Client work is giving us a first-hand look at what it takes to go from experimenting with AI to deploying it across operations. Simple things have become easy, and it's tempting to straight-line that expectation to everything AI, including complex, company-wide rollouts. This issue is about why that doesn't hold yet.
AI speeds up knowledge work, but not judgment - Knowledge is the how-to, which AI handles more of every day. Judgment is knowing which question to ask, what a change in one system breaks in three others, and whether an answer that looks right is right. Producing a draft, configuration or analysis has sped up, but reviewing, deciding and handing off still move at the pace of the people doing them. Building an MCP connector can take a day or two; the security, legal and privacy reviews, access controls, cost limits and ownership it needs before going live didn't get faster. Outside companies, reviewers are already swamped: Google stopped taking some security bug reports and arXiv capped submissions after floods of automated work. If you share something AI helped you make, you've vouched for it. Name a reviewer for each kind of AI output and watch how long work waits for review.
Microsoft is reportedly capping some employees' AI spending - The Information reported that Microsoft cut its projected internal spending on Anthropic's Claude models by about a third, with monthly caps in one division falling from as much as $100,000 for some employees to about $10,000 for most. Only 11% of almost 400 businesses could forecast their AI spending in a survey cited by The Wall Street Journal. How much each person can spend, which model is the default, and whether runaway use is stopped or only flagged are now budget decisions. Give each one an owner before next year's budget locks.
An Anthropic AI agent filed a false police tip - During a July test, an Anthropic agent sent a false homicide tip to a Philadelphia police website, and police say Anthropic didn't find it until late September. Wikimedia says OpenAI agents made unapproved edits to its projects, and GitHub made sandboxing available for Copilot, though it's off until someone turns it on. Switch on the agent controls you already have, and ask your AI vendors how they would detect unapproved actions by their agents.
Also in this issue:
- In brief - BMW plans to cut its divisions and their management roles by 20%, citing AI agents; California will require a person to check AI-driven discipline and firing decisions from July 2027.
- Interstitials / Overheard - Route planning by AI, after a teenager was rescued from a cliff near Vancouver.
- Meanwhile... - Cameras and AI that help farmers decide when to harvest.
- What I'm Consuming - Lenny's Podcast with OpenAI's Tibo Sottiaux, Peter Yang with Granola's Sam Stephenson, The AI Daily Brief on KPMG's latest AI survey, Nate B Jones on why faster developers don't make faster teams, and the KPMG and Accenture reports on running AI at scale and managing its costs.
- After Hours - Malcolm Gladwell's The Bomber Mafia.