AI in accounting: what to delegate and stay compliant

AI can take over much of the preparatory and checking work in accounting: reading documents, proposing postings, matching payments and flagging suspicious items. A person should still make the final call on what gets booked and on anything that affects individual people.

What AI can take over and what it cannot

The split below is general guidance, not legal advice.

The more an output affects taxes, financial statements or a specific individual, the more thorough the human review has to be.

Payroll and people: where GDPR Article 22 applies

GDPR is the EU data protection regulation. Its Article 22 says that nobody should be subject to a decision based solely on automated processing if it has legal effects on them or similarly significantly affects them. Pay and employee evaluation are typical examples, because AI starts to influence decisions about people.

The ban has exceptions. A decision is allowed if it is necessary for entering into or performing a contract, if EU or member state law authorises it, or if the person gives explicit consent. For contract and consent, the company must guarantee at least three things: the right to human intervention, the right to express a point of view, and the right to contest the decision.

According to the interpretation published on gdpr-text.com, a decision is not solely automated when AI only produces a recommendation and a person decides, taking other circumstances into account. A formal human signature is not enough, though. The reviewer must be able to change the decision and must understand the underlying data. If an accountant simply clicks through whatever AI proposed, it is still an automated decision.

A practical example (illustrative): AI proposes a deduction or a bonus for an employee. The payroll accountant recalculates the proposal, verifies the inputs and is able to change it. The employee knows who to contact if they disagree with the result.

E-invoicing: a PDF in an email is not enough

The advisory firm DAÚC, in its commentary on the topic, points out that a PDF sent by email is only a digital image of a document. A genuine electronic invoice is a structured data file under the European standard EN 16931, in practice in UBL or CII format, both based on XML (a data format that computers read without human help).

According to DAÚC, from 1 July 2030 selected cross-border business-to-business transactions move to structured electronic invoices under the EU's ViDA package (VAT in the Digital Age, the modernisation of VAT rules). The invoice is generally to be issued within 10 days of the supply. It will be followed by digital reporting, meaning invoice data is passed to the tax authority within a short deadline.

According to the same commentary, however, ViDA does not say that every invoice between two companies in the same country must be electronic from 2030. Domestic invoices will be decided by national legislation. As a Czech example, the Ministry of Finance lists the following as the legal framework for e-invoicing: Act No. 563/1991 Coll. on Accounting, Act No. 235/2004 Coll. on VAT, Act No. 134/2016 Coll. on Public Procurement, the EU invoicing directive and Government Resolution No. 347/2017 on mandatory acceptance of electronic invoices. Whatever your country, check which national accounting and VAT rules sit on top of the EU layer.

It already pays to know whether your accounting system can both read and issue structured invoices. DAÚC recommends involving your ERP vendor (the supplier of your company-wide information system) early, because this capability will be one of the main prerequisites.

AI Act: what applies today

The AI Act is the EU regulation on artificial intelligence. According to businessinfo.cz, a website prepared by the Czech Association of Small and Medium-sized Enterprises and Crafts, it concerns every company that uses AI, not only those that develop it. What matters is how the AI is used and the level of risk. Company size is irrelevant.

Three points matter for an accounting department:

• AI literacy (Art. 4). Applies from 2 February 2025. According to regulation-ai.eu, Regulation (EU) 2026/1744, the so-called Digital Omnibus, rewrote it: a company must take measures to promote literacy but does not have to guarantee a specific level for every individual. The date stayed.
• Transparency (Art. 50). Applies from 2 August 2026, so it is already in force. For generative systems that were on the market before 2 August 2026, machine-readable marking of outputs has a deadline of 2 December 2026.
• High risk. Businessinfo.cz points to HR tools, customer scoring and biometrics. According to regulation-ai.eu, obligations for stand-alone high-risk systems under Annex III move to 2 December 2027.

The Czech Republic is preparing an implementing act that will set out supervision, procedures and penalties. According to the Czech state business support portal (Ministry of Industry and Trade and CzechTrade), the AI Act is directly binding and does not need to be transposed into Czech law. Expect a similar split in other member states: the regulation applies directly, and national law fills in who supervises and what the sanctions are.

A practical sequence for your company

1. Map your tools. List every AI tool the accounting and payroll teams use. Count AI features inside ordinary software and tools that people found on their own.
2. Name an accountable person. It does not have to be a new role. Extending the duties of someone who understands accounting is enough.
3. Write simple rules. Which data may be entered into AI, what AI may propose and what a person must approve.
4. Train and document. Record what the training covered and who attended.
5. Separate preparation from decision. For payroll and evaluation of people, decide who has the authority to change an AI proposal.
6. Check your accounting system. Ask the vendor whether it handles structured e-invoices under EN 16931.

Questions for owners and directors

• Do you know which AI tools your accountants actually use?
• Who in the company is responsible for their use?
• Does the person who approves AI output have the authority and knowledge to change it?
• Can an employee request human review if a decision affects them?
• Is your AI training documented?
• Can your accounting system read and issue structured e-invoices?
• Does someone regularly check a sample of AI outputs against the original documents?

AI agent use cases in financeData security when you deploy AI in your company

Prepared by AI from the listed sources and checked automatically against them. Not legal advice.