GUIDE · AI FOR DEPARTMENTS

AI for administration and accounting: document checks and recording

After this guide you will know how to take a document from receipt to posting with an assisted step, which rules to write before you automate, how to handle uncertain documents, and which checks must always stay with a qualified person. The starting point is always a set of written rules.

Updated october 2026
7 min read
Altovar Technical Editorial Team

1. Where AI genuinely helps in administration

In accounting firms and administrative functions, a large part of the work consists of receiving documents, reading them, classifying them and posting them. Supplier invoices, receipts, expense claims, consultancy contracts, payment schedules: repetitive material, often in different formats. An AI system can extract the main fields, suggest a classification and flag incomplete documents. This reduces the time spent on manual entry and leaves the judgement-heavy work to the person posting.

The real benefit, however, depends on the quality of the input. A crooked scan or a rushed photo of an expense receipt produces errors the system does not always flag. The value is therefore also measured by the ability to say no: a document that cannot be read properly should be sent back, not half-posted. Ask the provider how the tool handles uncertain documents before choosing it.

2. Prerequisites: posting rules and responsibilities

Before you automate anything, the office must have written rules. Which expense categories exist, what the exceptions are, who approves postings above a certain threshold, how late documents are handled. If these rules exist only in one person's head, the system cannot apply them and the error will repeat. The chart of accounts and internal procedures are the starting point, not a detail to sort out later.

It is also important to clarify who is responsible for what. The system can propose a posting, but the professional sign-off and the responsibility towards the client or the company remain with a qualified person. Define this in writing. Finally, check which personal data the documents contain, such as payslips, bank details of suppliers and the contact details of private individuals, and who can access them. This too must be written down before you start, not reconstructed afterwards.

3. The operational steps

The process can be divided into short phases, each with a check. You do not need to automate everything at once: it is often better to start with a single document type, for example invoices from recurring suppliers, and widen the scope only after a trial. This approach makes it easier to see where the system goes wrong and who needs to correct it. Each phase should have a named owner.

Every phase must have an output that a person can read and a simple way to reverse a posting. If the system proposes a wrong category, the correction must be immediate and traced, not a complicated technical intervention. Keep track of corrections and their reasons: it is the most useful information for improving the rules. Review this information monthly with the owner.

  1. Establish the list of incoming documents and the channel they arrive through: post, direct upload, supplier portal.
  2. Define the mandatory fields for each document type, with written validation rules.
  3. Have the system extract the fields, with a doubt flag where the text is hard to read.
  4. Set the posting categories and the list of exceptions that always require a person.
  5. Set the threshold above which a posting requires approval before being confirmed.
  6. Run a periodic sample check on postings already confirmed.
  7. File the original document together with the posting, with a record of who confirmed it.

4. Mistakes to avoid

The first mistake is to confirm everything automatically to save time. A wrong amount or date extracted becomes a real accounting entry, and it is often discovered only at a later check. The second is failing to distinguish clear documents from doubtful ones: a digital invoice and a blurred receipt should be treated differently. Always create a separate queue for uncertain cases. Also decide who oversees it and how often.

The third mistake is changing rules without recording them. If a category is moved mid-period, the data of two periods is no longer comparable. The fourth is assuming the system knows your client's tax rules. It does not know them unless someone has provided them, and in any case tax decisions remain the professional's. Every change must be noted with the date and the person who approved it.

5. How to check the results

Verification should be done on documents the office has already posted, where the correct result is known. Take a sample of documents of varying quality, compare the extracted fields with the originals and classify the errors: field read wrongly, wrong category, document not recognised. Only after this analysis can you tell whether the problem lies in the input, the rules or the tool. Note the date of the trial, so you can compare results over time.

Ongoing checking is just as important. Even a system that worked well in a trial can deteriorate when suppliers, formats or staff habits change. Set a periodic sample check and a named person responsible for the result. A simple error log, updated every time, is worth more than a technical report nobody reads. The log also helps whoever has to explain an error after the fact.

6. Limits and compliance checks

Accounting entries and returns carry specific obligations that depend on legal form, tax regime and sector. There is no single answer for all firms or companies. Before using AI on documents that feed the accounts, clarify with your tax adviser which evidence must be retained and how the record of who confirmed each posting must be kept.

For documents containing personal data, data protection is the main issue. Verify the legal basis, the retention period and the instructions the tool provider receives, if it processes data on your behalf. If the system is used on clients' data within a firm, these points require a written agreement. Ask the data protection officer or a legal adviser to confirm.

7. How to start: a test on a single document type

This week, choose a single low-risk document type, for example invoices from recurring suppliers, and gather the posting rules into one sheet. Ask the head of administration to list the most frequent exceptions. No tool is needed for this step: you need the clarity that will later make the project controllable. Choose the document type also according to how many errors it currently produces.

Then run a sample test on documents already posted, using the error classification described above. The result will tell you whether it is worth continuing and which rules need better wording. Take the result to the person who signs off the postings, and decide together the scope of the next step. Share the result with the people who carried out the manual checks too.

Frequently Asked Questions

Can AI replace accounting staff?

As a rule, no, and it should not. Extraction and classification work can become faster, but verification, professional responsibility and the relationship with the client remain human. In many offices the change is in the type of work: less data entry, more checking and more attention to exceptions. Anyone reorganising the process should also consider training the people involved.

Can I upload a payslip into an AI tool?

Only after checking carefully where the data is stored, for how long, who can access it and whether the provider uses it for other purposes. Payslips contain particularly sensitive personal data and require stronger safeguards. In many cases the most prudent choice is not to use external tools for this document type without a written assessment from the data protection officer.

What do I do with a document the system cannot read well?

The safest rule is to put it in a separate queue and have it posted by hand. An uncertain document must not be forced into a category. Note the reason for the flag, to see whether the problem recurs with a particular supplier or sending channel. If it happens often, change the way the document reaches the office.

Do I also need to keep the original document?

Usually yes, and keeping it together with the posting is the tidiest practice. Retention rules depend on the document type and the applicable legislation, which should be checked with your adviser. The practical point is that every posting must be traceable back to the document it comes from, even when the system's summary is the only thing anyone reads.

How long does it take to see a result?

It depends on the volume of documents, the quality of the inputs and the time the owner can spend on checking. There is no point setting a deadline before defining the success criterion. A well-run trial on a single document type gives more useful guidance than a rushed full implementation. Ask the provider to describe the phases and what each one verifies.

About this content

General informational guide. Regulations and technical requirements change over time: always check the official sources and assess your specific case with a qualified technical or legal advisor.

GET IN TOUCH

Talk to our team.

We'll show how Altovar integrates into your infrastructure. No commitment, no sales pitch. Just a technical conversation.

BOOK A DEMO →