“Automate our customer support” sounds like a project. In practice, it is a collection of dozens of decisions, exceptions, systems, and promises to customers. Treating all of that as one AI workflow creates ambiguity faster than it creates value.
01 The ambition trap
The most common failure pattern is scope without a unit of value. A team knows a function feels inefficient, so it points AI at the entire function. The resulting prototype can look impressive while remaining hard to trust, evaluate, or operate.
A workflow becomes buildable when you can name the repeated decision it improves.
“Support automation” is broad. “Draft a reply to delivery-status questions using the order record and shipping policy” is a workflow. It has known inputs, a recognisable output, and a straightforward way to check quality.
02 Find the decision loop
Useful candidates usually share four properties: they happen often, consume attention, use information already available to the business, and produce an outcome someone can review. Frequency creates leverage; reviewability creates safety.
The first version does not need autonomy. A system that prepares the right context and a strong draft may already remove most of the work while leaving the final judgement with the person closest to the customer.
03 Draw a hard boundary
Boundaries are not a limitation of the product; they are part of its design. Define what information the system may use, what actions it may take, what uncertainty looks like, and when it must stop.
This is especially important for small businesses. Their workflows often live across email, spreadsheets, memory, and specialist tools. A good implementation makes the boundary visible instead of pretending the underlying process is cleaner than it is.
04 Let value compound
Once a small loop works, it becomes infrastructure. The accepted drafts become evaluation data. The exceptions reveal missing policies. The usage patterns show which adjacent loop deserves attention next.
This is how an AI capability grows without becoming an uncontrolled programme: one legible system at a time, each earning the right to expand.
Start with the smallest loop that creates a useful outcome every week. Make it observable. Make it reversible. Then improve it with evidence.
The goal is not to place AI everywhere. The goal is to make the business work better—and to build only as much system as that outcome requires.