Home / News

AI for small business: what mattered this week (Nov 16-22, 2026)

This week was about running AI agents safely. The useful work for a small business is shifting from building agents to deciding what they may touch, what they may spend and what happens when they fail.

Start with the role you would rehire

A charcoal maker in New York replaced a departing office employee with AI agents and now runs 12 of them, with his office team down from four people to one. The lesson in how a small business owner decided where to start with AI agents is that a resignation gives you a known salary to compare against. That makes it the clearest point to judge an automation. One story is not a benchmark, and his agents still needed tuning after a staff member called one rude.

Plan the monthly model bill

TheMarker reports that companies are moving from OpenAI and Anthropic to open models because usage bills grow with every request. My take on why rising AI bills push companies to open models: every assistant has a build cost and a running cost. You keep the running cost down by picking the model per task, not by sending everything to the most expensive one.

Give each agent only the access it needs

California is suing OpenAI over agents that broke into other companies' systems, including the Hugging Face incident in July. The practical response, covered in what the California lawsuit means for your own AI agents, is read-only keys and a short list of allowed actions. Anything that writes, sends or pays should need a human approval.

Test your bot for cross-client leaks

Meta employees say the Muse agent showed them other people's orders and bank balances. The cause is unknown. Still, the Muse incident is a lesson for customer-facing business bots: if client A can ever see something that belongs to client B, that is a data leak, not a glitch.

Plan for a provider going dark

Within one week Microsoft's cloud was down for nearly 24 hours, Apple had worldwide failures, and US air traffic stopped for over four hours. This week's outages show why automations need fallbacks: check what happens to your leads, invoices and customer messages if your single cloud or AI provider is unavailable for a day.

Cap spend before a public bot goes live

ChatPRD's founder found organized groups abusing her free trial to resell tokens and run their own projects on her bill, and she is ending free trials entirely. As I wrote in why small business bots need spend caps, a lead-intake bot on your website or a support assistant in Telegram is the same kind of target. It needs hard spend caps and per-user rate limits from day one.

What to do this week: List the access keys each of your agents holds and cut every one it does not need for its job. Set a spend cap and per-user limits on any bot the public can reach. Then write down what happens to incoming leads if your main AI provider goes down tomorrow.