The common thread across this week's news for small-business automation: the hard parts are no longer the model itself but routing, credentials, access and ROI math on real tasks.
Agent-friendly websites are shrinking
Delta, JPMorgan, Yelp and Amazon are turning away browser-based AI agents, and users of Grok, Dots and Instinct report the same locked doors. If your plan was to point a general shopping or booking bot at these services, the breakdown on sites blocking AI agents is a cue to move that integration onto official APIs or partner deals before building further.
A real ROI number on cash reconciliation
A Schlotzsky's and Cinnabon franchisee cut daily cash reconciliation time by 94% and menu repricing by 95% with ChatGPT Work, freeing roughly $100K in idle cash flow and skipping any ERP rollout. The franchisee ROI benchmark is a useful anchor when picking which weekly task to move to AI first.
Cheaper classification calls
OpenAI shipped the Decisions API: fast, low-cost inference for the narrow moments when you only need a label, a category or a yes/no answer instead of a full chat completion. For routing tickets, tagging invoices and sorting leads, the Decisions API shift on small automations lowers the entry bar enough to justify pilots at low volumes.
Credentials for agents without pasting secrets
1Password for Agents lets an AI assistant sign into your CRM, admin panel or billing portal without the password ever touching the agent, with per-login human approval in the 1Password app. Launch partners include Claude Duet, OpenAI's Dots, ChatGPT Work and Grok Bot, so 1Password's credential delegation for agents is the first production-grade pattern to put between an agent and a sensitive account.
A new agent aimed at owners
Anthropic launched Claude Duet, a research preview on Opus 5.5 that watches connected apps, sends morning briefs and runs tasks under your approval, shipping on iOS, Android and web for US Pro or Max subscribers. The Claude Duet research preview is worth a short test if your routine already lives inside Claude connectors.
When self-hosting an LLM is worth it
n8n released a routing framework comparing open-weight models like Qwen3 8B, Gemma 3 12B and Mistral Small 3.1 against cloud APIs. The core finding in the self-hosted versus cloud routing guide: a GPU usually costs more than a year of cloud model calls, so hybrid routing only wins at steady heavy volume or regulated data.
What to do this week: Audit any automation that logs into Delta, Amazon, Yelp or a bank through a browser agent and move the integration to an official API. Pick one weekly task like cash reconciliation or invoice tagging, run it through a cheap classification model for a week on real data, and compare total cost against the current process. If any agent needs access to a CRM, admin panel or billing portal, wire credentials through 1Password for Agents before handing over the keys.