Vai al contenuto

AI for accounting firms

Documents that classify themselves, deadlines under control and faster answers to clients - with the firm's data staying the firm's.

The workload peak is predictable

Filing deadlines concentrate a month's work into a few days. AI pays off precisely there, on the repetitive tasks that cannot be postponed.

The data stays the firm's

Tools that do not use your documents to train models, with access set by client and a record of who consults what.

The professional remains responsible

AI prepares and proposes, it does not decide. Every output passes through verification, and that is a rule written into the policy.

Where AI saves real hours in a firm

Not in the technical part of your profession, which is what clients pay you for. In the administrative part surrounding it, which consumes time without producing professional value.

Classifying what arrives. Invoices, contracts, payslips, correspondence from the authorities, documents sent by clients in every imaginable format. A system that recognises what something is, which client it belongs to and where it should be filed removes a job that today somebody does, every day.

Searching the archive. A letter from two years ago, a clause in a contract, a client document nobody can find any more. Searching inside the firm’s documents returns the exact point and the source.

Preparing recurring correspondence. The client letters that repeat, requests for missing documents, reminders. A draft already set up rather than a blank page.

Summarising long documents. A set of accounts to read, a contract received, a circular. The summary says where to look.

The constraint that matters: you hold the keys to hundreds of companies

This is the difference between an accounting firm and almost every other organisation. The data you handle is not yours: it belongs to hundreds of businesses who entrusted it to you, and your responsibility towards them remains yours even when you use somebody else’s tool.

That has three practical consequences.

The tool must not learn from your documents. Services free to the public may use the content entered to improve the models. The check is made by reading the contractual terms and asking in writing, not by trusting the marketing page.

Access has to be set by client. If the assistant searches the whole archive, anyone using it reaches documents belonging to businesses they do not act for. The system has to respect the same permissions the person has.

Every consultation has to be logged. It is a protection for you before it is a compliance matter: if a figure turns out to be wrong, or something leaves where it should not have, the record says what happened.

What AI must not do

It does not do the arithmetic. Language systems are not calculation tools: they can produce wrong numbers with the same confidence as right ones. The calculations stay with the practice software.

It does not decide. It prepares, searches, proposes. Professional responsibility does not become delegable because the text came from a system.

It does not receive everything. There are documents that should not be entered into any external tool, and the list has to be written beforehand rather than left to the judgement of whoever is in a hurry at month end.

The calendar decides when work can happen

An accounting firm does not have a steady workload: it has predictable and immovable peaks. From that follows an operational rule that applies to AI as to any other change: new things are introduced away from the deadlines.

It is also why AI pays off more in a firm than elsewhere. In the peak weeks the bottleneck is not technical competence but the time spent sorting documents, chasing missing attachments and answering repeated questions. That is exactly the part that can be made lighter.

How we do it

First we look at what is already happening. In almost every firm somebody has already started on their own, pasting documents into a public service. The first step is not to ban: it is to know.

Then the archive gets put in order, where needed. A document assistant works well over documents organised with some logic. That work is worth doing anyway, with or without AI.

Then we write one page of rules: what may be entered, who checks, with which tools.

Finally we train the people who will use it. The mistake that costs most is not technical: it is the person who, in good faith, pastes a client’s accounts where they should not have.

The rest of the firm’s IT

An AI project only holds up if there is order beneath it: individual logins rather than one shared account, client credentials kept in a manager rather than a spreadsheet, backups that genuinely restore.

See also IT, AI and security for accounting firms, AI for professional firms and AI Act, GDPR and AI governance.

The first step

An analysis of the firm’s document flows: where time is lost, which activities genuinely lend themselves to AI, and what is worth putting in order first. Free and without obligation.

Frequently asked questions

Can an accounting firm use AI on client documents?

Yes, provided the tool does not use the content entered to train the models and access is set by client. It is also worth remembering that a firm handles data belonging to hundreds of companies: the responsibility towards those clients is yours, so the boundary has to be defined in writing before any tool is enabled.

Can AI do the calculations or complete the returns?

No, and be wary of anyone who promises it. Language systems are not reliable calculation tools and can produce wrong numbers with great confidence. Where AI genuinely pays off is on the documentary and organisational side - classifying, searching, summarising, preparing correspondence - while the calculations stay with the practice software and the professional.

What can genuinely be automated?

The three things we see working best are classifying incoming documents, with automatic attachment to the right client and matter; searching inside the firm's archive, which returns the exact point and the source; and recurring client correspondence, which starts from a draft rather than a blank page.

How long before we see a result?

The analysis of document flows closes within a few days and says where the most time is lost. The first useful uses start within a few weeks. It is worth avoiding the deadline windows, though: introducing a new tool in the week of a filing deadline is a mistake nobody makes twice.

Does anything have to be put in order first?

Almost always yes, and it is the part nobody talks about. An assistant searching the archive works well if the archive is coherently organised; if documents sit in folders under different conventions accumulated over the years, that is the first useful job - and it is worth doing anyway, with or without AI.

AI in your firm, starting from what you actually do

Every firm has different repetitive work. Tell us about yours and we will say which parts lend themselves to this and which do not, without selling you tools you do not need.