Nobody in a company wants to learn another program. They want the work done

Someone who has kept a company’s books for twenty years is not looking for an assistant. They have email open, the ERP open, the bank open, and a list of things to do before evening. If you put one more window in front of them, with an artificial intelligence inside waiting for a question, the only question that comes to mind is: and now what do I do with this?

That is why so many AI projects in companies stop after the trial. Not because the technology does not work, but because it is packaged as a tool to be used before the work gets done. A copilot, a chat, a button. The work stays with the person; the AI offers to help, if someone remembers to ask.

Adoption is the bottleneck

When AI is a tool, value arrives only if people adopt it and use it well. And people, in administration more than anywhere, work the same way for an entire career. Habits are not a flaw: they are what makes an office reliable. But they make it almost impossible for a new program, however brilliant, to become the place where work happens.

So the most predictable thing in the world happens. Two people really try it, someone opens it now and then, the majority go back to doing things as before. Three months later the company has one more licence and the same work as before.

Software sold an interface. AI sells the result

When a company bought an ERP or a CRM, it bought a window. The interface was the product, because the program was there to help people do the work. With artificial intelligence the relationship flips: nobody wants help recording invoices. They want the invoices recorded.

A result does not need everyone to open a window. It just has to happen. It is a difference that changes the whole way of building: instead of creating a new place to use the AI, you bring it to the places where the work already happens.

Build where the work happens

In an Italian company administrative work has precise addresses. Invoices arrive through the SDI exchange system. Transactions sit in the bank. Documents arrive by email, often as the attachment to a three-line message. Deadlines live in the F24 tax form and in the accountant’s calendar. The ERP is the one chosen ten years ago and it is not changing.

The AI that works reads from there, acts there and writes into the system the company already treats as the truth. No migration, no “export everything into the new program”. The person who recorded invoices keeps looking at the same screen, with the difference that most of the rows are already in place when they arrive.

In the background, but not out of control

Putting the AI in the background has one risk: that the company feels it has less control, precisely because it no longer has to do anything. The remedy is not one more dashboard. The remedy is for control to sit inside the work itself.

One person governs: they see what has been done, they can stop, correct, remove a permission. And the AI, at every step, stops before the decisions that call for judgement. An invoice that cannot find its order, a transfer that covers two invoices but does not match to the cent, a credit note nobody was expecting. In those cases it does not decide: it asks, in the place where the person already is, be it email or WhatsApp, with everything needed to answer in thirty seconds.

Every company does things its own way

There is one last thing manuals do not say. The way an office really works is not written anywhere. It is in the heads of the people doing it: “that supplier’s invoices are always checked by Laura”, “if the order is missing I write to the manager, except for small expenses”. An AI built on the manual breaks on day one, because on day one it meets Laura.

That is why the real work is not choosing the most powerful model. It is sitting with people, understanding how they do things and why, and building the AI around that. The technology is general; administrative work never is.

What we build

Ditta was born from this idea. The virtual employee works on the sources the company already has, invoices, bank, documents, ERP, and does the repetitive work on its own: it records, matches, prepares. When it has a doubt it brings it to a person, on WhatsApp or in the app, with the numbers in front of them. Nobody has to learn to write requests to a machine. Those who want to ask, ask as they would a colleague. Those who do not, find the work done.

The starting point was an article by Varick Agents, who put AI agents inside large American groups. Here we rethought it for the scale and the habits of Italian companies.