AI in a fiduciary office: where it helps and where it does not

If someone retypes figures from receipts or follows up on missing documents, start with that workflow (A workflow specifies what information goes in, what happens to it and who checks the result.). AI can extract details and draft requests. Your staff check the originals before posting an entry or sending a message.

We are not fiduciaries and do not advise on accounting or tax. We examine workflows and set up AI for the people who use it. The examples in this article are invented.

Work AI can prepare

Choose a task where someone can check the result against the original document. These steps are worth examining:

• Recording receipts: AI extracts the date, amount and supplier, plus the tax rate if shown. It suggests an account and flags missing or illegible details for review.
• Drafting a document request: AI compares what has arrived for a tax return with last year’s file and drafts a request about documents that may be missing.
• Matching bank statements: AI matches payments on a bank statement to open invoices and flags discrepancies for a person to review.
• Filing: AI gives incoming files consistent names and suggests the appropriate client file.
• Tracking open questions: AI uses emails and notes to list questions still open on a client’s file.

Our examples of first projects use supplier invoices to show what AI can prepare and what a person checks and approves.

What your staff still decide

An AI suggestion is not a professional assessment. Your staff remain responsible for:

• Judging: Your professionals decide how to post an unusual item and apply a new rule. They also assess whether a deduction is allowed.
• Advising: Your professionals take account of the client’s circumstances and what they have agreed with them. An AI draft cannot replace that knowledge.
• Approving: A person has the final say before anything we build sends a message outside your firm or spends money. Decide who approves other actions too, such as posting an entry.
• Being accountable: Your office remains responsible to clients and authorities, even when AI prepares the draft.

AI can misread a figure or add something the document does not say. Have a person compare its suggestions with the source. A first task should be easy to check: you can verify an amount against a receipt. You cannot verify a tax judgement that way.

Three questions about your clients’ data

Salary certificates, bank statements and contracts contain client data. Before asking AI to process them, establish:

• Where are the working files? In an installed vault (The vault is a folder of your working files on your computer; we create it from the vault template.), your working files are on your machine. That alone tells you nothing about which content is sent to an AI model (A language model is the actual AI behind an AI assistant: it has learned from a great deal of text and writes an answer to an input.).
• Who receives the inputs? A cloud provider (The cloud means computers in a provider’s data centres on which programs run and data is kept, instead of on your own device.) receives the content sent to its model. If a model runs on your own hardware, inputs need not go to a model provider (An AI provider runs language models and sells access to them, for example OpenAI (ChatGPT), Anthropic (Claude) or Google (Gemini).). In a hybrid setup (Locally, the language model runs on your hardware; in a hybrid setup the working files stay local and only the excerpts needed go to a cloud model.), files can stay local while selected excerpts go to a cloud model.
• Are inputs used for training? That depends on the provider, the type of account and the applicable terms. Check the terms for the account and model you intend to use. Do not assume a personal account has the same terms as paid API (An API is an interface through which one program uses the data or functions of another, without a screen or a mouse.) access.

We do not require a particular model or provider. Before setup, we establish with you what content may be processed and which terms apply to your chosen account. Read our FAQ answer on training with inputs and the options for processing data on your machine or in the cloud.

The Swiss Federal Act on Data Protection also applies when you use AI to process personal data. You or your legal adviser must assess whether your client agreements permit the planned processing. We can show you what the installed system sends to each provider.

Before work begins, we sign a data processing agreement (A data processing agreement (DPA) sets out how a service provider handles personal data it receives on your behalf.) with you for the processing of personal data. We can also sign a non-disclosure agreement if you wish. During setup, AI agents (An AI agent carries out several steps of a defined task. A person decides before it sends anything outside your firm or spends money.) directed by Jules may process content from your work. We record in writing beforehand what they may send to each provider. At handover, we end our access. Please then change any credentials we knew.

Choose a small first task

Choose a recurring task you can check against the original document, such as entering one client’s receipts. First check what your accounting software can already do. ‘The cheapest solution might already be installed’ shows you how to examine those features.

‘Where does the time go?’ includes a worksheet for observing a workflow over a week. Its fiduciary example is invented too.

From the task to the interview

The free interview asks about your work, one question at a time, with no account; describe it without entering client data. After level two, your report shows where it pays to start, and each further level makes it more precise. You book a free 20-minute call with Jules on the contact page. After that, you can decide whether to go ahead with a package.

‘What does AI automation cost an SME?’ uses a worked example to distinguish the one-off setup price from possible running costs.