Most writing about AI for small business is speculative. This isn't. Here's what I actually shipped in the last year, what each thing does, and what I learned building it.
An operations layer for a home-services business
Miss Sparkle is a home-cleaning business on the Hibiscus Coast. It runs end to end on one custom system: marketing site, enquiries, quoting, scheduling, a field app for the cleaners, and the follow-up after every job.
The AI part is the least glamorous and the most valuable. Photographed job sheets, notes and routine admin get read and filed against the right customer automatically. Nobody retypes anything, and the same details never get copied between tools.
What I learned: the win wasn't the model, it was removing double entry. The AI is one step inside a system that was already designed properly. Bolt the same capability onto a business with no system of record and it has nothing to write to.
An assistant for a volunteer-run club
HBC Triathlon has trained on the Coast since 1993. The site holds a lot of changing information — events, training, results, conditions — maintained by volunteers who have day jobs.
Members can ask questions against the club's own information and get current answers, because the assistant reads from the same content the committee publishes. When they update a page, the answers update.
What I learned: grounding is everything. An assistant reading from your real material is useful. One reading from general knowledge is a liability, because it will confidently tell a member the wrong start time.
A family app that reads school notes
Oona is a subscription app with shared calendars, lists, chores and messaging. Its assistant turns a photographed school note into structured information — the date, the thing that's needed, who it's for.
What I learned: the interesting capability is rarely conversation. It's extraction. Turning a photo of paper into a record is worth more to a busy household than any amount of chat.
What it did to my own work
This is the part that changed my business rather than a client's.
A custom system like Miss Sparkle's used to mean a team. Design, build, integrations, content, search, then someone to manage all of them. With AI through my own pipeline I carry that work myself — the scaffolding, the boilerplate, the test coverage, the migrations, the parts that used to eat weeks.
That's not a marketing line, it's the reason my pricing looks the way it does. Custom software that used to be an enterprise budget now fits a small-business one, and there are fewer hand-offs to go wrong.
The honest part
Not everything worked.
The features I built because they were interesting got used less than the features I built because someone complained about a specific afternoon task. Every time.
And the boring plumbing beat the clever thing consistently. Extraction, routing and search paid off. Anything conversational needed far more care to be trustworthy than it looked like it would from a demo.
What I'd tell you from all of it
The pattern across every one of these is the same. Pick the specific job someone does every week and hates. Build that, properly, inside a system where the output has somewhere to go. Then do it again.
That's not a strategy document and it doesn't need one. It's just where the returns are.
If you want to talk through where that applies in your business, the first conversation is free.