▸ AI in practice · 4 min read
Practical AI: what to build first
When a non-technical founder asks me where to start with AI, I never start with the technology. I start with the most painful, most repeated task in their week.
That is the whole method:
- Find the loop. The thing you or your team do over and over that drains time and adds little judgement.
- Write it down as steps. If you can describe it to a new hire in plain language, you can hand it to an agent.
- Build the smallest version. One task, end to end, good enough to trust. Not a platform. A single useful thing.
- Live with it for a week, then widen it.
The mistake is reaching for the most impressive thing instead of the most useful one. The win that changes how you feel about AI is small, boring, and yours: the report that writes itself, the inbox that sorts itself, the quote that builds itself.
Get one of those working and you stop fearing the build. That is the real unlock.
More on AI in practice
Automation that clears the operational noise
How small teams use AI to kill the busywork that used to need a whole ops hire.
Part 1: Two AI tiers on one Mac Studio, and the routing between them
First in a series on running local and cloud AI together. The economics piece answered whether to own your AI. This one is the wiring: a 93.7 GB Flash Next model, Kev decision models, a hosted frontier model, and Hermes deciding which one answers. Measured on the machine, with the parts that are still rough named.
Is it feasible to own your own AI models?
Everyone is preaching 'run your own model.' In the middle of a memory-price crisis, is owning your AI actually feasible? The honest math, and when it makes sense.