SMB.co
Use cases for the AI agent
- Role
- Design Engineer
- Team
- Just me
- Timeline
- 2 weeks to launch
- Built on
- Cloudflare agents
I designed and built use cases for SMB.co’s AI agent: pages that each name a goal a small business owner already has, and agent skills that take them through it. It shipped in two weeks.
- Problem
- Owners used the agent a lot, but conversations ended without a next step, and owners who had never heard of SMB had no reason to start one.
- What I did
- A page template every use case is built from, the in-chat flow (live scan, report, goal and task cards), the integrations and the agent skills on Cloudflare.
- Key call
- Each use case is also a skill, so the agent can start it from any conversation. Goals and tasks are saved as cards in the chat, so owners never have to leave it.
- Result
- Three use cases were live when I left, and adding another takes a page and a skill. Owners went from 0 to 150 goals.
SMB.co is an AI agent for small business owners. It could already work with lists, rooms, goals and tasks, but owners had to know what to ask for. A use case starts from what they want instead: a page about one goal, like getting their café recommended by ChatGPT, that opens the agent with that goal already set.
I designed and built it, from the front end to the agent skills. It shipped in two weeks.
The problem
Owners used the agent a lot, but conversations ended without a next step. It answered questions without connecting them to goals, tasks or the tools that could do the work. And owners who had never heard of SMB had no reason to start a conversation at all.
What I built
- The use case pages: one template every use case is built from
- The in-product flow: the chat, the loaders, the scan and the cards the agent builds
- The integrations the agent works through, like Google, Search Console, Square and Gmail
- The agent skills on Cloudflare: when a use case applies, what to ask and which capability to call
Decisions
Every use case gets its own page
Owners don’t look for “an AI agent”. They look for a way to get their café recommended by ChatGPT. So each use case starts as a page that names that goal in their words, shows what they’ll get and drops them into the agent with the goal already set. Search engines and AI assistants read the same pages, so each one also brings in owners searching for that goal.
A wait you can watch
The check takes about a minute, long enough for a spinner to feel broken. So the agent shows its work instead: the business on a 3D map of its block, built with three.js on OpenStreetMap data, with each check ticking off beside it and the places assistants named instead showing up around it. By the time the report lands, the owner has already seen where it comes from.
Any conversation can start one
A page is only one way in. Each use case is a skill the agent can trigger whenever it notices an owner working toward that goal, from any conversation: what to ask, which capability to use and when to stop. It’s also why one use case can end by suggesting the next.
Owners never leave the conversation
When the agent creates a goal or tasks, the usual pattern is a link to the Goals or Tasks screen. That pulls the owner out of the chat at the moment they’ve decided to act. So goals and tasks show up in the chat as cards the owner edits and approves, and they’re saved there.
Outcome
Owners went from 0 to 150 goals created. The layer shipped in two weeks, and three use cases were live when I left, including Get recommended by AI and Find customers like your best ones.
Use cases run on capabilities the agent already has, so the next one is a page and a skill to write.