Hot take: "human in the loop" is often a way to avoid deciding what good output looks like. Write the acceptance rule first. Then you'll know which outputs actually need a human.

A client asked for an AI agent to answer supplier emails. We read 200 of them first. 170 were the same 6 questions. That's a template and a lookup table. The agent handles the other 30.

Most automations don't fail loudly. They keep running on yesterday's assumptions. A supplier changes a PDF layout. A field goes blank. Output still looks fine. Validate the input, not just the run.

Sounds like an AI we would train

A simple rule for AI in customer-facing messages: If you'd hate to see it screenshotted, a person approves it before it goes out. Everything else can go straight out. Then check a sample every week.

anthropic: an internal claude model generated 318,484 protein designs in 64 hours. 211 went to a lab. 17 passed. they published all 211, misses included. that's the part to copy. when a vendor shows you the hits, ask for the misses.

Replit made OpenAI the default for its agent. Claude is still in the menu. TechCrunch: most customers stick with the default. Until recently Replit was a top-five Anthropic customer. One default moved that spend. Pick your model on purpose. Keep it swappable in config.

IDC, 2,400 execs: 67% of companies with AI agents in production spend more than budgeted. nearly half can't measure the cost in real time. per-seat software had a fixed bill. agents bill per run. if you can't see cost per run today, the invoice will show you.

ornith's new open coding model: 122B, MIT licence. the model card says it "serves on a single 80GB GPU". the weights on that page: BF16, 125B params. that's ~250GB. the ollama build is 73GB. so, quantized. ask which version the benchmarks ran on.

Most companies don't need an AI strategy yet. They need a list of the tasks people hate most, sorted by how often they happen. Start at the top. Stop when it stops paying.

Anthropic studied its own staff. About 80% of interactive Claude session time is the agent waiting for a human. Reading everything their agents write would take non-technical staff half the work week. The bottleneck is the approver, not the model. Design fewer, better asks.

Five9 surveyed 3,000 consumers: trust in AI support is 26% with no human option. 55% when a person is reachable. 92% give the bot under 10 minutes before asking for a person. The "talk to a human" button isn't a failure. It's what makes people try the bot.

when not to use AI: the answer is one of a fixed list: a lookup table the rule fits in one sentence: an if statement the data already lives in a system: connect it use the model for the messy part. keep the boring part boring.

Microsoft's security chief, on its new agent monitors: "frontier models, regardless of their origin, cannot be trusted." Not malicious. Unpredictable. Nadella: "separate the supply of intelligence from the authority over it." Whoever supplies the agent shouldn't audit it.

Bank of America: average token prices are down ~50% since summer. Ramp data: US business AI spend has faded slightly since August while token volume keeps climbing. Bad news for AI investors. Good news if you're buying. Re-quote your AI line items every quarter.

dream labs' robot model API: $3 an hour. each call returns 1.6 seconds of actions, so about $0.0013 per action. their own launch page: a research beta that "has not been validated for safety-critical use." the cheapest worker on the floor. still needs someone watching it.

Anthropic: Claude writes ~90% of Claude Code's code. June to September, merged PRs more than doubled. The share reverted fell 40%+. Their goal isn't "a human read every line". It's less human time per PR, with reverts as the guardrail. If reverts rise, you only moved the work.

Holo, an AI marketing autopilot for small businesses: $5M ARR in 252 days, 20,000+ paying customers. That's under $250 a year each. About $20 a month. Small businesses don't buy "AI". They buy marketing done, with an approve button. At $20 a month it has to be boring.

Founders: what's the most expensive copy-paste in your company? Not the most annoying one. The one where a typo costs real money. Automate that one first. Not the one that would look best in a demo.

evan you: frontier models estimate a task at 2 weeks, then do it in 2 hours. a study this summer: coding agents overestimate runtime ~3-7x. given a tool that reports elapsed time, they got it right almost every time. give the agent a clock. don't let it quote your project.