The four-step process I use for every AI project, from huge orgs to one-person shops

The short version: every AI project I deliver runs through the same four steps. Diagnose, Plan, MVP, System. The process is identical whether the client is a global organization or a barbershop, because the failure modes are identical: people buy technology before they understand their problem.

Step 1: Diagnose

No technology in this step. We map what exists: which data you have, what you already tried, what worked, what did not, and where hours and money actually leak every week. Most businesses discover the leak is not where they thought. A restaurant thinks its problem is marketing; the diagnosis shows three hours a day answering the same twelve questions on the phone.

The output is one page: the problems, ranked by what they cost.

Step 2: Plan

We pick the problem with the highest return and define success in numbers before writing a line of code. "Save the office manager six hours a week" is a plan. "Use AI in the business" is a wish. This is also where I calculate whether the project pays for itself, and if it does not, I say so and we stop. Saying no early is cheaper for everyone.

Step 3: MVP

Build the smallest thing that delivers real value, fast, and run it on your actual work. Not a demo with sample data, not a slide deck: your invoices, your customers, your WhatsApp. Weeks, not months. The MVP earns the right to grow by proving value, and it also surfaces the real-world mess (weird file formats, edge cases, the one employee who does it differently) while the project is still cheap to steer.

Step 4: System

The MVP becomes a system that works quietly in the background, owned by you. Feedback loops are added, the team is trained, and the thing keeps working when I am not in the room. I stay available for improvements, but the goal is independence: infrastructure, not dependence on a consultant.

Questions I hear about this

Why not start with the tool everyone is talking about? Because tools change every month and problems do not. A good diagnosis survives every model release.

How long does the whole process take? Depends on the project: an MVP typically lands within two to six weeks of the plan. The diagnosis itself is one conversation plus a short review.

What if we already know our problem? Great, then the diagnosis is short. It exists to verify with numbers, not to create paperwork.