Private AI for accounting firms: tax questions answered from your own archive

The problem

An accounting firm answers the same questions every month. Is the cost of a company car deductible? How should a service bought from a foreign provider be treated? What applies to gifts for business partners? The firm has already written these answers, sometimes ten times over. They sit in e-mail threads, internal notes and the head of the most experienced accountant, who is busiest at month end.

General-purpose chat assistants write an answer fast, with two problems. First, the answer comes from the model's general memory, not from FURS guidance or your own opinions, so the accountant has to check it from scratch anyway. Second, once you paste an extract from a client's books or a contract into the chat, the document leaves your network and a third party processes it. For a firm that clients trust with payroll, contracts and bank statements, that is not acceptable.

What the private AI does

We set up a language model on the firm's server and connect it through RAG search to the material you choose: FURS guidance, your internal opinions, answers to clients and, where useful, individual clients' documents. When an accountant asks, the system first finds the relevant paragraphs, then writes the answer and cites the sources. If the material holds no answer, it says so instead of guessing.

A sample question and answer (a made-up internal note, shown for format):

Question: A client on the flat-rate sole trader scheme bought advertising from a foreign provider. What did we advise similar clients about recording and reporting it?

Answer: Last year we gave three clients in the same position the same procedure: record the invoice as a service received from abroad, check whether the client must file a separate report because of it, and inform the client in writing before the deadline. The internal opinion from March 2025 describes the procedure step by step and refers to a FURS guidance note you hold in the archive.

Source: Internal opinions/2025-03_services_from_abroad.docx, section 2; FURS guidance/services_foreign_providers.pdf, p. 4

The most common questions during a pilot:

• What did we write about taxing the benefit of a company car?
• Which FURS guidance notes deal with ZDoh-2 (Personal Income Tax Act) questions for sole traders?
• Find all our answers to clients about travel expense claims in the past year.
• Summarise what this client contract says about payment terms and interest.

The answer is a draft for the accountant, who checks and signs it. Tax decisions stay with the professional.

Data sources

We start the pilot with the one set that saves the most time. Most often that is:

• FURS guidance the firm already follows and stores, for questions under ZDDV-1, ZDoh-2 and other tax acts.
• Internal opinions and notes: answers to complex cases someone in the firm has already researched.
• Answers to clients: an e-mail or letter export, with anything that does not belong in the index removed.
• Individual clients' documents: contracts, resolutions and policies, accessible only to the accountant who handles that client.

We support Word, PDF (including scans), Excel sheets and e-mail. We do not connect the accounting software in the pilot, because ledger entries are numbers rather than text. If it makes sense, we add it in the production phase.

Architecture

Everything runs on the firm's server or in your own rented space in a data centre:

1. Indexing: documents are split into paragraphs and stored in a local search index; new guidance and opinions are added automatically from an agreed folder.
2. Permission-aware retrieval: each accountant gets answers only from general sources and from the documents of clients they handle.
3. Model: an open language model (we choose between the Llama, Mistral, Qwen and Gemma families depending on Slovenian-language quality and hardware) builds the answer from the retrieved paragraphs only.
4. Interface: a web chat, or an OpenAI-compatible API to connect tools you already use.
5. Log: questions and answers stay on your server for internal oversight.

The firm buys the hardware; we size it and do not resell it. For a team of up to ten people one server with a good graphics card is often enough; the options are on the private AI page.

Security and regulation

Because documents stay on your server, there is no transfer of client data to an outside model provider. The firm remains the controller or processor of clients' personal data, as it is today, and the General Data Protection Regulation (GDPR) applies to the AI just as it does to the accounting software.

During setup we agree on:

• access rights by client and by role,
• how long logs are kept,
• the procedure when a client leaves and their documents must be removed from the index,
• responsibility for security updates.

What the EU AI Act means for a firm that uses AI in its work is covered in our article The EU AI Act for SMEs.

Timeline and price

• Pilot (from 2,900 €, 2–3 weeks): one set of material, such as FURS guidance and internal opinions, a local deployment and an evaluation on 30 to 50 questions your clients actually ask.
• Production "Company knowledge" (6,000–15,000 €): several collections, per-client permissions, automatic ingestion of new guidance, monitoring and team onboarding.
• Fine-tuning (4,000–12,000 €): when you want draft replies to follow your firm's style.
• Support (290–790 € per month): updates, security patches and quality monitoring.

It makes sense to schedule the pilot outside the annual reporting season, when the team has time to test. In the first week we choose the material and collect 30 to 50 questions clients asked in recent months, together with the answers you sent at the time. In the second week the system answers the same questions, and the accountants compare its drafts with their own: is the answer correct, is the source right, how much editing would it need. In the third week you get a report with the results and an estimate of time saved per question. On that basis you decide whether extending it to more collections and the whole team is worth it.

Prices are final; the provider is not VAT-registered. What each tier covers is on the private AI for business page. Firms with a larger team often add an internal knowledge base for policies and onboarding new staff.

What we do not promise

• Tax advice. The system finds and summarises what is in your sources. The judgement and the signature come from the accountant.
• Up-to-date answers without maintenance. The model only knows what is in the index. If new FURS guidance has not been added, it does not know it. That is why automatic ingestion is part of a production deployment.
• Flawless answers. The model can link two cases incorrectly. The cited source is there so the error is spotted immediately.
• The quality of the largest cloud models on every task. Open models are very good at searching your documents and weaker at complex reasoning across several acts. We show the difference on your questions during the pilot.