Preparing annual accounts with AI: compilation file finished sooner

October. The new year-end work season is getting under way: the planning for the spring is set, the first files have been created and the request to clients will go out soon. Anyone who gets the preparatory work in order now will notice the difference later when everything has to happen at once.

Preparing year-end work with AI does not start with the financial statements themselves, but with the part underneath. A compilation file consists for a good part of lookup and reconciliation work, and that is exactly the work an AI assistant with direct access to the accounts can set up for you. Which steps those are follows directly from what belongs in the file according to the standard.

What belongs in the file according to Standard 4410?

NBA Standard 4410 is clear about the file content for compilation engagements. It must include a reconciliation of the compiled financial statements to the underlying records, in practice: the reconciliation of the financial statements to the trial balance. Significant matters belong in it, along with how you dealt with them. Think of estimation items, accounting policy choices and information that turned out to be incomplete or incorrect. And the file concludes with the final financial statements for which management has acknowledged responsibility, plus the compilation report.

Around that sits the work that precedes the compilation itself. The NBA report on Standard 4410 assumes that a compilation engagement starts at minimum with a trial balance in which the basic records have been processed. Before that is in place, the request to the client runs: bank statements, specifications of debtors and creditors, data on fixed assets and loans.

Line those steps up and you notice how much of it is lookup work. Checking the reconciliation, going through intermediate accounts, explaining deviations compared with last year, finding out which documents are still missing. Not technically exciting, but time-consuming. That is where the gain lies.

Let AI prepare the reconciliation work

With AI Connect you link your own AI assistant, for example Claude or ChatGPT, directly to the client's accounts via MCP (Model Context Protocol). You ask your review questions in plain English and get an answer from the real figures, with the entries included.

For the file this looks as follows. You open the accounts and ask which intermediate accounts still have a balance. You have the query account shown, with description and amount. You ask which general ledger accounts deviate by more than 25 per cent from last year, so that your deviation analysis does not start with a blank page but with a list of notable items. The ageing analysis of debtors and creditors is there alongside in a single question.

Every answer stays verifiable: you drill down into the underlying entries until you trust it. That drilling down takes seconds, not a new export.

Firms with clients on multiple packages do not have to choose. The same approach exists for Exact Online, Twinfield and AFAS. You select the accounts per client and repeat the same checks; the AI only sees what the connected user is allowed to see.

Draw up the client request more precisely

For many firms the information request is the stickiest part of the season. Documents arrive in pieces, half complete, or only after a few reminders. Part of that you cannot solve with software, but a targeted question does help: clients deliver faster when the list is about their situation.

So let the AI first gather the open items: suspense entries, interim accounts with a balance, deviations that call for an explanation, old debtors you want to check are still collectible. Then ask for a draft questionnaire for the client based on those findings. This way you replace the standard PBC list (prepared by client, the list of documents the client supplies) with a request that is about this specific client.

The AI sends nothing itself. The questionnaire is a draft that you review, supplement and send to the client yourself. Client communication stays where it belongs.

Compilation remains your judgement

Standard 4410 puts the core of the engagement with you: reading the compiled information against your understanding of the entity, discussing significant judgements with management and requesting further information when something is off. You do not outsource that judgement to a language model. What you do outsource is the groundwork leading up to it:

• Reconciliations and specifications prepared before the file really opens
• Deviations against last year as a starting point for your analytical review
• A draft request per client, based on what is genuinely outstanding
• The same control questions repeatable across every client administration

Do document your AI use. In their AI guidance the NBA and NOREA ask firms to include requirements for AI use in the quality system, and to record in the file which prompts you used, how you verified the results and what your conclusion was. AI output on its own never counts as sufficient support; you verify and you conclude. You can read more about that boundary in what AI does not replace in your bookkeeping.

On security: AI Connect is ISO 27001 certified and independently pen-tested. With Essentials we store no financial data; every question goes straight to the accounting package. With Analytics the larger analyses run on a secured copy of the administration within the EU, which we remove as soon as the client stops. For year-end work you can also set the connector to read-only: the AI can then change nothing in the administration.

Example prompts for the first file

Copy these questions into your connected AI assistant as soon as the client's administration is open:

Which interim accounts still have a balance? Show the amount per account.

Which general ledger accounts deviate more than 25 percent from last year?

Are there entries on the suspense account? Show them with description and amount.

List the reconciliation checks for the file: VAT categories, interim accounts, debtor and creditor ageing.

Based on these open items, draw up a draft questionnaire for the client.

Want to try it yourself? Start for free with Exact AI Connect →