AI that takes load off accountants and bookkeepers

Yes. In practice we work with Twinfield, Exact Online, AFAS, Visma, and Yuki. Most packages have an open API or a supported connection, so we can plug in invoice recognition, automatic booking suggestions, and reporting export without staff leaving their work environment. For packages without an open API we look at an intermediate solution with scan-and-recognise or email-to-booking. We only build something once the connection runs stably and the bookings are correct, otherwise it produces net extra work instead of less.

You remain professionally and legally responsible, that does not change. AI is a tool, not an accountant. We therefore always build in a review step: AI drafts the suggestion, a staff member checks and authorises. For VAT returns, annual accounts, and tax positions, a human is always in the loop before anything goes out. What AI does do: work out the first 80 percent so the senior or compiler does not start from a blank screen. That saves time without changing the chain of responsibility.

Client data is sensitive: financial figures, personal data, sometimes strategic information. We work with EU-only AI providers where data is not used for training and not stored after processing. For firms with larger clients or strict DPA requirements we build a setup where data does not leave the firm network. We document per use case which data is processed, by which model, and for how long. That is also useful for your own GDPR accountability and for questions from clients who want to know what happens with their figures.

Yes, and this is one of the places where it makes an immediate difference. AI can search UBO checks, sanctions lists, and adverse media, draft a first risk assessment, and record the outcomes in a file. The compliance officer or partner reviews the assessment, but the gathering and summarising work is done. Important: the final risk assessment remains human work, especially in elevated or unclear risk cases. We tighten the audit trail here so you can demonstrate to BFT or the tax authority during supervision which steps were taken and who reviewed what.

OCR reads text from a receipt. That is where it stops. AI can suggest, based on previous bookings, which general ledger account an invoice belongs to, recognise whether a supplier is new, flag a deviation between the receipt and what normally comes from that supplier, and generate descriptions consistent with your categorisation. The difference is in the prediction of the booking, not just in reading text. For an SME portfolio with many small clients this is the difference between "scan and type" and "check and approve".

Honestly: a general language model does not know what changed yesterday in tax law or in an NBA guideline. We solve this by connecting AI tools to validated sources, for example an internal knowledge base with current NBA publications, tax handbooks, case law via rechtspraak.nl, and internal firm positions. The AI then cites only from those sources, with direct references. For topics where rules move quickly (think box 3, business succession, CSRD implementation) we advise always letting a human do a final check, even if the AI seems to give a correct answer.

We work in phases. First a short scan to see where within your firm the most time goes into manual work, often invoice processing, standard correspondence, or file review. Then a focused pilot on one process, with a team that really uses it. Only when the pilot works and staff value it do we scale further. You evaluate at each phase whether we continue. This prevents a tool being rolled out that no one uses, and lets the firm gradually adapt to new ways of working.