The short answer: Entering received invoices and receipts into an accounting system is now one of the easiest tasks for a small business to automate. A combination of OCR and language models can reliably read ordinary machine-generated documents—including suppliers, amounts, VAT, variable symbols and line items—and prepare them for import into an accounting system. However, the system should not make decisions on its own: handwriting, poor-quality scans and non-standard formats still cause errors, so human review and an approval workflow must remain between extraction and posting. Implementation is not about buying an off-the-shelf product. It is about integrating automation into the document workflow the company already uses: collection point → extraction → validation → approval → accounting import.
In most small businesses, received documents are still processed in much the same way: invoices arrive by email as PDFs, receipts accumulate in wallets, and some documents arrive on paper by post. Every so often, someone sits down and enters the details—suppliers, amounts, due dates and variable symbols. Someone else then approves the documents by forwarding emails, and at the end of the month the company tries to work out what is missing.
This article examines one specific issue: what machines can realistically handle today, where their limits lie, and what step-by-step implementation looks like. If you are looking for a broader overview of which business processes are worth automating, we have covered that in our article on AI automation for businesses—here, we take a closer look at the process that ranked first on that list.
Manual Data Entry vs. Data Extraction: What Is the Difference?
Manual document processing involves more steps than people tend to realise: opening an email, saving the PDF, reading the header, entering the supplier, IČO, amounts and VAT rates into the accounting system, copying the variable symbol and due date, classifying the document, sending it to someone for approval, and filing the original. Every step creates an opportunity for a typo—and an incorrect variable symbol or amount may not become apparent until payments are matched or the records are reviewed.
Document data extraction means that software reads the same information from the document and passes it on as structured data. Technically, this involves two layers:
- OCR (optical character recognition) converts an image or PDF into text. Software has been able to do this for decades—but text alone is not enough, because it does not tell you which number is the VAT-inclusive amount and which is the variable symbol.
- Data extraction turns the text into a structured record: the supplier, IČO and DIČ, net amounts and VAT by rate, total amount, currency, issue date, DUZP, due date, bank account number, variable symbol and, where applicable, individual line items. In the past, this required a separate template for every supplier. Today, language models can handle invoices the system has never seen before because they understand the document’s context rather than merely the position of a field on the page.
A third layer that is often overlooked is validation. Extracted data can be checked automatically before anyone sees it: do the net amounts and VAT add up to the total? Does the IČO exist in the ARES register, and does it match the company name? Is the bank account number in the correct format, and is it the same account the supplier used on previous invoices? Is there already a document in the system with the same supplier, variable symbol and amount—in other words, a duplicate? These checks are not performed by „artificial intelligence"; they rely on ordinary rules—and those rules catch most problems.
What Automation Can Realistically Do Today
For the routine workload of a small business, the technology is already highly capable. The following work reliably:
- Machine-generated PDF invoices—the vast majority of documents received by email. It does not matter whether the invoice was created by a large ERP system or a self-employed supplier’s invoicing service.
- ISDOC—the Czech electronic invoicing format. If a supplier sends an ISDOC file, there is nothing to extract: the data is already structured and can be imported directly. It is worth actively asking suppliers for ISDOC files, as many systems can generate them with a single click.
- Email collection. A mailbox such as invoices@company.com, to which suppliers and colleagues send documents, can be monitored automatically—the attachments are collected, classified and submitted for extraction without human intervention. The same collection point can also receive scans of paper documents.
- Receipts photographed with a phone—subject to the limitations discussed below. A flat, legible receipt photographed immediately after a purchase will usually be processed successfully; a crumpled thermal receipt that has spent a month in someone’s wallet is a different story.
- Suggested classification. Based on the supplier and the document’s contents, the system can suggest a cost centre, project or preliminary account assignment according to how similar documents were posted in the past. This is only a suggestion—the accountant confirms it.
- Purchase order matching. If the company records purchase orders, a received invoice can be checked automatically: do the supplier, line items and amount match what was ordered? A match can be approved easily, while any discrepancy is highlighted.
The practical consequence is that people do not disappear from the process, but their role changes. Instead of entering data, they review prepared information and decide how to handle exceptions.
Where the Limits Are and Why Human Review Remains Essential
An honest answer is just as important as an enthusiastic one. Today’s automation still cannot reliably handle:
- Handwriting. Handwritten receipts, added notes, crossed-out text and stamps placed over text. Handwriting recognition is improving, but it cannot be relied upon for accounting documents where every digit matters.
- Poor-quality scans and photographs. A receipt photographed at an angle, overexposed or cropped, faded thermal printing, or a document scanned through a plastic sleeve. OCR produces poor-quality output from poor-quality input—and the extraction system may then confidently process incorrect text.
- Non-standard formats. Advance invoices combined with final settlements, corrective tax documents, foreign-language invoices with unusual VAT structures, and documents with discounts allocated across individual line items. The system may read them, but interpreting them requires accounting judgement.
- Certainty. A language model will always provide an answer—even when it is unsure. A robust implementation must therefore use confidence scores: data extracted with high confidence and successfully validated can proceed directly to approval, while uncertain data is highlighted and held for human review. A system that cannot say „I am not sure about this" does not belong in accounting.
