The first Monday of September you sit down at your desk and realize the summer was good for something. You know what an agent is. You know why the context window isn’t an infinite box. If someone tells you “we’ll train it on your data”, you no longer nod automatically.
Then you open your email, and there’s a vendor proposing to connect AI to the CRM. By the end of the month, he says.
And that’s when you notice the other thing: understanding isn’t deciding.
It’s not your fault, and it’s half mine.
All summer I told the story of AI in the company through cartoons. Rocco, a man in a tank top who knows everything about everything, and his four friends, who turned up under his beach umbrella at Ferragosto (Italy’s mid-August holiday) with real problems: one joke and one trap at a time. It worked, and I don’t regret it: a cartoon gets read, remembered, shared, and reaches people who would never open a serious document. But it also has a built-in limit, and it’s worth saying so now that the series is over.
A cartoon makes you understand. It doesn’t make you decide.
That’s why the series doesn’t end with an article. It ends with a 45-page document that I’m giving you today, and that you’ll find below. But first I have to tell you why you need it, otherwise it’s just another PDF you won’t open.
Three layers, and they don’t take the same effort
The day-to-day of anyone who has to bring AI into a company is made of three things, and they sit one on top of the other.
Words: you understand what people are talking about. It’s the level a summer of posts gets you to, and it’s also the easiest. Three weeks of attention and you’re there.
Questions: you notice you’re missing a piece of information, and you ask for it. That’s already a different job. Knowing what RAG is doesn’t tell you that you need to ask by what criterion the chunks of documents are selected, and how many. A post teaches you the word. A meeting gone wrong teaches you the question.
Let me tell you about a meeting like that, and it isn’t mine: it’s the one anybody who’s been doing this job for a few years describes. Perfect demo. The vendor asks the system a question, the system answers by citing the right manual at the right page, and everyone in the room relaxes. Nobody asks how many documents the demo was running on. There were 12. In production there are 4,000, and 300 of them are old revisions of the same thing: the system will pull from there with exactly the same confidence it pulled from the 12, and the wrong answer will have the same calm tone as the right one. The word “RAG” doesn’t save you from this. The question does: how many documents did you test it on, and what did you remove before testing it?
Use cases: you combine words and questions into a system that works on Tuesday, and on the Tuesday after too. Here nothing you know is enough anymore: you need decisions, in a precise order, made by specific people. And the most expensive decision is almost never technical—it’s who answers when the wrong answer has already reached a client.
Nobody argues anymore about whether AI belongs in the company. The September question is no longer whether to do it. It’s how not to get bogged down.
How far there is to go, really
Let’s make a deal before the list: what follows isn’t an inventory of the episodes. It’s the measure of the distance between understanding and deciding.
There are 25 words that end up in vendor quotes, and they fall into the three parts every AI quote is made of.
The ones that describe the mechanism they’re selling you: prompt and prompt engineering; training and cut-off; RAG, searching your documents; embeddings and search by meaning; fine-tuning; memory; system prompt; tools and function calling; context engineering.
The ones that describe what it costs, and how you test it: token; context window; reasoning; hallucination and bias; loop; evaluation and test cases; determinism and guardrails.
The ones that describe who answers if it’s wrong: agents and autonomy; GDPR and the data controller; where the data goes; training on your data; AI Act, and where you sit in it; traceability and the action log; what stays yours.
There are 11 places where AI really gets into a company that makes or sells something. In the technical area: company knowledge management, analyzing tender specifications, reusing past projects and jobs, technical support. In the commercial area: technical proposals for custom jobs, price-list quoting and configurators, draft replies to clients and requests, integration with the CRM, automated customer care. In operations and administration: extracting data from incoming documents, operational agents on the ERP.
Thirty-six entries. For each one there’s a decision to make, someone who has to make it, and a moment after which making it is no use anymore.
