Custom AI systems
for Italian SMEs.

The enquiry that came in on Saturday morning and got a reply at nine on Monday.

They arrive on WhatsApp, through the website, by email, by phone. The first to answer wins the order.

The quote that was sent and never followed up. The customer who bought only once.

Nobody said no. That contact was simply never given a next step.

The March invoice, paid in May.

Overdue invoices get chased when someone remembers, and someone remembers when the day leaves room.

Data retyped from one spreadsheet into another, with the odd error that slips through and turns up later.

A machine would do it exactly the same way, and it wouldn't skip a line when the phone rings.

They search for exactly what you do, and click on someone else's website.

It happens on search engines and inside AI assistants, which increasingly give an answer instead of a list of results.

The renewal spotted the day after, the grant you read about once it had closed.

Nobody checks them, because they're nobody's job.

A hundred real enquiries: orders, reorders, complaints, customs questions, reminders. Four channels, five languages.

A hundred real enquiries across four channels and five languages: this is the time from arrival to a draft ready for approval, as a median over the hundred. Twenty that arrived in the same second were all closed, every one of them, in twenty-one seconds, which is exactly when a person would have to choose which one to ignore. A hundred in, a hundred out with their replies, and the right priority assigned a hundred times out of a hundred: complaints and customs at the top, payments and junk at the bottom.

By hand, the same enquiry takes four to ten minutes, and that's a stated estimate: read, understand, look up the data, write, reread. The seconds are timed; the minutes aren't.

Thirty real business documents: invoices, delivery notes, spec sheets, orders. Five languages, three different layouts.

Two hundred and thirty-four fields right out of two hundred and thirty-seven, and the document type recognised thirty times out of thirty: invoice, delivery note, spec sheet or order, in five languages and three different layouts. We had deliberately hidden an unreadable one in the pile, and it set it aside rather than making up the numbers.

The three missed fields have a reason, and here it is: twice the same issue, a zero amount left blank instead of written as zero, and once a country that wasn't on the document and had to be inferred. None of the three is an invented value.

Open the PDF, find the fields, key them into the business software, double-check.

Six point three seconds per document, median across thirty. By hand, the same document takes three to eight minutes, and that's a stated estimate: the seconds are timed, the minutes aren't.

A contact writing on WhatsApp, read and queued with the next step already assigned.

Median across nineteen real contacts, not simulated ones. No one is left without a next step, and that's the point: leads don't go cold on their own, they go cold when nobody touches them.

The founder of a management SaaS for the beauty industry was looking for Italian professionals to collaborate with, within a specific follower range. Five professions, six cities, no bought lists.

Twenty-two searches, three hundred and eighty-nine names collected, ten delivered in priority order: 2.6%. In between, two filters, the thresholds he set himself, and a profile-by-profile read of the thirty left. One hour and four minutes from the first search to delivery. By hand, the same job takes a week, and somewhere around the two-hundredth name you stop reading and start scrolling.

Three of the ten had already put the problem in writing, in public: one had asked customers to message him again because he'd lost some of their messages, and two take bookings by hand. For those three, the first message was already written, built on what they had posted.

Five public exhibitor lists. No contacts bought, no paid lists.

Twenty minutes in total, for work that takes a salesperson a whole day. Four hundred and twenty-eight names read, sixty-four websites found and confirmed one by one, and ten emails written, each built on a flaw verified on that company's own website: none of the ten could have gone to anyone else.

A maker of tube-bending machines, a restaurant, a domain for sale, a Chinese gaming page.

Nineteen companies had the right name and someone else's website. They were caught before they got onto the list, not after the email had gone out: the step that caught them checks each website against its content, and it's the most valuable step in the whole chain.

An external check asks the machine whether the systems are up, every five minutes, day and night.

An external check has asked the machine whether the systems are up every five minutes, day and night, for a hundred and twenty-four days: roughly thirty-five thousand seven hundred checks. It runs from outside, so it doesn't depend on the machine it's watching.

Underneath, every single process is logged: when it started, what it worked on, how it ended. Three hundred and fifty-six log entries between 18 May and 26 August. No engine runs in the dark, and when something breaks, a person responds.