AI in tax firms: starting sensibly and keeping responsibility

The first sensible AI pilot doesn’t have to touch a tax return or a client engagement. Better suited are internal knowledge structures, preparing training, a template for recurring organisational processes or evaluating deliberately anonymised sample data. That way a team can learn to write good inputs, check results and recognise a tool’s limits.

Only once a use case is reliably understood is it worth asking whether, and under which conditions, client-related information may be processed. That is a governance and legal question, not a setting you sort out on the side in a chat window.

A convincingly worded AI output can be incomplete, wrong or unsuitable for the specific facts. With tax assessments, deadlines, calculations and client communication in particular, a clear checkpoint is therefore needed: who assesses the result, which sources or files are consulted and who approves the statement?

That is not just caution. It is what makes the benefit reliable in the first place: AI saves time in preparation, while the firm invests the time gained in assessment, quality and the client relationship. An automatic professional decision is not a good first use case.

German tax advisers are subject to a comprehensive duty of confidentiality. When external services are used, the service provider’s mere access to client-related facts can already be relevant under professional law. The German Tax Advisory Act only permits the use of service providers under specific conditions; selection, contract, confidentiality and the extent of access are not side issues.

Providers outside Germany or the EU and services that use further subcontractors deserve special attention. For services that directly serve a single client engagement, the law additionally requires the client’s consent. What is required in an individual use case should be clarified by the firm with its chamber of tax advisers, data protection advisers or legal counsel before real client data reaches an AI system.

A practical tool approval describes the specific use, not just a provider’s name. It records which accounts are used, which categories of data are permitted, where data is processed, who gets access, which contracts apply and how results are checked. With personal data, GDPR requirements are added, including technical and organisational safeguards and, where applicable, the rules on transfers to third countries.

A good process can look different: some firms work with a clearly defined, contract-reviewed service. Others start exclusively with data that has no connection to any client engagement. What matters is that the choice is documented comprehensibly and doesn’t rest solely on a tool’s marketing claims.

The European AI Act requires providers and deployers to take measures to ensure a sufficient level of AI literacy among the people who operate or use AI systems on their behalf. What is appropriate depends, among other things, on experience, training and the context of use. For a firm this means: not everyone needs the same training, but everyone must understand the approved tools, data rules, quality limits and escalation routes.

A short, documented start is more effective than a one-off product demo: explain approved use cases, discuss safe and unsafe examples, name a contact person and practise typical error routes. The rules are then developed further from real questions.

A pilot should deliver insight, not change the whole firm straight away. It ends with a decision: expand it cleanly, adapt the process or deliberately not continue.

As soon as a project touches real client data, a decision with significant effect, new service providers, processing abroad or automated external communication, a general guide is not enough. The specific design must be reviewed professionally. This article is guidance for planning and conversations, not legal or tax advice.

Good digitalisation strengthens the firm when it makes quality and accountability visible. That is exactly what distinguishes a sustainable use from a quick but risky tool experiment.