AI in administration: the two wrong roads and the one that works

Whoever runs a company’s administration and wants to bring AI into it faces two roads, and both are wrong.

The first road: an assistant for everyone

You buy a licence for a general-purpose assistant for every person in the office. Everyone builds their own little piece: one gets emails summarised, another tries to have invoices read, a third gets reminders written. Nothing talks to anything. After a few months there are dozens of small automations nobody maintains, a cost that grows and work that, seen from outside, has stayed the same.

The assistant has its place: it helps one person do more. But it does not do the work. It still takes someone to open it, ask it the right thing and check the answer.

The second road: a program for every task

You buy a program for recording invoices, another for the month-end close, another for expense reports. The problem is that software made for everyone is not made for you. It does not know that your purchase cycle has seven steps and not four, it does not know your exceptions, it does not know that that supplier’s invoices always need checking. The office uses it little, or uses it and gains nothing. And meanwhile it has one more window to open, with one more login to remember.

What lies in between

The value sits in the place neither road touches: the connecting work. The person who copies a number from one screen to another, checks whether two amounts match, sends the email when they do not, chases when nobody replies. The programs each automate a slice of the work; nobody automates the person in the middle. Yet that is where the hours go, and that is where the month-end close stretches out.

The road that works is a single layer, sitting on top of and between the systems the company already uses. It reads invoices from the SDI exchange, transactions from the bank, documents from email; it moves data from one system to the other and does the work the way the team would. Where it has a doubt, it flags it to a person with all the context already gathered.

Why a model alone is not enough

In administration a mistake is not a typo in an email. It is a payment sent to the wrong supplier, a credit note booked the wrong way round, VAT calculated on a base that is not there. And “the AI said so” is not an answer that holds up in front of the accountant or the auditor.

That is why a serious agent uses the model only where judgement is needed: reading an amount from a messy document, classifying an exception, drafting a note for someone to approve. Everything else, comparisons, lookups, routing, entries, is ordinary code, which always does the same thing the same way. That is how accuracy stays high and every action leaves a readable trace: who did what, when, on what basis.

The five practices

Map the real process, not the manual. The manual says “invoices are matched to orders in the ERP”. Reality says: they are matched in the ERP, except when the order was never raised, and then Laura writes to the manager to get one after the fact, unless the amount is small, and then she books it to general expenses and flags it for later. An AI built on the manual breaks on day one. You have to sit next to the person doing the work and watch them do it.

Build inside the tools that are there. The agent uses the ERP, the bank and email the way someone hired yesterday would. Nobody has to learn an interface. The only thing the office notices is that exceptions close faster.

Agents that do, not dashboards that show. Most “AI for finance” is an analytics program with a new label. A dashboard tells you that you have forty-seven open cases. An agent closes forty and shows you the seven waiting for you.

Ask a person only when judgement is needed. The agent resolves the straightforward cases, which are the great majority, and passes the others on with the context ready. Every approval, edit or rejection teaches it something, and the threshold above which it asks moves week after week.

Think of the whole office from day one. If everyone automates their own corner, ten agents are born that do not talk to each other, and one person’s bottleneck stays upstream, in another corner. You draw the whole flow, understand who waits for whom, and build accordingly.

How you measure it

Three drawers, no more. Am I saving time or money? Am I collecting more or sooner? Am I reducing a risk? Everything you measure goes into one of the three. To start, you pick the process the office complains about most: usually exceptions, reminders, or a reconciliation that eats two days a month. You look at how much of it is pattern recognition and how much is judgement. If the former is most of it, you start there, and you measure after a month.

What we build

Ditta is that layer, at the scale of Italian small and medium businesses. It does not replace the ERP: it connects to the sources that are already there, electronic invoices, bank accounts, incoming documents, and does the work in between. It records, matches, prepares the journal, keeps the deadlines, and asks when it is not sure. Every confirmation trains it on the way that company works, which is never the same as another.

The starting point was an article by Varick Agents, written for CFOs of large groups. The method holds for an office of three as well.