THE TAKEAWAY

The most useful starting point is the moment a person needs help—and what should happen when the AI cannot finish the job.

Start with one piece of work

A useful AI project starts with a specific job. “Help our sales team” is too broad to design or evaluate. “Prepare a response to an inbound service inquiry using our approved service information” is concrete enough to discuss. Name the person who benefits, the information available, and the next action they need. The clearer the job, the easier it is to tell whether the system is useful.

Write the boundary before the prompt

Decide which information the assistant can read, which actions it can prepare, and which actions it can complete. A customer assistant might explain an offering and collect a question. It should not invent a price, promise an unconfirmed delivery date, or expose another customer’s information. Write these boundaries in business language first. The implementation should make them real.

Give uncertainty somewhere to go

The assistant will encounter missing documents, conflicting information, unusual requests, and people who need a human. Design the handoff as part of the main experience. Who receives it? What context travels with it? How does the customer know what happens next? A handoff that ends in an unattended inbox is an unfinished workflow.

Evaluate the whole experience

Build a small set of representative questions, ambiguous questions, and requests the assistant should decline or escalate. Review whether answers use the right information, whether the next step is appropriate, and whether the human receives enough context. Keep the distinction between a convincing demonstration and a dependable customer experience.

Choose the first release

Begin with a narrow, valuable workflow and an owner who can review real behavior. Capture recurring failure patterns and improve the knowledge, rules, or interface that caused them. Expansion makes sense when the original job works reliably enough to support it. A broader agent is not automatically a more useful one.

About this field note

Original MKAIN editorial framework for planning a project. Examples are illustrative, with no claim of measured client performance. Published October 28, 2026; no corrections recorded. Send a correction or question.

Put the idea
to work.

Bring us the customer problem or workflow you want to improve.

Talk about your project