For most tax advisers, artificial intelligence is no longer an experiment but a daily tool. That is now measured rather than assumed. The Dutch Association of Tax Advisers (NOB) and the Foundation for Tax Adviser Education (SOB) published a trend monitor on the impact of AI in tax services, conducted by research agency BusinessMonitor among 799 respondents: 700 AI users and 99 AI strategists.
The findings on AI for tax advisers leave little room for doubt. Around 80% of junior tax advisers use AI daily, roughly 70% of mediors and 53% of seniors. It is used mainly for text review and editing, summarising, searching legislation and case law, tax research and document analysis. Firm size barely affects how often it is used.
Yet a second figure sits alongside it that changes everything: 88% of respondents name trust in the quality of AI output as the single largest barrier. Asked specifically about AI-generated tax analyses, 39% rate reliability as low and a further 18% as very low.
Adoption and trust are not growing at the same pace. Adoption is far ahead.
The unavoidable question: who checks the work?
The most uncomfortable pattern in the research lies in two findings combined. The group using AI most intensively — juniors — is precisely the group least equipped to spot when it goes wrong. Respondents say so themselves: AI output is riskiest for less experienced advisers, because they cannot always recognise hallucinations, invented sources, incorrect conclusions and subtle errors of interpretation.
Dennis Jongbloed of Jongbloed Fiscaal Juristen puts it well: ask a complex question without the underlying tax expertise and you may receive a beautifully written answer that is nonetheless legally wrong. A misquoted provision, a ruling that does not exist, a conclusion that sits just beside reality.
That is not an argument against AI. It is an argument against AI that refuses to show its work.
Critically assessing AI output is named by 92% of respondents as the most important competence to develop, well ahead of prompt engineering (71%). The four-eyes principle remains the formal quality safeguard at most firms, and AI output is almost never treated as a finished product — more like the preparatory work of a junior colleague: usable, provided someone checks it.
The practical consequence is often overlooked. If every answer has to be verified anyway, then how fast a system answers is not its most important property. How fast you can verify the answer is. An answer with the citation attached, traceable to the statute or the ruling, takes an experienced professional minutes to check. A fluent paragraph without sources costs half an hour — or worse, slips through.
Why the gap between firms is widening
The trend monitor exposes something else. At large firms, around 70% of tax advisers use internally developed AI solutions. At mid-sized firms that figure is 18%. Mid-sized firms turn more often to specialised tax software from external vendors (39%), while smaller firms rely mostly on generic tools.
You can see that difference in the toolset. At mid-sized and large firms, Microsoft Copilot is the most used generic tool at roughly 73%, followed by ChatGPT (28%) and Claude (14%). At smaller firms, Copilot and ChatGPT are used almost equally (57% and 54%) and Claude appears relatively more often (30%).
Generic assistants are good at language. They are not built to answer a tax question from current legislation, parliamentary history and published case law with a citation at every step. Anyone using such a tool for professional work verifies not just the conclusion but the foundation — every single time.
The NOB therefore expects AI to widen the differences between firm segments: in productivity, innovative capacity and scalability. That is the real concern. Not that AI will replace the profession, but that only the largest firms can afford a system that meets the professional standard. A firm of ten should be able to get the same quality of support as a firm of three hundred, without building an AI department of its own.
From time saved to better judgement
The productivity gain is real: 45% of respondents save one to three hours per week, and 11% more than five. Even so, time saved is the least interesting part of this story. If research that used to take three hours can be prepared in one, the remaining two hours can go into thinking harder about the case. Which alternative structure is available? What happens on emigration or death? How is the tax authority likely to view this? Is a theoretically optimal structure also sensible in practice?
That matches what the NOB writes about the business model. As research, analysis and standard advice get faster, economic value shifts from production towards interpretation, judgement and strategic advice — strengthening the move towards fixed fees, subscriptions and value-based pricing. Clients do not enjoy paying for minutes spent at a screen. They pay for certainty about a tax position.
Four design requirements for responsible use
A few concrete requirements follow from the research for any firm serious about AI.
- Citations are not a nice-to-have. Without a traceable source, an answer is unusable for professional work, however well it reads.
- Human-in-the-loop belongs in the process, not in good intentions. The four-eyes principle applies especially when AI wrote the draft.
- Client data demands an explicit choice. Confidentiality and GDPR are named as a barrier by 42%. Where does the model run, who has access, what is retained?
- Policy has to be known, not merely to exist. Open answers show staff often do not know their own AI guidelines; at smaller firms, 38% report no AI policy at all. Regulation compounds this: 80 to 90% of tax advisers say they are not or barely familiar with the content of the European AI Act.
Verifiability as a design choice
PrudAI builds LEO as a digital colleague for legal and tax professionals. LEO works from more than 150 validated sources — legislation, case law and official publications — and returns the citation with every answer, so that checking takes minutes instead of half an hour. LEO makes no decisions and replaces no adviser; professional judgement stays where it belongs.
The tax practice edition of LEO emerged from co-creation with Dutch tax law firm Jongbloed Fiscaal Juristen, which uses the system daily for case law research, preparing and reviewing analyses and advice, and residency investigations. That co-creation is exactly why the product does what it does: tax professionals defined which questions need answering and when an answer is actually usable, and PrudAI built the system around that. Tested against real work by real tax professionals, not against a demo.
On information security, PrudAI is following the formal route: for ISO 27001:2022 and NEN 7510 the certification audit was completed in August 2026, with the certificate expected in the fourth quarter of 2026.
The question is no longer whether tax professionals will use AI. They already do, daily. The question is whether they use a system that lets itself be checked.
Would you like to know how LEO helps your tax or legal practice work faster with answers you can verify in minutes? Get in contact with us for a tailored demonstration.