AI agents in the finance department – what to automate in invoices, approvals and reports

An AI agent is a program based on artificial intelligence, similar to the one in ChatGPT or Copilot, that does not just answer questions but carries out steps in company systems by itself. It retrieves data, compares it, prepares a document and passes it to a person for a decision.

Most finance departments already use rule-based automation: the accounting system codes fixed costs to ledger accounts according to a template, and a Power Automate flow sends an invoice for approval. Rules work well as long as the situation is predictable. An agent is useful where content needs to be understood: the description of a service on an invoice, an email from a supplier or an employee’s question asked in their own words.

The table below helps you judge whether a given problem needs an agent at all. A simpler, cheaper tool is often enough.

KSeF, Poland’s National e-Invoicing System, is the state platform through which companies issue and receive invoices. Every company has had to receive invoices through KSeF since 1 February 2026, and since 1 April 2026 most sellers have issued them there. So invoices from Polish suppliers usually arrive as structured data. Retyping and reading scans are no longer the main problem.

The work has not disappeared; it has moved. Someone still has to assign the cost to an account and a department, check the invoice against the purchase order, find the approver and resolve differences with the supplier. These steps are described in the list below.

Take a hypothetical trading company with 90 employees. The agent sees a transport invoice with no order number, checks this carrier’s earlier invoices and proposes the logistics department and the right account. The accountant accepts or corrects the suggestion, and the agent records the correction for future suggestions.

The cheapest mistake in purchasing is the one that never gets past the request. An agent can check a purchase request before it goes for approval. Is there a budget? Does the company already have a contract with a supplier for similar goods? Does the request contain all the necessary details?

Suppose a manager submits a request to buy 15 laptops from a new supplier. The agent notices that the company has a framework agreement with another supplier for this equipment and adds that to the request. The manager and the approver see this information before they make a decision.

There is one condition: the agent must have access to budgets, contracts and order history in the ERP or purchasing system. Without that, all it can do is answer questions about procedures, which is useful but does not change the process.

At month-end close, an agent helps in two ways. It compares data from different sources, such as a bank statement with the ledger or intercompany balances, and flags items without a match. It also prepares the first draft of commentary on variances from budget.

Microsoft offers Finance in Microsoft 365 Copilot for this – an add-in for Excel and Outlook connected to the ERP system. It requires a Microsoft 365 Copilot licence, and in October 2025 data reconciliation was the only feature generally available. If your reports are built in Power BI, Copilot in Power BI will answer questions about the data in a report. However, it requires paid Fabric or Premium capacity, which is separately paid computing power for Power BI.

This is where the biggest risk lies. AI can write convincing commentary on a figure that was calculated wrongly. So before you bring an agent into reporting, write down how the company calculates margin, unit cost or departmental results, and have a person approve every figure in reports for the management board.

Every day the finance department answers the same questions: has an invoice been paid, how do I settle a business trip, up to what amount can a manager approve a purchase? An assistant in Microsoft Teams can answer them based on policies in SharePoint and statuses from the accounting system.

The assistant does not learn from company documents. It searches them for the answer at the moment a question is asked and shows which document it took the answer from. Microsoft states that company data accessed through Microsoft Copilot is not used to train models.

Check whether it pays off with a simple estimate. If the accounts team gets 60 such questions a week and each takes 5 minutes, that is 300 minutes, or 5 hours a week. The assistant will not handle every question, but at this scale it makes sense. With only a few questions a day, it is better to update the policy page on the intranet.

An agent in finance works with money and with data that the statutory auditor checks. So it must be subject to the same internal control rules as an employee. The most important of these is segregation of duties: neither the same person nor the same account may both prepare a payment and approve it.

Microsoft Copilot only shows data the user has permission to see. That is an advantage but also a risk: if the payroll folder is shared too widely, the assistant will find it faster than a person would. Review permissions in SharePoint and Teams before you launch the agent, not afterwards.

Since 2 February 2025, the AI Act (Article 4) has required companies that use AI to take steps that support their employees’ knowledge of the technology. In finance, this mainly means one skill: the accountant knows when an agent’s suggestion must not be accepted without checking.

Start with measurement, not with the tool. Without baseline data, you will not be able to prove to the management board that the agent has changed anything. Three figures from the last three months are enough, as described in the first step below.

The order matters. Purchase invoices involve a lot of repetition, clear rules and an effect that is easy to measure. Reports and analyses need well-organised data, so it is better to move on to them after some initial experience.

Not every company needs an agent. With a few dozen invoices a month from regular suppliers, rules in the accounting system and a simple approval path are usually enough. The cost of preparing and maintaining an agent would then exceed the time saved.

The project will fail if the company has no written approval rules, ERP data is incomplete or every department describes costs in its own way. The agent will then speed up the chaos instead of removing it. The process has to be put in order first.

Be cautious, too, about promises of rapid headcount reduction. The real effect is usually less work on documents and more time for analysis and exceptions. If the only goal is to cut staff within a few months, disappointment is very likely.