AI Disclosure

Where the service uses a machine-learning model, exactly what is sent to it, and how to turn it off.

Version 1.0 · In effect since August 19, 2026

There is exactly one place in vivaBooks where a machine-learning model is used, and it is described here in full. If you are looking for a general statement about our approach to artificial intelligence, this is it: we use one model, for one narrow job, and it is off unless you turn it on.

What it does

When a bank or credit-card statement is uploaded in bulk, the service has to work out which account and which month it belongs to. It first tries a deterministic parser that reads the statement’s own header. When that does not produce a confident answer, and only then, the extracted text can be sent to a language model to identify the bank, the account type, the last four digits and the statement period.

What is sent:up to the first 4,000 characters of the extracted text, the file name, and a list of the account-and-month slots the file might belong to. For a bank statement, the first 4,000 characters typically include the account holder’s name and address, the account number, and the opening transactions.

What is never sent: employee records, Social Insurance Numbers, pay data, T4 or Record of Employment content, and any document other than the statement being filed.

What it does not decide

The model reads; it does not choose. Which slot a statement is filed into is decided by our own rule from the dates the model reported, because the model repeatedly overrode the correct period with common filing conventions. A match above a high confidence threshold is filed without a person reviewing it, and is labelled in the interface as auto-accepted so it can be found and reversed.

Nothing here makes an automated decision about a person. No one is hired, paid, assessed or refused anything by a model in this product.

Training

Your data is not used to train models — not ours, and not our provider’s. We do not build models from customer content.

Turning it off and on

The feature is off by default. A firm that never enables it never has a document sent to a model.
A firm administrator can enable or disable it at any time in firm settings.
With it off, statement filing still works — the deterministic parser and file-name matching handle it, with more files landing in the unsorted pile for a human to place.

Provider

The model is operated by Anthropic, in the United States. It appears on our Subprocessorspage with the same detail as every other provider. Firms subject to Quebec’s Law 25 should treat enabling this feature as a transfer outside Quebec requiring a privacy impact assessment — see the Data Processing Addendum.