is exactly what this new generation of models is built for.

The real question now is whether AI can materially improve the decisions that drive P&L - what to write, how to price it, and where to grow - with enough reliability, traceability, and oversight to earn a seat at the table.Last year, I wrote about the difference between AI models and AI systems - and why professional-grade AI requires layered architecture, not just a smarter model. The reasoning gains in this generation of models make that systems approach even more powerful. Better reasoning capability, embedded within systems that enforce controls, learn from outcomes, and maintain auditability, is where the compounding value happens - whether or not a human is directing every step.

The models available today are categorically more capable than what most insurance AI proof-of-concepts were built against in 2022 and 2023. If your organization has written off AI based on a POC that underperformed three years ago, the technology has moved significantly. If you have a CTO or AI team, ask them specifically how they are evaluating these new reasoning models against your use cases.

The One Thing to Take Into Your Next Board Meeting

Read these stories together and a single theme emerges: AI has stopped being a question of capability and started being a question of implementation.

The models are good enough. The infrastructure is there. The regulatory direction - while still unsettled - is becoming clearer. What separates organizations extracting real value from AI from those still running pilots is not access to better technology. It is whether they have connected that technology to the decisions and processes that actually drive results.

If your AI investments are still operating as point solutions - a tool here, a pilot there - the competitive gap being opened by peers who have moved into integrated workflows will be difficult to close.

The window is narrowing. This should be a board-level conversation.