An accountant in an accounting firm answers dozens of questions a day that begin with "can we": can VAT be deducted on a lease of the client director's car, are business trip per diems exempt from personal income tax, how to account for gift vouchers for staff, what to do with an invoice from another EU country without VAT. The answer is sought in the text of the VAT and income tax legislation, in tax authority rulings and commentary, in colleagues' chat in Teams and in the firm's internal procedures in SharePoint, which describe how the firm treats such cases the same way for all clients. A junior accountant asks a senior one, the senior one looks for a ruling they read last year, and nobody is sure it is still valid after the latest amendment. Half an hour disappears on a single question, and the answers differ between clients.
The assistant answers from sources, not from general knowledge. The accountant asks in Teams in plain language, and the assistant searches the current versions of the VAT and income tax legislation, the tax authority rulings and commentary the firm has connected, and the internal procedures in SharePoint. It cites the article and section, the document title and the date with each statement; if the firm's procedure and the tax authority's commentary say different things, it shows both. If a ruling is older than a later amendment to the law, the assistant flags it and suggests checking. If there is no source, it says so instead of guessing. The assistant does not make tax decisions and does not draft the answer to the client: it shows where the rule is written, and the accountant decides how to apply it to a specific client. We deploy it with Claude through Amazon Bedrock in an EU region, because the questions include clients' situations: data is processed in the EU, is not used to train models, and every query and answer with its sources is logged in line with GDPR. The procedures stay in your SharePoint and are maintained by the chief accountant, not by us.