Insights

OpenAI offered Washington a 5% stake — about $42.6B — and wants the rest of the industry to do the same. Meanwhile Google quietly rationed Gemini capacity to Meta, forcing its engineers to conserve tokens. Model supply now has politics and priorities attached. Neither of them are yours.

The 2026 Thomson Reuters report shows legal AI adoption nearly doubled in a year and now frees lawyers roughly 240 hours each. The capability question is settled. Only 17% of legal professionals trust AI to give advice and just 18% measure its ROI — which means the real advantage is no longer using AI. It's proving you can rely on it.

Microsoft's new Frontier organization puts $2.5 billion and 6,000 experts behind a single idea: enterprises don't need more AI tools, they need help getting AI into production. That's the clearest signal yet of where the value has moved — and it's good news for anyone still stuck at the pilot.

OWASP's 2026 report puts prompt-injection attacks up 340% year over year. A backdoored LiteLLM package sat on PyPI for three hours and was pulled ~47,000 times, poisoning CrewAI, DSPy, and Microsoft GraphRAG. The models aren't getting more gullible — we keep handing them untrusted input and privileged tools in the same breath.

In a single month, Visa completed hundreds of secure agent-initiated payments and Mastercard launched Agent Pay for Machines with 30+ partners. Juniper puts agentic commerce at $8B this year, growing to $1.5 trillion by 2030. The infrastructure for agents that transact has arrived — and the controls that make it trustworthy came baked in.

JPMorgan runs 500+ AI use cases in production. Klarna's assistant added an estimated $40M to profit doing the work of 700 agents. Enterprise deployments now average 171% ROI. The agents that win aren't running better models than everyone else — they're built on a discipline you can copy.

Gartner says that by 2027, 40% of enterprises will demote or decommission their autonomous AI agents — not because the agents failed, but because governance was binary: locked down or fully trusted. The fix isn't more control or less. It's control proportional to what each agent can actually touch.

Anthropic's Claude Fable 5 reached general availability on AWS on June 9. On June 12 a US export-control directive forced it offline for every user worldwide. The model was perfect; the dependency was the risk. That distinction is the whole job.

Amazon's April 20 announcement of a $25B Anthropic investment is being read as compute-arms-race news. The line that matters for builders is buried three paragraphs in: full Claude Platform, same AWS account, same bill, no separate contracts.

JPMorgan's late-February data shows agentic AI adoption at companies over $1B jumped from 11% to 26% in a year. Smaller companies are not seeing the same curve — and the reason matters.

Anthropic's November 12 announcement of a $50B custom data center buildout signals something bigger than a compute race. It is a bet on who owns the AI stack — and where the leverage lives.

MIT's 'GenAI Divide' report says 95% of enterprise AI pilots fail. The headline is true. The framing — that AI is failing companies — gets the causality backwards.