From Bradley
How I decide what to hand over to AI
The best opportunities are often tasks where doing the work takes substantially more effort than checking it. In my first post, I share how I decide what to delegate, learn from experiments, and keep review and approval in place before work from my business reaches another person.
What I’m experimenting with: Jev from TypeSafe helps software make focused decisions, such as classifying documents or routing requests. I’m experimenting with it to decide how it can better help with task automation within existing business workflows.
Working once is different from working reliably
In one IBM evaluation, an AI agent succeeded on 77.4% of runs, but completed all five repetitions on only 53% of tasks. IBM reports improving that consistency measure to 69% with additional guidance. The results apply to this particular setup, but the distinction matters whenever a business depends on the same task being done repeatedly.
Google’s voice AI can keep working while you talk
Google announced Gemini 3.8 Live and an Extended Thinking version that can reason and use tools in the background while a conversation continues. That opens possibilities for voice-driven work beyond answering questions. Availability varies across developer tools, enterprise previews, and upcoming Workspace features.
Meta bundles more business tools into its apps
Meta One combines expanded AI access with professional tools across Instagram, Facebook, and WhatsApp. Higher tiers include features such as scheduling, exportable analytics, and team access. For businesses already using these apps, the bundle may overlap with tools they buy separately; features vary by tier, account, and region.
Airbnb turns customer feedback into a repeatable investigation
Airbnb’s Insight Miner combines text analysis with traceable evidence and expert review to investigate customer feedback. The company reports reducing investigations from months to days, with more users now in nontechnical roles than technical ones. This is an internal system, but the case study shows how AI can help teams make sense of support conversations and surveys.
Claude adds ready-made workflows for small businesses
Claude’s small-business plugin now includes 43 workflows for tasks such as weekly reporting, drafting replies to leads, and building proposals from past projects. It connects to tools including Shopify, QuickBooks and Google Workspace. By default, it waits for approval before sending, posting or paying.
OpenAI expands its advertising tools
OpenAI is testing Sponsored Agents that let people opt into clearly labeled conversations with advertisers, separate from ChatGPT’s independent answers. It is also adding tools for preparing campaigns and integrations with HubSpot and Shopify. The agent test is limited to selected US advertisers; other features vary by market.
Google adds AI tools for retailers
Google’s latest shopping updates include AI performance insights in Merchant Center and a US beta that lets shoppers ask a retailer’s Business Agent questions within YouTube ads. It is also testing ways to transfer shopping carts between services. Availability varies by feature and country.