AI in Accounting Firms - From Invoices to Tax Filings: What to Automate in 2026

The owner of a nine-person accounting firm in Poznań, 240 clients, is facing two deadlines at once in February 2026. The first: starting February 1, her firm - like every business in Poland - has to be able to receive invoices through the National e-Invoice System (KSeF). The second, less formal but just as pressing: a competing firm across town has been sending clients automatically generated monthly reports for six months already, while her team is still manually retyping line items from PDFs into the accounting system after hours to make the 20th-of-the-month deadline. This isn't a hypothetical - it's exactly where most Polish accounting firms find themselves right now, in the fall of 2026.

73% of accounting firms worldwide have already rolled out some form of AI automation, and among tax advisory firms adoption jumped from 9% in 2024 to 41% in 2025. At the same time, starting February 1, 2026, every business in Poland must be able to receive invoices through KSeF (the national e-invoicing system), and from April, issue them too. That's not a coincidence: mandatory e-invoicing and AI automation reinforce each other - just not the way most accounting firms assume. Here's the process map, what AI already automates well, what has to stay with a human signing the filing, and how to calculate ROI before you sign a vendor contract.

This post pulls together pieces I've covered separately before, one process: reading invoices from email into an ERP, OCR and document data extraction, AI integration with ERP systems, text-to-SQL for your own company database, and automated reporting with alerts. As with AI in recruiting and HR, individual techniques only click into place once they're mapped onto one industry's actual process - here, accounting, with one extra layer HR doesn't have: mandatory e-invoicing arriving on a fixed government timeline, and professional liability that doesn't disappear just because a language model did the work.

Mapping the accounting firm's process - four stages, four different automation opportunities

Before you buy any tooling, it's worth breaking the firm's work into stages, because each carries a different risk profile and a different automation ceiling:

1. 1.Inbound documents. Cost and sales invoices, receipts, bank statements, contracts - today a mix of PDFs, scans, phone photos, and, increasingly, structured XML files from KSeF.
2. 2.Coding and bookkeeping. Assigning a document to the right account, cost category, VAT rate, counterparty - the work that eats the largest share of a bookkeeping assistant's time today.
3. 3.Filings and tax returns. Month-end close, VAT filing (JPK_V7 in Poland), CIT/PIT returns, submission to the tax authority - the stage with the highest cost of error, because this is where the document's journey ends at a regulator's desk.
4. 4.Client communication. "How much do I owe in social security," reminders about missing documents, monthly summaries - work that doesn't directly bill, but consumes a disproportionate share of the team's time.

The biggest mistake I see firms make when they start automating: they focus entirely on stage 1 (because that's where the most visible tools live - OCR, scanning) and ignore stages 2 and 4, which is where the most hours actually leak out.

What AI already automates well today

The market data shows this isn't a fringe experiment anymore. Per AdAI News' 2026 roundup, the share of tax and accounting advisory firms using AI jumped from 9% in 2024 to 41% in 2025 - one of the fastest adoption curves in professional services. Large firms (51+ employees) are already at 89% adoption; small firms are still catching up at 68% - which tells you the first-mover advantage here is real, not cosmetic.

Teams running more advanced, agentic automation (not just OCR, but a system that carries out the next steps of the process on its own) report up to 70% faster invoice cycle times and a 76% reduction in processing costs, with touchless rates - invoices that go through the whole process without a human touching them - above 70%. That's no longer a cosmetic gap between the firm that invested and the one still waiting.

KSeF as an accelerator, not a replacement for automation - and a trap worth avoiding

The rollout schedule is already set in law, so it's worth having it in front of you:

Here's the thinking trap I run into with many firm owners: since invoices will start arriving as structured XML files instead of scanned PDFs, OCR must be becoming obsolete. That's only half true. KSeF removes the need for OCR for one category of documents - domestic B2B invoices issued after the obligation kicks in. It doesn't cover receipts, foreign invoices, simplified invoices under PLN 450 (during the transition period), historical documents predating the rollout, or notes, contracts, and bank statements. In other words, KSeF doesn't close the need for document automation - it relocates it: less work reading domestic invoices, more pressure to get everything outside KSeF's scope under control. A firm that invests solely in KSeF integration in 2026 and calls the topic closed will discover in 2027 that it still needs automation - just for a narrower, more scattered set of documents.

Penalties for skipping KSeF - what's real risk in 2026 vs. what only bites from 2027

Worth separating two levels of risk here, since there's a fair amount of confusion circulating. The administrative penalties under the VAT Act - up to 100% of the VAT shown on an invoice issued outside KSeF, or up to 18.7% of the gross amount for invoices without VAT - are suspended through December 31, 2026. Poland's Ministry of Finance has been explicit that the first year of mandatory KSeF is meant to be an implementation year. That doesn't mean 2026 is risk-free, though: Fiscal Penal Code (KKS) provisions apply immediately and allow for penalties around improper invoice issuance regardless of the VAT Act sanctions being paused. The full administrative sanctions take effect January 1, 2027 - that's the real deadline by which a firm needs more than a technical KSeF connection; it needs an orderly process covering every document type, not just domestic B2B invoices.

