The first 90 days of an AI engineer in an Italian SME

Italian SMEs do not lack ambition. They lack legibility. Most of the company runs on Excel, on tribal knowledge, on the founder remembering which supplier had the best price three years ago. Walking in as the new AI engineer with a notebook full of model ideas is the fastest way to ship nothing.

Week one: do not write code. Sit with the office manager, with the production planner, with sales. Watch them work. Ask, when their attention flags, "what just took you twenty minutes that you wished took two?" The answer is rarely "an AI". It is usually a single Excel formula nobody has time to fix. Note it. Do not solve it yet — you are mapping.

Week two and three: produce one diagram. A single sheet of A3 showing how an order, or a quotation, or a delivery, moves through the company. Every department signs it. The moment everyone agrees on the diagram is the moment the AI roadmap becomes possible. Before that, you are building on sand. The diagram will reveal three things: a single workflow everyone hates and uses, a piece of information that exists in two places with two different values, and a decision that one person makes from gut feeling. Those are your first three projects.

Week four to six: build the prototype the office manager can run by hand. Not a chatbot. Not a model. A spreadsheet, a Streamlit app, a Python script with a CSV in and a CSV out. Show it to the people who do the work today. Ask them where it lies. Adjust. The point is not to deploy — the point is to compress six months of requirements discovery into six weeks of weekly demos. Every week you bring something they can break.

Week seven to ten: the integration. Now you need a backend developer or you need to be one. The model goes behind a .NET or FastAPI endpoint. The Angular dashboard learns to call it. The result lands inside the existing ERP screen where people already look — not in a new tab, not in a new product. This is where many "AI features" die. They get built as a parallel product and nobody opens them. The rule is: the AI lives where the workflow lives.

Week eleven and twelve: ship the first version, and make sure it can answer "why" in plain language. The factory operator who sees a recommended production plan deserves to know what drove it. The sales manager who sees a predicted delivery date deserves to know which assumptions are inside it. If your model cannot say "because" in plain Italian, you have built a black box and you will lose the operator the day it is wrong.

The biggest mistake AI engineers make in SMEs is not picking the wrong model. It is not respecting that the team running this company has survived for forty years without you. Your job is not to replace them. Your job is to make their tribal knowledge legible — first to themselves, then to a model that can amplify it. Skip that step and you ship a demo. Do it well and you become the first AI engineer they hire, and the last consultant they fire.