This week's four stories are about things leaders often can't see: AI bills that outrun the forecast, a vendor of your vendor holding your team's chats, AI-built apps shared with the whole internet, and new rules for which AI agents get through the door. Here's what happened, why I think it matters for your business, and what I'd do about each one. If you want a second set of eyes on any of these, talk with the Spej team.
Most large companies say their AI bills ran over budget.
On November 12, Capgemini published a snapshot of a September survey of 500 CXO-level executives at organizations with more than $1 billion in revenue: 59% said AI spending in the past 12 months was significantly higher than budgeted, those organizations ran 24% over on average, and only one in five calculates the verified value of an AI-assisted outcome after the full cost of delivering it safely. Capgemini sells AI consulting, so weigh that accordingly, then list every AI line item with one owner, count review and rework time, and track the cost per completed task on one workflow before you shop for a cheaper model.
A breach at an AI infrastructure vendor reached Notion AI chat transcripts.
Fireworks AI said a third-party cloud provider hosting its research machines had an incident on October 23, and the next day an unauthorized party reached Fireworks development machines; Notion then said the stolen data included Notion AI chat transcripts used for its Insights feature, affecting fewer than 1% of workspaces. Ask your top software vendors with AI features which providers process your data, what gets stored, and for how long, and tell the team in writing that passwords and API keys never go into an AI chat.
Thousands of AI-built work apps were open to anyone who looked.
Ars Technica reported on November 10 that security firm Claroc found 43,601 links for apps made with Grok Build, and more than 35,000 were public, including worker medical exams, a hospital staff roster, a sales call sheet, and customer proposals. Claroc said link sharing was set to anyone with the link by default, which a Cursor representative disputed, so ask every department which AI-built apps they have shared and what data is inside, then switch sharing to invited users or take the app down.
Platforms, airlines, and regulators started setting terms for AI agents.
On November 11, Google's Workspace MCP servers began rejecting requests from any app whose OAuth client is not marked as an AI agent, United said it will restrict unauthorized third-party agents from booking flights or using MileagePlus miles, and the European Commission proposed that a shopper choose whether an agent confirms every purchase, only some, or none. List the agents connected to company email, files, and calendars, give each a named owner and its own limited credentials, and set a dollar limit before an agent can spend.
First published on Spej's Substack.