What AI shouldn't do without a human - professional liability doesn't automate away

This is the single most important boundary in this whole post, so I'll state it plainly: the tax filing is signed by a natural person - a bookkeeper, a tax advisor, or the business owner themselves - and that person carries full liability for the accuracy of the data, regardless of which tool helped prepare it. Tax authorities don't accept "the AI did it" as an excuse - the standard of due diligence applies regardless of the tools used. A tax advisor whose automation makes an error that lands in a client's filing can face civil and criminal liability for failing to supervise the tool they used.

In practice this means a specific division of labor, not a vague "AI assists, human decides":

• AI proposes, a human approves - coding, cost categorization, and draft filings always go through review by a licensed person before moving forward.
• An audit trail is mandatory, not optional - every AI decision (why this document landed on this account, why this transaction didn't trigger an alert) needs to be reconstructable, because that's the first question an auditor asks when something goes wrong.
• Submission to the tax authority is always a deliberate human action - never an automatic step without a checkpoint, no matter how well the tool has performed over the previous twelve months.
• Clients have a right to know where AI sits in the process - that's not just good practice; it also touches the AI literacy obligation under Article 4 of the EU AI Act if anyone at the firm operates these tools.

ROI per FTE - a concrete calculation before you sign anything

Take the firm from the intro: 240 clients, nine people, a realistic figure for the Polish SMB market of roughly 900 cost invoices a month to process manually (10,800 a year). At a manual pace of around 6,100 invoices per FTE per year (Ardent Partners' figure), that's work eating a noticeable share of a full-time position on invoice entry alone - not counting coding, corrections, and client questions. With automation pushing throughput to 23,000+ invoices per FTE per year, that same volume can be handled in a fraction of one person's time - and the freed-up time can shift to work clients actually pay a premium for: advisory, tax planning, optimization conversations, not retyping numbers off a PDF.

Numbers for your own calculation, instead of promises from vendor marketing materials:

1. 1.Count the real monthly document volume - cost invoices, sales invoices, receipts, separately for each category, because each has a different automation ceiling.
2. 2.Measure the current time per document - from receipt in the inbox to being posted, including corrections and follow-up questions to the client.
3. 3.Subtract the time automation won't eliminate - review and approval always remain; count the savings on entry and coding, not the whole process.
4. 4.Weigh it against the tool and implementation cost - license, integration with the accounting system, team time to learn and tune it in the first weeks.
5. 5.Measure the return in freed-up senior bookkeeper time, not junior headcount savings - a firm that automates to cut people loses institutional knowledge; a firm that automates to shift people toward advisory work grows revenue from the same team.

I break down the general version of this calculation, transferable to any process, in AI Automation ROI - this post just adds the concrete numbers for the accounting vertical.

GDPR and professional confidentiality - where accounting automation has hard limits

An accounting firm processes especially sensitive data from a GDPR standpoint - clients' payroll and HR data, bank account numbers, sometimes health-adjacent data tied to social security filings. Before connecting any AI model (especially a cloud one) to process client documents, you need exactly the same set of questions I lay out in GDPR and AI - DPIA, personal data in prompts: does the model provider train on your data, where is it physically processed, do you have a signed data processing agreement, and is a DPIA actually required at this scale of processing. For a firm serving dozens or hundreds of clients at once, this isn't a box to tick - it's the condition for keeping your professional credentials in the event of an audit.

Rollout plan - where a small firm should start, without flipping everything at once

Sequencing matters here - starting with the most impressive thing (full end-to-end automation) instead of the most cost-effective one is the most common reason automation projects in this field get abandoned:

1. 1.Start with one document type, not the whole process. Cost invoices from recurring, predictable vendors are the safest starting point - patterns are predictable, error risk is low. I walk through the full loop in AI reads invoices from email and enters them into your ERP.
2. 2.Only then expand to unstructured documents - scans, phone photos, receipts - where OCR and data extraction needs more tuning.
3. 3.Wire the output into your accounting system through [AI integration with your ERP](/en/blog/ai-erp-integration) - without this step, automation stops at "a nice-looking PDF with extracted data," and the team retypes it anyway.
4. 4.Add an ad-hoc question layer over the data through text-to-SQL - so a bookkeeper can ask "which invoices from client X are unpaid" without waiting for a report.
5. 5.Automated reporting and alerts come last, not first, as I cover in automated reporting with AI alerts. It's the most visible result to clients, but it only makes sense once the underlying data is already trustworthy.
6. 6.In parallel, not as an afterthought - prepare for KSeF and document where AI's authority ends in the process, before, not after, full sanctions take effect in January 2027.

---

I help accounting and finance firms design automation for the process from invoice to filing - with integration into Optima, Symfonia, or enova365, a control layer that respects professional liability, and a plan ready for full KSeF sanctions from 2027. I do this through AI consulting and AI automation engagements. Get in touch - I'll start by mapping your current process and pointing out where automation pays back fastest.

Worth reading next:

• AI Reads Invoices from Email and Enters Them into Your ERP
• AI Integration with ERP - Optima, SAP, Subiekt
• GDPR and AI - Personal Data, DPIA and LLM Vendor Contracts
• AI Automation ROI - How to Calculate Returns
• AI in Recruiting and HR - Automating CV Screening