The firm does not sell software. What is described here grew out of obligations it has had to discharge by hand, month after month, which is why it knows where a machine must not be allowed to decide. The repetitive half of administration can be handed to a procedure. The half that needs judgement stays with a person, and the procedure reports to that person rather than replacing them.
This is an accounting and business consultancy that automated its own work first and offers the same thing to clients second. There is no product on this page, no licence to buy and no platform to be onboarded onto. Every procedure described below began as a duty the firm had to perform itself, with its awkward months, its documents that did not fit the pattern and its one supplier who invoices differently from everyone else. Performing a task by hand for two years is the only reliable way to learn which part of it is safe to hand over.
Dott. Olindo Cervi is an economist who spent years in information technology at IBM before building this practice. That is an uncommon pairing and it is what lets the firm stand between two worlds that rarely talk to each other: settled administrative method on one side, and on the other the tooling now reshaping it, artificial intelligence included. Someone who has only written software does not know which exception matters. Someone who has only kept books does not know what a machine can safely be asked to do.
So the first question is never which tool to adopt. It is which part of your month is genuinely identical every time, and that is answered by watching the work rather than by reading a brochure about it.
Each procedure exists because somebody here had to do the task manually first. The exceptions were met before the code was written, which is why the code knows to stop at them.
One task, automated properly and watched for a quarter, beats a redesign of everything. A procedure nobody trusts gets worked around within a month and then quietly costs money twice.
If the accounting package and the bank feed do their job, they stay. Most of the benefit is in the joins between systems, which is where the retyping and most of the errors live.
Your documents and your data remain exportable in ordinary formats. An arrangement you cannot leave is not a service, and we will not build one.
The tasks worth automating share one property: they are the same every time. The inputs arrive in a recognisable shape, the rule that governs them has not moved in years, and a competent person performing the task would reach the same result on every occasion. Those are the hours that can be given back without giving anything up.
The second list is shorter and much more important, because getting it wrong is expensive in a way the first list never is. Automation that removes control is worse than no automation at all: a task done badly by hand is visible and gets corrected, while a task done badly by a procedure is invisible and gets repeated, silently, until the year is closed on it.
The same document type, read and posted the same way every time. Purchase invoices, bank movements, expense receipts, payroll entries. High volume, low variation, no discretion.
The Italian administrative year is a fixed calendar. What falls due, for whom, on which date, with what reminder in advance, is a table and a table can be kept by a machine.
Bank against ledger, customer balances against invoices issued, supplier statements against invoices received. The matching is mechanical. What is left unmatched is the interesting part.
Naming, sorting and storing so that a document can be found years later by someone who was not there when it arrived. Dull, and the single largest time saving in most offices.
The figures somebody rebuilds every month from the same sources in the same order. Turnover by customer, ageing of receivables, position against budget, VAT position for the coming settlement.
Payment reminders, confirmation that a filing has been accepted, a note when an expected document has not arrived. Sent by rule, worded once, reviewed before anything difficult goes out.
Whether a cost is deductible, whether a claim is worth contesting, whether a customer is in trouble. A machine can assemble the evidence. It must not reach the conclusion.
If four files in ten need special handling, a procedure adds a layer to inspect on top of the work rather than removing any of it. Those tasks are better written down as a checklist and left to a person.
Some obligations change every budget season. Automating them means committing to update the procedure each time, and a procedure that encodes last year's rule is more dangerous than no procedure.
A checkpoint that exists on a diagram and not in somebody's week is not a checkpoint. If no one has the time to review the output, the task stays manual and visible.
Electronic invoicing is the clearest case, because Italy made it compulsory and in doing so turned invoices into structured data. An invoice between Italian parties travels as an XML file through the Sistema di Interscambio, the SdI, the exchange system operated by the Agenzia delle Entrate. The SdI validates the file, delivers it to the recipient and returns a receipt, and that receipt, not the paper copy, is the evidence that the invoice exists and when. Purchases arrive through the same channel, which means the data needed to post them is already machine readable on the day it lands.
The terms around it are precise and unforgiving. An immediate invoice is transmitted within twelve days of the operation it documents. A deferred invoice covering a month of deliveries goes by the fifteenth of the following month. A file rejected by the SdI has to be corrected and resent within a short window or the invoice is treated as never issued. None of that is hard. All of it is easy to miss in a busy week, and a procedure that watches receipts and flags a rejection the same day removes an entire category of penalty.
Deadlines are the second front. VAT settlements fall on the sixteenth of the month, withholdings on the same date, IMU on 16 June and 16 December, Intrastat summaries by the twenty-fifth, annual filings on their own calendar. A client with an Italian company, an Italian property and a foreign residence is subject to three overlapping calendars and no single office tells them so.
Reconciliation is the third and the least glamorous. Its value is not the matching, which any machine does well, but the residue. The five bank movements that match nothing, the customer whose balance is two invoices older than the rest, the supplier who has been paid twice. Automating the matching is worth doing precisely because it makes the unmatched visible.
