Either AI runs in the systems you have, or nobody uses it.

Not with the technology. I ask you to point me at a process with three properties at once: it costs someone time every week, the data it needs already exists somewhere, and whoever runs it today can tell whether the result is right or wrong. Without the third it is not a use case, it is an experiment nobody will be able to judge.

The first question I ask is not "what do you want to automate", it is what happens when it gets it wrong. A draft reply a person reads before sending is an acceptable risk. A classification that releases a payment is not, and gets built differently. That distinction is what separates a project that reaches production from a demo that stays one.

Not having a structured AI project yet is not a delay: in 2025 only 15.7% of Italian SMEs used at least one AI technology (Istat, Imprese e ICT, December 2025). The real risk is not starting late, it is starting from a use case chosen because it was the one in the demo.

Your people already use generative tools, with or without your permission, and they paste in whatever is at hand: price lists, quotes, contract text, sometimes customer data. It is called shadow AI and banning it does not work: it only moves the use onto personal tools, where you can no longer see anything.

What can be done is to map it and give it a shape. Which tools are in use, on what information, and under which contractual guarantees: a business tool with a data processing agreement, exclusion from training and data kept in the European Union is one thing; a personal free account is another, and that is the one that actually exposes you. The rule that holds without banning anything: what the AI produces is a draft, not a decision, and the higher the impact the tighter the review.

The regulatory part follows from this rather than leading it: the European AI regulation asks you to know which systems you use and what for, and that knowledge is the same inventory you need in order to decide. Anyone selling you compliance before the inventory is selling the roof without the walls.

The threshold is simple and it is not technical: it depends on whether the work needs judgement. If the flow is "when this arrives, do that" and the rules fit on a sheet of paper, you need ready made tools like Zapier or Make, and it would be dishonest to charge you for code that replaces them.

An agent earns its place when someone today has to read, understand and classify before deciding: cases arriving in different formats, mail that has to be routed by working out what it asks for, documents to pull data from that is never in the same place. Rule-based automation breaks at the first exception there, and the exceptions are half the work.

How I build one: human oversight wherever a mistake costs, a record of what it decided and why, and the ability to switch it off without stopping the business. An agent you cannot switch off is not an agent, it is a dependency.

95% of generative AI pilots produce no measurable return (MIT NANDA, The GenAI Divide, 2025). Why that happens I have written out in full, with the sources, in the post on getting from pilot to production: here I answer only the operational question, which is what I do once the project has already stopped.

I always start from the real data, because that is where almost everything breaks: the prototype ran on an extract cleaned by hand, and what is missing is the current source nobody ever built. Next comes where the output has to land, and the right answer is almost never a new dashboard: it is the ERP or the inbox, somewhere someone uses it without changing their habits. Then the question that kills more projects than any technical problem: who inside the company answers for this still working next month. Without a name the system switches itself off, and nobody notices until it is needed.

Then I stay in the field until it really runs, which does not mean until the demo convinces anyone: it means until someone uses it on a Monday morning without calling me.

Because an AI assessment cannot be quoted before the state of your data is known. If the numbers live in five systems and disagree, anything built on top inherits that, and the quote would be a made-up figure.

So that is where it starts: Data Foundations has its prices published, and at the end of that work an AI engagement can be quoted on facts. If the data is already in order, the first call is free and half an hour is enough to see whether the use case holds.

There is also a free way to form your own view: the AI Readiness Assessment is a 24-question questionnaire across six areas, and it tells you where you stand before you talk to anybody.

There are two cases where I am not the right person. If you want a model trained on your data to sell as a product, that is research work and it needs a research team, not me. And if what you need is somebody to explain AI to the board across thirty slides, there are firms that do exactly that and do it better: I turn up when there is something to build that then has to be standing on Monday.