What can AI do in an accounting office, and what should it not do?
It can help with preparation and review: store a document, sort it, pull out data and flag what is missing. It should not post entries on its own, interpret tax rules or give professional opinions. A person checks every suggestion and decides.
This guide says nothing about rules in Bosnia and Herzegovina, including e-invoicing. What is allowed, and what you owe your clients, is a question for a specialist.
Which repetitive job should you automate first?
One, and one whose result is easy to check. Three come up most often: document intake, chasing missing documents and matching bank statements to documents.
- it repeats every week or more often
- it has a known input and a clear output
- there is a person who can tell whether the result is good
- a mistake can be spotted before the work moves on
Leave the rest for later, until the first job works.
How do documents arrive, and where do they get lost?
Usually through several channels: email, messaging apps, scanned paper, a shared folder. Before any automation, write down the path of a few real documents: where each arrives, who sees it first, where the original stays and who notices when something is missing.
The original always stays unchanged. Whatever the system extracts or suggests is stored separately, next to it.
What do rules check, what may AI suggest and what does a person decide?
Rules check what is simply true or false, such as whether there is an attachment or the file is a duplicate. AI may suggest what needs interpretation, such as the document type or an amount it has read. A person decides everything that moves on.
| Step | What a rule checks | What AI may suggest | What a person decides |
|---|---|---|---|
| Intake | Is there an attachment, is the format allowed, is the file a duplicate | Nothing; this step does not need AI | What to do with a message that has no attachment |
| Sorting | Is the sender a known client, is the date in the period | Document type and client, when a rule cannot tell | Confirms or corrects the sorting |
| Extracting data | Are required fields filled in, is the amount a valid number | Values read from a scan or photo | Compares with the original, accepts or corrects |
| Missing-documents list | What is expected but has not arrived | A draft message to the client | Edits the draft and sends it |
| Matching statements and documents | An exact match of amount and date | A possible match when it is not exact | Decides whether the match is right |
| Handing off | Is every item marked as reviewed by a person | Nothing | Posts in their own software; the system posts nothing |
If a rule can do the check, you do not need AI. Add it only where a rule cannot, and always with review.
What does Midaz know from its own work?
Midaz built an internal tool for its own use that compares bank transactions with invoices. An AI advisor in it only suggests a possible match, and a person decides.
It was built for our own work, not as a product for clients. We have not measured any time saving. Midaz does not replace an accountant or give accounting advice. We mention the tool only as an example of the division of work in the table.
What does a missing-documents list look like, and who follows it?
A list of what is expected but has not arrived, with the name of the person who follows it up. A rule can build it, but a person writes to the client.
- client and period
- what is missing
- how long it has been outstanding
- who is responsible
- what has already been done
Without a name, the list is one more list nobody reads.
What could one flow look like?
A fictional client sends an email with three attachments: a PDF invoice, a photo of a receipt and a scanned statement.
A rule saves all three attachments unchanged and records who sent them and when. It sees that the same PDF invoice arrived last week and marks it as a possible duplicate. AI suggests that the photo is a receipt, but the amount is not legible, so the item is marked “needs review”. A rule then sees that one expected document is missing, and the system drafts a message to the client.
The person at the office edits and sends the draft, reviews the flagged items and decides what moves on. Posting is done by that person, in their own software.
This is not a description of a deployed solution or of a real client.
What does a small pilot look like, and what happens when it fails?
A pilot is one limited process on a sample of your documents, not the whole office. The sample needs both clean cases and difficult ones: scanned, incomplete, duplicates and documents from the wrong client.
Measure the whole task, including preparation, waiting, review and corrections, not only the speed of the first result. If review takes longer than the old way, the pilot has not shown a benefit. Agree in advance where a failed run shows up, who gets the task and how the work continues by hand.
We agree a pilot as separate work, with a written quote: we walk through today’s process, agree the inputs, the human review and the success criterion, build a pilot of limited scope and review the result with your team. See also enquiry handling and where to start with AI automation.
Who may see your clients’ documents?
Only the people whose access was agreed in advance. Before a build we agree which data is needed, who has access and which outside services the solution uses, and we explain that before anything is connected. Business data stays yours.
The documents belong to your clients, so check your contracts before sending them to any outside service. That is a question for a lawyer.
What do we need to prepare a quote?
A description of one job and a few real examples. The first conversation is free, and you receive the scope, terms and price in writing before work starts.
- what the job is and who does it today
- how documents arrive and in which formats
- roughly how many arrive each month
- which tools you use and who has access to what
- what a good result looks like
- what must never be automated
- a few real examples, including bad ones
We promise no time savings or error counts in advance. The build, upkeep and the cost of outside tools are shown separately in the quote.
Frequently asked questions
Can AI post entries instead of an accountant?
No. Nothing is posted automatically. A rule checks, AI only suggests, and the person at the office posts in their own software.
Do I need special software?
We start from the tools you already use and check what their integrations support. We do not claim support for any specific accounting program.
How many documents does a pilot need?
We have no number that would hold for every office. What matters more is that the sample contains every kind of case. The size is agreed when the scope is set.
Who is responsible for a mistake?
The decision stays with the person who reviews, and we do not promise that mistakes will not happen. Who answers under contract and law is a question for your contracts and a specialist.
Does it work with documents in Bosnian?
We cannot promise that in advance. A pilot checks it on your sample: Bosnian text, scans, photographs and different formats. Midaz has not published a test on Bosnian documents.
How much does it cost?
The quote is based on the scope. You receive a written quote, and the first conversation is free.
Describe one repetitive job in five sentences: how a document arrives, who sees it first, what is checked, what is missing and who decides. That description is enough for a first conversation.
See also the guide to an AI chatbot for a business, the difference between an AI assistant and automation, AI and automation or contact.
How Midaz can help ↗