AI in accounting and accounting firms after KSeF: what to automate and where to start

KSeF, Poland’s National e-Invoicing System, has above all changed how invoices reach the firm. Most purchase invoices from Polish companies arrive as structured data, not scans or PDFs. The accounting program downloads them directly, with no retyping.

For the firm to download a client’s invoices, the client must grant it permissions in KSeF. In practice, this is a new task for every new client and whenever staff change at the firm. It is worth keeping a register of permissions, because a missing one only comes to light at month-end close.

KSeF does not cover everything. Invoices from foreign suppliers, receipts, tickets, invoices to consumers and documents that are not invoices – contracts, bank statements, notes – remain outside the system. Until 31 December 2026, companies with sales of up to PLN 10,000 including VAT a month may also issue invoices outside KSeF.

We describe the deadlines and contingency modes in detail in a separate article on invoice automation after KSeF. Here we focus on what this means for how an accounting firm organises its work.

The table below shows the tasks where AI has a real use today, and what stays with the accountant in each. The order reflects how much time each task usually takes in a typical firm.

Note the last column. In none of these tasks should AI post entries on its own or send a client a reply on a tax matter.

An invoice from KSeF already has correct data, but someone still has to decide which account to post it to and which cost category to assign it to. In a firm with dozens of clients, each has a different chart of accounts and different habits. This is where AI delivers the most, because it learns from each client’s previous postings.

For example, the same fuel supplier is posted for one client as a passenger car cost with limited VAT deduction, and for another as a truck fleet cost. AI sees this difference in the history and suggests the right option. The accountant approves it with one click or corrects it, and the correction goes into the history.

It is worth setting a confidence threshold. Suggestions that AI rates as certain and that concern known suppliers are reviewed by the accountant in bulk. New suppliers, high amounts and unusual rates go into a separate queue for full verification.

Take a hypothetical accounting firm with 12 employees that codes 3,000 documents a month. Suppose a manual coding decision takes 1.5 minutes on average, and approving an AI suggestion takes 0.5 minutes.

The saving is 1 minute per document, i.e. 3,000 minutes, or 50 hours a month. Over a year this adds up to 600 hours, roughly three and a half months of one full-time employee’s work.

This is an estimate, not a promise. Plug in your own figures and remember two adjustments. First, at the start some suggestions will need correcting. Second, the time recovered only brings a benefit if the firm uses it to serve new clients or provide advice.

OCR is a technology that turns a scan or photo of a document into text and data. Before KSeF, it was the foundation of automation in accounting. Today its role is narrower, but it has not disappeared.

In a firm serving companies that trade internationally, invoices from foreign suppliers can make up a large share of documents. On top of that come receipts from business trips, tickets, bills from online platforms and contracts. Modern AI tools read these better than the old templates, because they understand the layout of a document, not just the position of the fields.

Before you invest in such a tool, count how many documents from outside KSeF you process each month. If it is a few percent, a function built into the accounting program or ordinary manual handling is enough.

In many firms, the greatest frustration comes not from the bookkeeping itself, but from collecting documents. AI compares the bank statement with invoices and flags payments that lack a document. It then prepares a single, clear request to the client listing what is missing, instead of several emails in the last week of the month.

The second part is client questions. Some are administrative, such as the tax payment deadline or the status of a return, and AI can prepare a draft answer to these from the firm’s data. Tax questions, such as whether something can be treated as a deductible cost, should always be decided by an accountant or a tax adviser.

It is worth writing this split into the firm’s rules. A wrong tax answer sent automatically makes the firm liable and undermines the trust on which the whole client relationship rests.

Many business owners would like to see a simple picture every month: revenue, costs, margin, receivables and payables. Preparing such a summary manually for every client is too labour-intensive. A Power BI report fed from the accounting program does it automatically.

AI can add a short comment to the report, for example: “transport costs rose by a fifth compared with the previous month, mainly due to two invoices from a new carrier”. The accountant checks the conclusions before sending. For the firm, this is a way to offer a higher-value service that clients may be willing to pay more for.

An accounting firm processes its clients’ financial data and the personal data of their employees and business partners. For HR data, it usually acts as a processor under a data processing agreement. If AI is to have access to this data, the tool provider becomes another entity in the chain, which must be reflected in the contracts.

Accounting services are not listed in the annexes to the NIS2 Directive. So, as a rule, an accounting firm is neither an essential nor an important entity under the amended KSC Act, in force since 3 April 2026. However, clients covered by the Act may require security measures from the firm in their contract, because they must look after the security of their suppliers.

An accounting firm is also an obliged institution under anti-money laundering (AML) legislation. AI can help pick out unusual transactions that need explaining. However, the risk assessment and any report to the authorities are always made by a person.

First, measure the current state: how many documents a month, how many of them come from outside KSeF, how long account coding takes and how long it takes to collect missing items. Without this, you can assess neither a vendor’s offer nor the result of a pilot.

Run the pilot on a few clients with different profiles, for example a service company, a trading company and a construction company. Only when the suggestions work well for them should you extend to more clients.