An AI bookkeeping agent in an accounting firm collects the documents of every client in the portfolio, reads them and prepares them for booking. It downloads incoming invoices from SPV, matches bank statements to documents and proposes entries in SAGA or WinMentor. The accountant reviews, corrects and validates. Classification stays with them.


Primary bookkeeping is the part of the month no client sees and no firm can bill at its real value. Receipts photographed on the hood of a car, invoices sent through three channels, statements downloaded in different formats from different banks, handwritten receipts. Before an accountant can judge anything, someone has to collect them, read them and put them in order.
The AI agent does exactly this work, for every client, every month.
Receives documents on WhatsApp, by email or in the shared folder and sorts them by client, by month and by document type
Reads the data on each document, including from crooked photos, and flags what is illegible instead of filling it in from assumptions
Downloads incoming invoices from SPV through e-Factura for every client who has given you access and links them to the rest of the month’s documents
Matches the transactions in bank statements with the corresponding invoices and receipts and lists whatever is left unmatched
Prepares the proposed entries in the import format of the accounting software, for the accountant to review
Tell us what slows you down. We talk, then decide together whether and how we work.

Take a small services business in the portfolio. During the month, the owner sends photos of fuel receipts and supplier invoices on WhatsApp, whenever they remember. The AI agent receives them, reads them on the spot and files them in the month’s folder. If a receipt is cut off in the photo, it writes to the owner the same day asking for another picture, while the receipt is still in their pocket.
In parallel, the AI agent regularly downloads from SPV the invoices the business has received. These are the reliable source for purchase invoices, so it compares them with what the client sent and sees the differences right away: an invoice that came through SPV but that the client never mentioned, or an invoice sent by the client that doesn’t appear in SPV.
At the start of the following month, the AI agent asks for the missing statements and matches every payment and receipt. Then it prepares the package for the accountant: sorted documents, proposed entries and a short list of questions. The journal is not closed by the AI agent. The accountant goes through the proposals, changes what needs changing and validates.

The biggest waste of time isn’t data entry, it is waiting. The accountant starts on a client, discovers three invoices and a statement are missing, writes to the client and moves on to someone else. Four days later they come back and start all over again.
The AI agent knows which documents each client usually has, month after month. Rent, utilities, the phone contract, the main supplier. If this month the rent invoice is missing, it asks the client before the folder reaches the accountant. The list of gaps goes out early, with exactly what needs to be sent, not a vague message along the lines of “anything else?”.
For clients who keep a cash register book, the AI agent also checks continuity. If last month’s closing cash balance doesn’t tie to the first document of the current month, or if receipts are missing for cash takings, it asks early. These are details that, left for the last day, turn a simple month-end close into a whole evening of searching through paper.

SAGA, WinMentor or Ciel, through the file import each program accepts, for the proposed entries
SPV ANAF, with the access granted by the client, for downloading incoming invoices through e-Factura
Clients’ bank statements, in whatever format each bank exports
WhatsApp Business and email for receiving documents and requesting what is missing
Google Drive for the month’s archive, by client and by document type
We don’t ask the firm to change its accounting software. If a program doesn’t accept imports for a certain document type, the AI agent prepares the list in a form that is easy to enter and we say clearly what stays manual.

It doesn’t file returns, doesn’t sign anything and doesn’t submit to SPV on the clients’ behalf. It doesn’t close the month, doesn’t produce the trial balance and doesn’t prepare financial statements.
It doesn’t answer clients’ tax questions. When the owner asks whether an expense is deductible or what VAT applies, the AI agent forwards the question to the accountant and tells the client it has reached the right person.
It doesn’t fix bad documents. If an invoice has wrong details, the AI agent flags it and tells the client what to ask the supplier for. It doesn’t fill in information that isn’t on the document.

We start with a few clients, usually the ones who send the most documents on WhatsApp, because that is where the difference shows fastest. For each one we set the channels they send on, their usual documents and the SPV access.
In the first month, the AI agent sorts and prepares, while the accountant keeps entering data as before and compares with the proposals. Wherever the AI agent gets it wrong, the rule is corrected. Once reviewing the proposals takes less time than entering everything from scratch, we move to imports and add more clients.
Everything that means accounting judgment stays with the accountant. The AI agent suggests an account for an invoice based on that client’s history, but the suggestion is only a starting point. The accountant decides the classification, the VAT treatment, deductibility and when the entry is recorded.
The accountant also validates any unusual document: an invoice from a new supplier, a receipt without a tax ID, a payment to an individual, a foreign currency transaction, a document that doesn’t seem to belong to the business’s activity. The AI agent doesn’t book them on a guess. It sets them aside, with the reason it stopped.
The AI agent prepares and matches. Classification, validation and professional responsibility stay with the accountant.
The most important link on the page. From here you go up to the agent we sell.
Sideways, on the same level, so you can move easily between pages.
It prepares the import file in the format the program accepts, and the accountant reviews and imports it. We don’t write directly into the accounting software’s database. That way the accountant sees every entry before it reaches the books and can reject any line that isn’t right.
It reads the legible ones, even if they are crooked or poorly lit. When a figure or a tax ID isn’t clearly visible, it doesn’t fill it in from assumptions. It asks the client for another photo or marks the document for the accountant. A misread receipt costs more than one extra question.
It compares them with the documents the client sent and shows you the differences in both directions. An invoice received in SPV but not sent by the client may be a forgotten purchase or an invoice wrongly issued to their business. The AI agent doesn’t decide which it is. It puts the list in front of you and can ask the client about it.
It suggests an account based on each client’s history, for suppliers and document types that repeat. For any new or unusual document, it suggests nothing and leaves it to the accountant. The suggestion is a starting point, not a classification. Validation and responsibility stay with the accountant.
No. Clients send documents exactly the way they do now, on WhatsApp, by email or in the shared folder. The AI agent picks them up from there. The only thing that changes for them is that they get the message about what is missing earlier, with the documents named precisely, instead of being called on the last day before the deadline.
Tell us what slows you down. We talk, then decide together whether and how we work.