AI for fiduciary and accounting firms – where it helps and where it does not

Where the time goes

In fiduciary and tax firms, work piles up before year-end closing and before tax return deadlines. Much of it is preparation: reading and assigning emails, filing attachments, typing up receipts, chasing missing documents and answering the same questions again and again. Two things help here. Automation means a workflow runs by itself according to fixed rules. AI comes in where content has to be read and classified that never looks quite the same, such as freely worded emails or invoices with different layouts.

Where AI and automation help today

Incoming client emails

A workflow reads new messages in the shared mailbox, recognises the type of request, for example receipts, queries or requests to extend a deadline, and assigns each message to the right client file and the responsible person. Attachments go into the client folder, and the sender receives an acknowledgement with a standard text. Anything the workflow cannot assign with confidence goes to a person.

Receipts and documents

AI extracts the supplier, date, amount, VAT and payment details from invoices and receipts and prepares them for your accounting software. Not everything needs AI: with a QR-bill, the payment information is already held in digital form in the Swiss QR Code and can be read directly. Common Swiss accounting packages offer interfaces or import formats; whether and how your software can be connected is something we check case by case. A booking proposal remains a proposal until a qualified person approves it.

Collecting documents and sending reminders

For year-end closing and tax returns, the system keeps a list of expected documents for each client and records what has already arrived. From this it prepares personalised reminders that a person only needs to check and approve. The status is visible at any time instead of being scattered across individual mailboxes.

Drafting standard replies

Many enquiries recur, for example which documents are needed for the tax return. AI drafts replies based on your own templates and saves them as drafts in the mailbox. Nothing is sent until a person has reviewed the draft.

Internal knowledge

A knowledge assistant answers staff questions from your own guidelines, checklists and templates and cites the source of each answer so it can be checked. It should only access approved documents and respect existing access rights. If you work with Microsoft 365, consider Microsoft Copilot for this too; see rolling out Microsoft Copilot.

For example: a fiduciary firm has n8n pre-sort the messages in its shared mailbox. When the workflow recognises a receipt, it files the attachment in the client folder and writes the extracted data to a list for review. A query produces a draft reply in Outlook. In the morning, the responsible person sees what is new, corrects where necessary and approves.

What AI does not do

• AI does not replace professional judgement: tax assessments, advice and responsibility for the accounts stay with your professionals.
• AI makes mistakes. It misreads amounts, assigns receipts wrongly or produces statements that sound plausible but are false, known as hallucinations.
• AI does not always behave the same way: the same input can produce different results. Where fixed rules are enough, they are the better choice.
• AI lacks the context your staff have: agreements made on the phone, the history of a client relationship, a family's circumstances.
• AI needs checks: plausibility checks, such as whether net amount, VAT and gross amount add up, plus spot checks and clear responsibilities.

Principles for using it

A person decides

A qualified person approves anything that goes to a client or into the books. This can be built in cleanly: in n8n, for example, a workflow can use the "Send and Wait for Response" operation to wait for approval via Outlook or a Teams chat before it continues.

Confidentiality and data protection

Client data is confidential, and tax documents often contain sensitive personal data within the meaning of the Federal Act on Data Protection, such as medical expenses or religious affiliation. The revised act (FADP) has been in force since 1 September 2023, and the FDPIC states that it applies directly to AI-supported data processing. We do not give legal advice, but you should clarify these points before you start:

• Processing by a service provider is only permitted if no statutory or contractual duty of confidentiality prohibits it (Art. 9 FADP).
• Does the provider bring in further subcontractors? That requires your prior authorisation (Art. 9 para. 3 FADP).
• Will personal data go abroad, and is protection there adequate or secured by contract (Art. 16 FADP)?
• Is a data protection impact assessment needed? The act requires one where there is a high risk, in particular for extensive processing of sensitive personal data (Art. 22 FADP).
• Do your privacy notice and terms of engagement cover the use of AI?

Client data does not belong in freely available chat services whose terms you have not checked.

Know where the data is processed

A workflow passes through several stations: mailbox, automation platform, AI model, file storage and accounting software. For each, you should know where it runs and who has access. An automation platform such as n8n can, for example, run in your own Azure subscription in the Switzerland North region. For the AI model, the processing location depends on the provider and the deployment type; see AI with your data in Switzerland.

Logging

Record what came in, what the AI proposed, who approved it and what was sent or booked. This lets you trace errors and answer queries. Set the retention period deliberately: by default, a self-hosted n8n instance deletes finished executions after 14 days at most, which may be too short for an audit trail. At the same time, a log should not hold more personal data than necessary.

Start small and measure

Choose one process, record the current state beforehand and compare afterwards. Useful measures are the time per case, the number of cases a person still had to handle and the corrections to AI proposals. Only with these figures can you decide whether expanding is worthwhile.

How to get started

1. Choose a process that is frequent, follows clear rules and has a clear result, for example assigning incoming client emails.
2. Record the current state for two weeks: volume, time spent and typical errors.
3. Clarify data protection and confidentiality: types of data, processing locations, contracts with service providers, confidentiality duties and the need for a data protection impact assessment.
4. Define the control points: where a person approves, and what the system must never do alone, such as sending to clients or booking.
5. Run a pilot with a limited scope, for example one team or some of your clients, and log every step.
6. Evaluate the pilot, then decide whether to expand, adjust or stop.

We support you from the first process to day-to-day operation – see AI & automation. An AI potential analysis shows where AI brings your firm the most benefit. Our article n8n or Power Automate explains which tool suits which workflow. To discuss your first process, reach us via our contact page.