How to automate document management in a Swiss fiduciary firm

Every month, a mid-sized Swiss fiduciary firm receives hundreds of documents. Tax returns, bank statements, invoices, contracts, AVS forms, communications from the AFC. Each in a different format. Each one to be opened, read, classified, with the relevant data extracted and entered into the management system.

The team spends hours every week on this work. Not because it lacks skill — but because no one has yet built a system that does it automatically, respecting Swiss privacy constraints and integrating with the management systems already in use.

This article describes how a document intelligence pipeline for Swiss fiduciary firms works: what it automates, what it does not automate, how it stays compliant with the nLPD and how quickly it can be operational.

If you run a fiduciary firm and still spend hours processing documents manually, this article is written for you.

1. The real problem of document management in fiduciary firms

The problem is not the quantity of documents. It is that every document requires human attention for an operation that a machine can do in seconds.

Opening a PDF, identifying what type of document it is, extracting the relevant data — invoice number, amount, date, tax code, reference period — and entering it into the management system. Repeat for every document. Every day.

In fiduciary firms this work has specific characteristics that make it particularly costly:

Heterogeneity of formats. Clients send documents in any format: native PDFs, scans of variable quality, photographs of receipts, Excel files, digitized paper forms. There is no standard.

Seasonality of volumes. Peak periods — tax closings, AVS deadlines, VAT returns — concentrate high volumes into narrow time windows. The team gets overloaded exactly when it has the least time.

Risk of human error. A data point entered incorrectly into a tax management system is not just an operational problem — it can have legal and professional liability consequences.

Cost of double entry. Data extracted from documents often has to be entered into multiple systems: the main management system, the archiving system, possibly the client portal. Each step adds time and risk of error.

The sum of these factors produces a significant operating cost that recurs every month, invisible because it is spread across many hours of the team’s work.

2. What an AI pipeline can automate — and what it cannot

It is important to be precise on this point, because unrealistic expectations produce costly disappointments.

What the pipeline automates:

• Receiving documents from email, client portal or shared folder
• Automatic classification by type (invoice, tax return, account statement, contract, etc.)
• Extraction of the relevant data for each type of document
• Validation of extraction quality with a confidence level per field
• Entering the data into the management system via API
• Archiving the original document in the correct folder
• Notifying the team for documents that require human review

What the pipeline does not automate:

• The professional assessment of the extracted data — that remains with the accountant
• Tax decisions and interpretations of regulations
• Documents with a scan quality too low for reliable extraction
• Anomalous cases or situations that require context not present in the document

The principle is simple: the pipeline handles the mechanical work, the team handles the professional work. This is not about replacing staff — it is about freeing them from work that does not require their expertise.

3. How it works technically — without the jargon

From the point of view of the firm’s team, the flow is this:

The client sends the document. By email, through a web portal or by uploading it to a shared folder. They do not have to do anything different from what they already do.

The system receives it and processes it. Within a few seconds the document is opened, the system identifies what type it is and extracts the relevant data.

The data enters the management system. Automatically, via API. Without anyone having to open the management system and enter the data by hand.

The document is archived. In the correct folder, with a structured name, accessible to those who need it.

The team is notified only when needed. If the confidence level of the extraction is below threshold — poor quality document, unrecognized type, anomalous data — the document is flagged for human review.

4. nLPD and Swiss data residency

This is the point on which Swiss fiduciary firms have every right to be demanding.

The new Federal Act on Data Protection (nLPD), in force since September 2023, imposes precise obligations on the processing of personal data. For a fiduciary firm that processes its clients’ tax and financial data, the requirements are stringent.

A document intelligence pipeline that complies with the nLPD must meet these criteria:

Swiss data residency. The documents and the extracted data are processed and archived on infrastructure with servers physically located in Switzerland. The data never crosses borders towards foreign servers during processing.

No use for training. Client documents are never used to train third-party artificial intelligence models.

Data minimization. The system extracts only the data strictly necessary for the operation.

Full traceability. Every operation is logged with a timestamp, the type of document, the extracted data and the actions taken.

Atenek implements these guarantees as standard for all Swiss clients — not as a premium option.

5. How long it takes to go into production

The honest answer is: 2-4 weeks from the initial analysis session, for a basic pipeline on standard document types.

The time depends on three variables:

Variety of document types. If the firm mainly receives standard invoices and tax returns, configuration is quick. If the documents are very heterogeneous or in different languages, more calibration time is needed.

Number of systems to integrate. An integration with a single management system is faster than one with three different systems.

Availability of sample documents. Before go-live, the pipeline is tested on the firm’s real documents.

No structural IT intervention is required from the firm. The pipeline connects to existing systems via API — no installation, no migration.

6. What to expect in the first month

The first month is used to calibrate the system on the firm’s real documents.

In the first two weeks the team monitors the pipeline’s output, flags any classification or extraction errors and provides feedback. The system is updated on the basis of this feedback.

From the third week onwards, for most firms, the pipeline reaches a stable level of accuracy that allows monitoring to be significantly reduced.

What to measure in the first month:

• Percentage of documents processed automatically without human intervention
• Extraction accuracy per document type
• Average processing time per document
• Number of documents flagged for human review

FAQ

Does the pipeline work with documents in German, French and Italian? Yes. The document intelligence engines support the three Swiss national languages.

What happens if a document is too illegible to be processed? The system flags the document for human review with an indication of the problem detected.

Is the firm’s client data used to train AI models? No. Never. The processed documents remain operational data.

Does the pipeline replace the management software already in use? No. It integrates with the existing management system via API.

How much does it cost? The cost depends on the volume of monthly documents and the number of systems to integrate. The first analysis session includes an estimate of the expected ROI.