Range, a wealth-management firm, had a choice to make about who writes the instructions its AI agents follow. Engineers know the technology. Financial planners know how a good cash-flow analysis actually gets done. They picked the planners.

Six months in, Matt Ragone told AI Engineer NYC, staff had written 24 of these instruction sets, and the habit had spread from planning into sales, product and marketing without anyone mandating it.

His point applies to every small business: the AI can already do most of the work, and the bottleneck is someone writing down how your business does it.

Step 1: pick the right job

Start with work that repeats every week and that you can judge at a glance. Ragone’s advice was to pick the least complicated task that eats the most manual time, not the most impressive one. Reconciling a weekly report beats “handle all customer service”.

Step 2: teach it on real cases, out loud

Take three or four real past examples and walk the AI through them the way you’d walk a new hire through them, explaining why at every step. “I quote higher here because the access is difficult.” “This customer always asks about delivery, so I answer it up front.”

Ragone called this narration “the whole game”. A list of steps gets you something that runs. Your reasoning gets you something that makes the same calls you would.

Step 3: give it questions and limits

Two short lists turn a draft into something you can trust:

His line on this: write only the steps and you get something that runs by itself; write the questions and the limits and you get something you can trust.

Step 4: test it on your own past work

Run it five to ten times on real cases before anyone else relies on it. The best test material is work you’ve already finished. Rayan Krishnan of Vals AI made the same point from the other side: public AI benchmarks tell you very little about your business, while your own past emails, quotes and reports show exactly what good looks like. Give the AI an old inquiry and compare its reply with the one you actually sent.

Step 5: give it an owner and a second pair of eyes

Every instruction set gets one person who owns it, and a different person who reviews its output. “You don’t grade your own work,” as Ragone put it.

The owner matters because instructions go stale. Prices change, policies change, a supplier disappears. Ragone noted that an instruction set that was right in March can be wrong by December. Somebody has to notice.

Why this suits a small business

Large firms have to coordinate dozens of teams to do this. In a business of five or ten people, the person who knows the job best is usually a few desks away, and often it’s the owner. The technical setup is the part you can hand to someone else. The judgment about how your business does things can only come from inside it.

That’s how we work with clients: you bring the know-how, we turn it into an assistant that follows it, then keep it running. If that sounds useful, book a call.