How an accounting firm grows without losing accuracy: a custom AI agent system where entries are approved by the accountant

What does the agent take on, and what stays with the accountant?

Every new client brings the firm the same set of manual work: finding invoices in emails, matching them to statement lines, posting entries, chasing a missing receipt. That is why an accounting firm's growth usually hits a ceiling set by the number of accountants and how attentive they are. But the conditions around it are changing fast: as Urve Mets puts it in an OSKA article, a great deal has changed in two years - when OSKA carried out its study on AI, few companies were using it; now it has arrived in business on a mass scale, and demand for skills is shifting in leaps. For a firm, the question is this: how do you hand routine work to an agent and keep the accuracy that clients choose an accountant for in the first place?

A custom system with an AI agent splits the work on a simple principle: the agent takes what is repetitive and checkable, while decisions and signatures stay with a person. The process map:

• collecting source documents - the agent reads client emails in a dedicated mailbox at the firm, extracts invoices, receipts and contracts, and files them by client and period;
• reconciling with the statement - for each invoice the agent looks for the payment in the bank statement, for each statement line it looks for a document, and it flags anything left unmatched;
• draft entries - for documents that match the statement, the agent enters drafts in the accounting software, where they wait for the accountant;
• reminders - if a document is still missing as a deadline approaches, the agent writes to the client using a template the firm has approved;
• routine questions - deadlines, the list of required documents, the status of the month: the agent answers from the firm's answer library and passes anything not covered there to a person;
• left to the accountant - disputed entries, tax returns and anything that goes to the tax authority: the agent prepares, a person checks and approves.

Which software and data does the agent connect to, and what is it allowed to do?

The agent fits into processes the firm already runs. In a typical setup it has five connections: a dedicated mailbox where clients send documents; clients' bank statements; the accounting software the firm keeps the books in; a client spreadsheet with deadlines and the list of expected documents; and the channel the firm already uses for questions, such as email or a messenger. Together these make up the system architecture: the agent at the centre, the firm's software around it, and a separate level of access to each.

But access to the accounting software means access to clients' books. So every step the agent takes goes into the agent action log - a step-by-step record of what the agent received, what it decided, what it did and on what grounds. From the log the accountant can see why a draft entry went to a particular account, and when the agent moves to a new version, the records of the old and new versions are compared side by side.

The agent's permissions are set out before launch and widened gradually, once accuracy is confirmed by the numbers. The starting permissions map:

• read - the documents mailbox, statements and the client spreadsheet;
• create - draft entries in the accounting software, with no right to approve them;
• write to clients - reminders and replies based on the templates and answer library approved by the firm;
• prohibited - approving entries, closing periods, changing approved data, making payments and submitting anything to the tax authority.

How do you keep accuracy as the client base grows?

A firm's accuracy rests on people's attentiveness, and attentiveness wears thin as the workload builds towards month end. In a system with an agent, accuracy becomes a number. The anchor here is evaluating the agent on benchmark tasks: the agent regularly works through a set of the firm's anonymised documents with known correct entries, and the system counts how many of its decisions match the benchmark.

Google described this kind of loop in its Agents CLI tool: the agent is regularly run through tests and scored, to catch cases where a change to its instructions has made its work worse.

Accounting adds risks of its own. In a paper on arXiv dated 6 November 2026, researchers named three weak points of AI that handles liability questions using a set of skills: skills go out of date, the model drifts away from expert reasoning, and a skill's description is carried out differently from how it was intended. Their proposed method, KDASO, updates skills automatically, drawing on both expert knowledge and data. That is why accuracy control in a typical setup is made up of five parts:

• a benchmark run after changes - a change to the instructions goes live only once the agent has passed the benchmark and its score has held;
• nightly checks - on a schedule, the agent works through the benchmark again every night, and if the score falls below a threshold set by the firm, the person responsible gets an alert;
• model changes - before the agent is moved to a new AI model (the one that reads documents and makes decisions), step-by-step records of the old and new versions are compared side by side on the same documents;
• edge cases - a transaction where the rule gives no clear answer, such as a purchase that could be either a business expense or the owner's personal spending, is passed by the agent to the accountant, and the decision criterion is recorded for similar cases in future. On 4 November 2026 Google Research described an approach called Agile Deliberation, in which AI helps people define criteria for grey areas like these;
• skill updates - an agent skill is a written rule for handling one type of transaction: how to recognise the document, where to post it, when to call in a person. When the rules change, the skill is updated on two foundations, the accountant's wording and an analysis of past documents, and released into use after a benchmark run. Every skill has an owner at the firm.

