I've been bullish on mass AI experimentation inside companies, but I also see practical limits to how far it should go, and ways to channel some of it into patterns that help the next experiment. That's the main essay this week.
Most AI experiments won't pay off. Make sure they leave something behind. - Experimentation without boundaries produces dashboards, command centers and other things that demo well, but few automations that change how processes run or move a number the CFO tracks. In Accenture's survey of 300 technology leaders at large companies, 70% use AI mainly to make individuals more productive, and only 37% use agents to run work end to end. Every disconnected experiment also builds its own plumbing, and some of it leaks: one software company found about a tenth of its AI bill went to stray API keys and workloads nobody had switched off. A better test is whether the tenth experiment is cheaper and safer than the first. That takes four unglamorous pieces of AI infrastructure: who and what can act, what the AI can trust, what it can touch, and a record of what it did and what it cost. Build them as your experiments show you need them, and back the experiments that change a KPI or add one of the four.
AI use is turning into a rationed budget line - Apple has reportedly cut some engineers' daily AI spending caps from $300 to $200, OpenAI will pool ChatGPT and Codex into one usage allowance per seat from December, and Hightouch cut the cost of an AI quality check by up to three-quarters by sending only the cases a cheap model was unsure about to an expensive one. Expect your AI bill to need the same management as cloud spend.
Who carries the risk when AI agents act alone - In Visa's test of 22 AI models on shopping tasks, agents failed 43% of runs, and about half of those failed runs still ended in a purchase. A New York law firm bought $5 million of insurance for its AI agents' mistakes so some work can go out without lawyer review, while most of the largest liability insurers have filed to exclude AI losses from standard policies.
Also in this issue:
- News Bits - The FT's John Burn-Murdoch finds the link between AI, productivity and job losses mostly inside US tech, with falling junior software hiring the clearest signal across countries; Claude can now read bank and card accounts, read-only, in a US beta.
- Overheard on X - A 3 a.m. call to Anthropic about Claude usage limits.
- Meanwhile... - A Mass General Brigham AI tool estimates a cancer patient's chances of survival from a photo of their face.
- What I'm Consuming - Anthropic on agent memory, Oskar Groth on running a one-person software factory, Spawn's founder on an AI-heavy company, Nate B Jones on agents meeting in the wild, and IBM on five concepts for coding with agents.
- After Hours - Blake Crouch's Dark Matter.