This leads to the fundamental rule we follow in every automation project: the machine prepares, a person approves, and the system posts. The goal is not to remove the accountant from the process, but to eliminate manual data entry while leaving review and exceptions in their hands. We discuss the principle of human oversight in more detail in our article on what is genuinely worth automating in a business.
What Step-by-Step Implementation Looks Like
Implementing document extraction is not simply a matter of installing software; it requires adjusting the document workflow. In practice, it involves six steps:
- Map the current process. Identify how documents enter the company (email, paper, supplier portals or employee receipts), who enters them today, who approves them and which system ultimately receives them. Without this step, you merely automate chaos—and automated chaos is still chaos, only faster.
- Create a single entry point. Set up a central mailbox for documents and agree that everything must flow through it: suppliers receive the new address, paper documents are scanned, and employees submit photographs of receipts to the same place. This organisational step may seem minor, but it determines the success of the entire project.
- Extraction and validation. Deploy the extraction system together with a set of checks: mathematical checks (totals and VAT rates), register checks (IČO against ARES and VAT payer reliability), historical checks (the supplier’s bank account number and usual invoice amounts), and duplicate detection. Define what can proceed automatically and what must stop for manual review.
- Approval workflow. Define who approves the substance of each document—typically the person who placed the order—and the amount above which a second person must also approve it. Approval moves from forwarded emails into a system that shows the status of every document and preserves an audit trail of who approved what and when. Simple rules are sufficient for a small business; what matters is that the rules exist and cannot be bypassed.
- Accounting integration. Once approved, the document is imported into the accounting system—Czech systems such as POHODA, Money and ABRA provide interfaces for machine import (XML, API or ISDOC). This is where the distinction between „automation„ and „semi-automation" becomes clear: if someone still has to re-enter the output into the accounting system manually, half the benefit has been lost. At the same time, the company and its accountant agree on what remains the accountant’s responsibility: checking preliminary account assignments, period-end procedures and non-standard documents.
- Pilot and parallel operation. During the first month, the new process runs alongside the old one: documents are extracted automatically, but the accountant checks each one against the original. Error types are recorded—which suppliers, formats and fields cause problems. The rules are refined accordingly, and only then is manual review limited to uncertain and non-standard documents.
The implementation is more about integration than conventional IT work: connecting the mailbox, extraction service, approval workflow and accounting system so that documents pass through without being transferred manually between them.
When Automation Pays Off—and When It Does Not Yet
You can make the decision without inventing figures by asking several straightforward questions:
| Signal | What it means |
|---|---|
| A specific person enters documents manually, and it regularly takes up part of their working week | A clear candidate—the time can be calculated from your own data: number of documents × time per document |
| Document volume is increasing due to an online store, more suppliers or more employees submitting receipts | Automation scales with volume; manual data entry does not |
| Payment-matching errors, overdue invoices or lost documents keep recurring | The benefit is not only time saved, but also better organisation and traceability |
| Approval relies on forwarding emails and assuming „we can find it somehow" | An approval workflow will deliver value even without data extraction |
| The company processes only a handful of documents each month | Extraction is not worthwhile—ISDOC files from suppliers and an organised mailbox will suffice |
| The process has no rules and nobody knows who approves what | Agree on the process first, then automate it |
One more factor belongs in the assessment: where the data goes. Invoices contain information about business relationships, prices and personal data. When selecting an extraction service, it is reasonable to ask where documents are processed and what data the service retains. For more sensitive workloads, processing under your own control—locally or on your own infrastructure—may be worth considering. As with any business use of AI, this is a choice between the convenience of the cloud and control over your data.
How We Do It at ITHOPE
Document data extraction is one of the automation services we provide to businesses at ITHOPE—alongside connecting systems that do not „talk" to each other and building AI assistants for company data. We do not start with a tool; we start by mapping the document flow: what arrives through which channel, who works with it and where the current process runs into problems. We then design the solution around the systems the company already uses, with validation and human oversight where they belong—and with a pilot that allows everyone to verify the process using their own documents before relying on the automation. We describe the service in more detail on our invoice automation page. If you would like to discuss what it could look like in your business, contact us—for an initial conversation, all we need to know is how many documents you process each month and which accounting system you use.
Frequently Asked Questions
Can the system extract data from invoices issued by a supplier it has never seen before?
For machine-generated invoices, usually yes—extraction based on language models does not require a predefined template for every supplier. However, it is sensible to review the first documents from a new supplier manually and confirm that the fields are being read correctly.
Will automation replace the accountant?
No. Automation takes over data entry and routine checks; the accountant remains responsible for correct posting, exceptions, non-standard documents and period-end closing. In practice, their work shifts from typing to decision-making.
What About Paper Documents and Receipts?
Paper documents are scanned into the same collection point as email attachments, while receipts are photographed with a phone—ideally straight away, while they are still legible. Illegible and handwritten documents remain subject to manual processing; the system should flag them rather than silently guessing.
Will We Have to Change Our Accounting Software?
Usually not. Common Czech accounting systems support machine-based document imports (XML, API and ISDOC), so the extraction solution can be connected to the software the company already uses. Changing software only makes sense if the existing system offers no integration interface.
How Long Does It Take Before the Automation Can Be Trusted?
A sensible implementation includes a pilot period during which every extracted document is checked against the original and the rules are refined using the company’s actual suppliers. Only once the error rate on the company’s own documents has fallen to an acceptable level should review be limited to uncertain cases—trusting the system from day one is a mistake.