Now reread that list and ask yourself how many of those 36 you could say something about to a vendor. Not explain: demand. They’re two different verbs, and the summer only taught one.
This newsletter is too short
It’s not modesty, it’s arithmetic. Thirty-six entries, with the decisions inside, who makes them and how you notice you haven’t made them, don’t fit in 1,500 words. I tried: what came out was a list of headings, which is exactly the shape of the easy solution I spent the summer taking apart.
So I didn’t write it here. It’s the document I promised you at the start, and now I’m giving it to you.
It’s called “Managing AI in your company, without the noise”. It’s 45 pages. It has the 25 words and the 11 cases, one card each, all built the same way, and at the end 24 questions to ask a vendor, ordered by time instead of by topic: the ones to ask before you start, before you sign, before going into production, and then at every change. You print it, keep it on your desk, pass it to a colleague.
It’s a cheat sheet. A cheat sheet is what you write to get by without having studied—and this one is 45 pages long and has months of work in it. It’s a cheat sheet that isn’t a shortcut, and it couldn’t have been anything else: in this subject shortcuts are the problem, not the solution.
Download it here: Managing AI in your company, without the noise—PDF, 45 pages. It’s free, and now I’ll tell you why.
I didn’t write it
AI wrote it.
It’s not a provocation, it’s what happened. And before it sounds better than it is: what I put in isn’t just the signature.
I decided the thesis, the structure, which words went in and which stayed out, the order of the decisions inside each card—which is chronological, because time decides how much it costs you to change your mind. And I brought the raw material, which is the part no model could have.
The technical support card doesn’t come from a model: it comes from asking people who do that job what a service report really looks like. The answer is that half the time it says “board replaced” and nothing else—and as long as the reports look like that, no AI will help, because there’s nothing to learn from, and that’s the sentence that makes the difference between a useful card and a brochure. The one on CRM integration came from a list of six real annoyances, dictated by the people who deal with them every day, not from a feature list. And on one card I had to correct a statement that took for granted that suppliers constantly change their product codes: that’s not true, codes rarely change. What changes are the catalogs, and new products come out. It’s a difference that shifts the whole design of the use case, and only someone who works inside that trade knows it.
Then I rejected. A lot. I rejected more drafts than I approved.
A whole section rewritten from scratch because the heading of a group needed to be translated in the next sentence, and a label that has to be explained is a wrong label. A sentence cut because it defended a choice against an objection the reader hadn’t raised: it was our discussion, not the reader’s. A number deleted because I hadn’t measured it, and in a document that teaches you to demand evidence I couldn’t afford that. A whole page in which I declared the document was serious, thrown away: a document that asks the reader to distrust claims without evidence can’t make one on its last page.
It put in the time. I put in the no’s.
And that’s why it’s free: the price says how much it cost me, not how much it’s worth to you. Writing, laying out, drawing the illustrations, redoing everything 24 hours later because I’d changed my mind—that cost, which is usually what decides whether a document gets made or not, wasn’t there this time.
And what stays with me isn’t the PDF. It’s the method. Now I have a system that produces them, and the bottleneck has moved: it’s no longer writing, it’s having something to say and recognizing when what comes out isn’t that thing yet. I might start enjoying it, we’ll see.
And it applies to you too
Because look at what will happen when AI enters one of your processes. You won’t be the one writing the replies, the proposals, the emails to clients. Your job will be deciding which ones aren’t good enough—and it will be the only job left that you can’t delegate to anything.
It’s also the part you can’t learn by reading, and no document will spare you from it. But to say no to something, you need to know what the right thing looks like. That’s what the 36 entries are for: they’re not general knowledge, they’re the yardstick.
Download “Managing AI in your company, without the noise”—45 pages, free, to print. If you’re missing a word, or you have a case you can’t fit into any of those pages, write to me: we’ll think it through together and the guide will get better.
The summer cartoons are still on my profile, if you want to see where it all started.
AI puts in the time. You have to put in the no’s.