How can an accounting firm use AI without exposing clients' tax data?

An accounting firm can use AI without exposing clients' tax data when the language model runs on a server owned by the firm, consults only authorized folders and records what was done. The AI helps read documents, check filing obligations and organize routine work; the signature, the final calculation and the professional responsibility remain with the accountant.

How many client spreadsheets, bank statements and invoices have already been pasted into a public AI tool just to "save time"? This article explains where AI helps an accounting firm, why the place where it runs matters, and how to start with a small, verifiable scope.

DOUVRAS's local AI vertical for law and accounting firms lists accounting as a specialization, focused on bookkeeping, balance sheets and ancillary tax obligations. For this area, the Brazilian tax support product covers calculations, tax-saving opportunities and ancillary obligations, always as support for the accountant's work.

In day-to-day work, this becomes tasks like these:

Reading documents. Locating, in a batch of documents sent by the client, the information the accountant needs to check, and showing which file each item came from.

Checking obligations. Building, from the client records and the rules the firm itself defines, a list of what each client needs to file in the period, for the team to check.

Querying the firm's own knowledge. Answering internal questions based on the procedures, templates and guidance the firm has already written, citing the source document.

Items for review. Flagging entries, amounts or documents that look out of pattern, so that someone looks at them before closing.

What the AI does not do is sign a balance sheet, decide a tax classification or answer the client on behalf of the firm. A language model makes mistakes, sometimes in a very confident tone. In the design we use, its output is a draft for review.

An accounting firm holds some of the most sensitive data of a company and its partners: revenue, payroll, employee data, bank statements and tax returns. Much of it is personal data, and what is not is usually strategic information about the client.

The Accountant's Code of Professional Ethics (NBC PG 01), approved by Brazil's Federal Accounting Council in 2019, states in item 4, subitem (c), that the accountant has a duty to keep confidential what they learn in the course of their professional work, except in cases provided for by law or when requested by competent authorities. The LGPD, Brazil's data protection law, requires processing limited to what is necessary (art. 6), security measures (art. 46), and only allows international data transfers under the cases in art. 33.

DOUVRAS's page on local AI for law and accounting firms starts from these guarantees: local processing, with open models running on the firm's server and no external APIs; access control by user and by folder; encrypted communication and local backup with the firm's own keys; and usage records for internal auditing.

This does not satisfy the LGPD or the duty of confidentiality on its own. Purpose, client contracts, access control and staff training remain the firm's obligations. But it removes the most common risk: not knowing where the client's document went.

In Terry Gilliam's film Brazil (1985), a fly falls into a government printer, the name "Tuttle" becomes "Buttle" on a form, and the bureaucratic machine arrests the wrong man. No one along the way checks where the data came from; each department just passes on the paper it received.

Accounting firms live with this risk: information copied by hand from one system to another, from an email to a spreadsheet. DOUVRAS's automation and integrations service starts from exactly this problem, information that crosses the company through manual copying, memory and scattered messages, and delivers a map of the current flow and its failure points, integrations with retry handling, logs, alerts and contingency.

With local AI, the flow looks like this:

Input. The client's document goes into a folder on the firm's server, with access defined per user.

Reading. The assistant reads only what that task authorizes and extracts the requested information.

Answer with source. Each piece of information comes with the file name and the passage it came from, so the accountant can check it in seconds.

Record. What was asked, which documents were read and what was answered is recorded.

This design follows the scope of the AI agents and assistants service: a knowledge base with declared sources, tests, guardrails, an escalation flow, logs, costs and shutdown criteria.

AI speeds up reading and checking. Professional responsibility remains with whoever signs. In our design, three moments always stay with a person:

Before. Someone decides which folders and which clients the assistant may read, and who on the team may use it.

During. The accountant checks each piece of information against the cited document. Information without a source does not go in.

After. No deliverable, tax form, return or reply to the client goes out without review by a professional at the firm.

The common mistake is trying to put AI into every routine at once. A safer path:

Pick one routine. A repetitive task with a clear pattern, such as checking whether a group of clients' documents for the month arrived complete.

Define what a good result is. For example: the assistant flags the same missing documents the team would flag, and shows the source of each one.

Test with months already closed. Compare what the AI flags with what the team had already checked. This makes it clear where it gets things right and where it fails.

Define when to switch it off. If the error rate exceeds what the firm accepts, the assistant goes back for adjustment or out of use.

DOUVRAS serves professional services, such as firms and consultancies that sell trust and knowledge. In a local AI project for an accounting firm, we map the routine, install the model on the firm's server, connect the assistant only to authorized folders and deliver tests, operating limits, an escalation flow, logs and shutdown criteria.

DOUVRAS is not an accounting firm and does not provide tax consulting. The technical part is ours; the accounting and tax judgment belongs to the firm. Each firm has its own confidentiality and security needs, so the project starts with a diagnosis. If you want to see how the same design works in law, read how to use AI in a law firm without exposing client data.

And here is the question: in your firm, how many hours a month go into finding and copying information, rather than analyzing it?

Douglas Henrique, DOUVRAS Labs.

Does AI replace the accountant?

No. It reads documents, organizes information and flags what deserves review, always with the source. The signature and professional responsibility remain with the accountant.

Does client data leave the firm?

In the architecture we use, the model runs on the firm's server, with no external APIs, and consults only authorized folders.

Is running the AI inside the firm enough to comply with the LGPD?

Not on its own. It reduces sending data to third parties, but purpose, access control, security and record-keeping remain the firm's obligations.

Brazil. Law No. 13,709 of August 14, 2018 (General Personal Data Protection Law). planalto.gov.br

Federal Accounting Council (CFC). NBC PG 01 of February 7, 2019: Accountant's Code of Professional Ethics. cfc.org.br

Gilliam, Terry (dir.). Brazil. 1985.

DOUVRAS Labs. Local AI for law and accounting firms. douvras.com