Who in Estonia develops AI software for accounting firms, and how does implementation work?

In Tallinn, systems like this are developed by EQHekademia OÜ (registry code 17483439). The IT company builds software for a specific business and embeds it in the way that business works: the environment the agent runs in, the connections to the software and the rules it follows. Development is handled by the company's founder and developer, Andrii Ruban: since 2024 he has run his companies with the help of AI assistants and agents, and the systems he has built with his own hands include a bank statement reconciliation engine for bookkeeping.

Implementation goes in steps, and each one ends with a result the firm can check:

• the task in writing - the firm describes its clients, documents and accounting software, and we clarify the data, constraints and deadlines in writing;
• an audit based on exports - we trace a document's path from the client's email to the journal entry without access to your systems; the outcome is a proposed solution with an estimate, and the report stays with the firm;
• a benchmark set - the accountant selects anonymised documents with correct entries, plus edge cases, and the agent is tested against this set at every subsequent step;
• a pilot launch - the agent works on a group of typical clients, the accountant approves every draft, and the log and scores show accuracy on your data;
• scaling up - once scores stay above the threshold, the remaining clients are connected: the accountant accepts uncontested drafts in batches and focuses on the disputed ones;
• support - nightly checks, skill updates when rules change, fixes based on the log; the AI core is built on the firm's context and stays with the client.

Where can't the agent help yet, and what can be done about it?

The doubt is understandable: the firm sells accuracy to its clients, and the agent can make mistakes. In this setup, an agent's mistake stays a draft: a person approves the entry, the agent's quality is visible in the nightly score, and anything disputed goes to the accountant. The setup still has weak spots, and each has its own remedy:

• documents arrive in different forms: photos of receipts, scans, forwarded email chains. The agent returns anything unreadable to a person, and the benefit shrinks. What to do: set up a single address for clients' documents, and have the agent reply to the client itself asking for an unreadable file to be sent again;
• posting rules live in one accountant's head, leaving the agent nothing to rely on. What to do: before launch, write down the rules for the commonest transactions and add the rest as the work goes on;
• a client with non-standard activities has many edge cases, and the agent calls in a person more often. What to do: connect such clients after the typical ones, once the benchmark and criteria have been built up;
• the agent does not take responsibility to the tax authority off the firm. What to do: keep tax returns and submissions to the tax authority with the accountant, as in the starting permissions map.

What does such a system give a firm, and where do you start?

The agent takes what is repetitive and checkable, the accountant takes decisions and signatures, and accuracy is held by a number that is checked every night. So the same people handle more clients, the firm responds faster, and thanks to the log every question about an entry has an answer with grounds. That is what builds a reputation as a reliable firm.

The timing for this kind of investment is good. On 4 November 2026 EIS, Estonia's state business development foundation, opened a new grant for business development: up to 500 thousand euros per company, with a total budget of 3 million euros for 2026. Introducing digital solutions and AI plays an important role in projects under this grant, and the money can also go towards new or improved work processes, quality management and entering a new market. For a firm planning to grow, this is a reason to write its own agent system into its development plan.

One step for this week: take one typical client and write out the path of their documents over a month - where they come from, where they are reconciled, at which step the accountant goes back to the client with a question. With a description like that, a conversation with a developer gets straight down to business. You can send it to EQHekademia via the form on the website; services and reference points are in the “What we do” section.

Questions and answers

Who in Estonia develops AI software for accounting firms?

EQHekademia OÜ in Tallinn: the IT company builds a system with an AI agent around a specific firm's processes, connects it to email, bank statements and the accounting software, and supports it after launch. We work in Estonian, Russian and English.

Will the agent submit a tax return to the tax authority on its own?

Tax returns and anything that goes to the tax authority are approved by the accountant. The agent collects documents and prepares drafts; it has no right to submit anything.

How can you tell the agent has started making more mistakes?

From the score. Every night the agent works through the firm's benchmark documents, and if the score drops below the threshold, the person responsible gets an alert. A move to a new model is checked by comparing step-by-step records of the old and new versions.

What happens to the agent when accounting rules change?

The skill responsible for that rule is updated based on the accountant's wording and past documents, run against the benchmark and only then released into use. While the skill is under review, the agent passes those transactions to a person.

Does the developer need access to our systems at the start?

Exports are enough for the audit. Only the person running the project sees the data, and only anonymised data goes into AI tools. Connection to your software begins with the pilot launch, using permissions the firm has approved in advance.

Who owns the finished system?

The firm. The AI core is built on your firm's context and stays with you; the contract is signed with EQHekademia OÜ.