Issue, transmission, receipts and rejections watched as a flow rather than as a pile. The rejection you notice on the day costs nothing. The one you notice in April is a penalty.
One calendar per client, built from what actually applies to them, with the document needed for each date named in advance rather than hunted for on the morning.
Movements matched to entries automatically, the remainder presented as a short list for a person to resolve. The list is the product, not the match rate.
Invoices issued against payments received, with ageing that updates itself. An overdue balance visible in week two is a phone call. In month five it is a legal file.
Invoices received against payments made and against what was ordered. Duplicate invoices and quiet price increases are both found by the same comparison.
The monthly set rebuilt from source each time, on the same day, in the same shape, so that two months can be laid side by side and the difference means something.
The order matters more than the technology. Almost every failed automation we are asked to repair was built in the opposite order: a tool was chosen, then the work was bent to fit it, then the exceptions were handled outside the system in a spreadsheet nobody maintains.
Measurement closes the loop and is the step most often skipped. Before anything is built we record how long the task takes and how often it goes wrong. Three months later the same two figures are taken again. If they have not moved, the procedure was not worth building and we say so rather than defending it.
Performed as it really is, by the person who really does it, including the parts they have stopped noticing. Most of the waste is in the handovers between people, not inside any one step.
A plain description of inputs, rule, output and exceptions. If the task cannot be written down clearly it cannot be automated safely, and the written version is already worth having on its own.
The portion that is identical every time. The rest is left alone deliberately, and the boundary between the two is documented so that everyone knows where the machine stops.
A named person reviews what the procedure produced, on a fixed rhythm, with the authority to switch it off. Time taken and errors made are compared against the figures recorded before it existed.
Control is not a feeling, it is an arrangement. You keep the original documents, in ordinary formats, retrievable without asking anyone's permission. You keep the credentials to your own accounting system, your own banking and your own entries in the Agenzia delle Entrate portals, and where the firm operates on your behalf it does so under a delegation you granted, that is written down and that you can revoke. Nothing here depends on a password only we hold.
Where the data sits is a question with a real answer rather than a reassuring one. Material relating to an Italian activity is kept on services hosted inside the European Union, and before your documents are loaded into anything you are told in writing which services those are and what each of them does. Access is individual and logged. Italian law obliges accounting records to remain available for years after the event, so the exit route matters as much as the setup: if a procedure cannot hand your archive back in a readable form, it is the wrong procedure.
Artificial intelligence is used inside this perimeter and nowhere else. It reads, extracts, classifies and suggests. It does not sign, file, pay or decide, and material that identifies a client is not sent to a service that would train on it. The distinction is not caution for its own sake. A tool that is right ninety-five times in a hundred is excellent at reading and useless at deciding, because the five remaining cases are precisely the ones that would have needed a person.
Invoices, contracts, receipts and filings stay yours in standard formats. A procedure is a way of handling them, never a place they are locked into.
Credentials to your systems and your portal positions remain in your name. Where we act for you it is under a written delegation with a scope and an end.
Hosting for client material stays within the Union, with the services involved named before anything is loaded, and access granted person by person rather than shared.
Records exportable in readable form, and retention long enough to satisfy the Italian obligation to produce them years later. Tested when the procedure is built, not when it ends.
Usually neither. The firm does not sell software and has nothing to place with you. Most of what is described here is built around the tools a business already runs, the accounting package, the bank, the invoicing channel, and consists of connecting them and removing the retyping between them. Where a new component is genuinely needed we say which one, why, what it costs and what happens to your data if you later stop using it.
No. It is used where it is good, which is reading: pulling figures out of a document, suggesting how an entry should be classified, spotting that this month's invoice from a supplier looks nothing like the previous eleven. Every one of those is a proposal that a person confirms. A model that is right ninety-five times out of a hundred is useful for reading and unacceptable for deciding, because the five remaining cases are exactly the ones that cost money.
Sometimes there is not, and we say so. The test is simple arithmetic: how many times a month the task is performed, how long each repetition takes, and how long the procedure will take to build and to keep working as the rules move. Below a certain volume the honest answer is that a written checklist and a shared calendar will serve you better than anything automated.
Accounting data for an Italian activity stays on servers inside the European Union, and the firm tells you in writing which services are involved before any of your material is loaded into them. Access is personal and logged rather than shared through a common password. Italian law also requires accounting records to be retrievable for years after the event, so a procedure is only acceptable if you can still get your documents out of it when the arrangement ends.
It is treated as part of the work rather than as an accident. A procedure is written down when it is built, so there is a document saying what it does and where it stops, and the checkpoints report when the inputs stop looking like what was expected. A rule change that arrives without warning therefore surfaces at the checkpoint rather than three months later in the accounts.
The ordinary year that automation is meant to make quieter, not invisible.
What a supervised procedure looks like once it is running in a real office.
The order of work, from the first document to the procedure that keeps it.
One repetitive task, described honestly, is enough to say whether it is worth automating, what the stable part of it is and where the human checkpoint would have to sit. The first assessment carries no obligation and no fee.