AI in accounting: what it really does for firms today

What does AI concretely do in an accounting firm today?

Today AI mainly takes over the repetitive work in pre-accounting: reading documents, proposing bookings, linking payments and flagging missing information. These are tasks with high volume and lots of history, where patterns repeat: the same suppliers, the same kinds of costs, the same payments. A model can learn from that and make proposals an employee would otherwise make by hand.

What AI does not take over is professional judgement. Whether an item is an investment or an expense, whether a receivable is still collectable, or which tax choice works out best for a client remain the accountant's decisions. AI therefore shifts the work from data entry to review, and gives the team more time for review and advice. Concretely, the applications are:

• Recognising: reading supplier, date, invoice number, amounts, lines and VAT, including unfamiliar layouts.
• Coding: proposing GL account, VAT code and cost centre based on history.
• Matching: linking bank transactions to open items on amount, reference, IBAN, contact and date.
• Asking: flagging missing receipts or unknown payments and preparing a question for the client.
• Flagging: marking duplicate invoices, deviating VAT or unusual amounts.

How do confidence scores and booking proposals work?

A confidence score indicates how certain the AI is about a proposal and, together with a threshold, determines whether a booking is processed automatically or reviewed by a person first. An invoice from a regular supplier that is booked the same way every month gets a high score. A first invoice from an unknown supplier with mixed lines gets a lower score and ends up with an employee.

What matters is that a score is explainable. A good booking proposal shows what it is based on: similar earlier bookings, the VAT code, applied rules or instructions. That lets an employee judge quickly whether the proposal is right, and when something goes wrong the firm can see where the error came from. Corrections feed back into the system, so the next proposal for the same situation is better. A score without an explanation, by contrast, is hard to check and therefore hard to trust.

What is agentic AI and what can an AI agent like Atlas do?

Agentic AI is AI that does not just make proposals, but completes tasks on its own within boundaries a person has set. Where ordinary automation stops at recognising and proposing, an agent carries on until the task is done: booking an invoice, sending a question and processing the answer, or matching and reconciling a payment. Whatever falls outside the boundaries or needs human judgement, the agent escalates to the team.

Atlas is Autoboeker's AI agent. Atlas reviews and books invoices, sends client questions via the client portal and processes the answer into the booking, and matches and reconciles bank transactions, within the firm's confidence thresholds and rules. Every action by Atlas is recorded in the audit log, with the context and the rule it acted on, and the firm defines which actions always require a person. Atlas is available to firms with 25+ FTE as part of the Large firm plan; availability per package is on request.

Where do people stay in control?

People stay in control in three places: the threshold that determines what may be automated, approval steps before or after booking, and audit logs in which every action is traceable. With a threshold per firm or per administration you choose between suggestions only, smart confirmation or fully automatic booking. You can also specify that certain amounts, GL accounts or actions may never be automated.

Approval steps make sure a client or colleague sees a document before it is booked, or checks it after booking. Audit logs record who or what created, changed or approved a booking. That is needed not only for the firm's own quality control, but also to show a client or accountant how a figure came about when they ask. The layers of control:

• Thresholds: anything below the confidence threshold stays a proposal.
• Rules: specify which amounts, accounts or actions always require a person.
• Approval: approval before or after booking, in one or more steps.
• Audit logs: every action by people and AI is traceable.

What are the risks of AI in accounting?

The main risks are errors that go unnoticed, privacy and decisions that cannot be explained. An AI that proposes a wrong VAT code with high confidence can repeat the same error many times if nobody is watching. That is why you start with a cautious threshold, spot-check automatically booked items as well, and regularly review the corrections the team makes.

Privacy matters because invoices, bank data and client correspondence contain personal data. Under the GDPR you need a data processing agreement with every vendor that processes this data for you, and you want to know which subprocessors are involved and whether data is used to train AI models. At Autoboeker, data is not used to train public AI models. Explainability is the third risk: if it is impossible to find out why a booking was made, the firm cannot stand behind it either.

How do you introduce AI in a firm step by step?

AI is best introduced in phases: start with proposals the team reviews, increase automation only once the proposals are demonstrably right, and then extend to other processes. Choose a few administrations with lots of history and regular suppliers to start, because that is where proposals become accurate fastest. Involve the team from the start, so it is clear that the work changes from data entry to review.

Record working agreements that now live in people's heads as firm-wide rules or instructions, so the AI can apply them. Agree who reviews corrections and exceptions, and measure what changes: how many bookings go through automatically, how many client questions are open and where the team still spends time. Only once that picture is stable do you raise the threshold per administration. A step-by-step plan:

1. 1Choose two to five administrations with lots of history to start with.
2. 2Have the AI make proposals only at first, and review all of them.
3. 3Record working agreements as firm-wide rules or instructions, and exceptions per client.
4. 4Move to smart confirmation, so the team only reviews deviations.
5. 5Switch on fully automatic booking for administrations where proposals are consistently right.
6. 6Spot-check, follow the audit logs and extend to bank matching and client questions.