Professional secrecy and AI: what auditors and fiduciaries need to know
Artificial intelligence saves a lot of time in auditing and fiduciary work. But auditors and fiduciaries are bound by professional secrecy: whoever hands engagement data to an AI tool potentially discloses secrets to third parties. This page explains what applies in Switzerland (Art. 321 SCC, Art. 730b CO, revised FADP), why the US CLOUD Act matters even with “Swiss” cloud services, which questions to ask every provider – and how pruvia processes vouchers without any original ever reaching the AI.
As at 13 November 2026
What professional secrecy means for auditors and fiduciaries
For auditors, professional secrecy is backed by criminal law: Art. 321 of the Swiss Criminal Code expressly covers auditors bound to confidentiality under the Code of Obligations; the statutory auditor is additionally subject to the duty of confidentiality under Art. 730b CO. Fiduciaries without an audit engagement are generally not subject to criminal law, but to contractual confidentiality and the revised Federal Act on Data Protection (FADP).
In practice this means engagement data may only be processed in a way that preserves the secret. Professional secrecy is the stricter standard, above data protection. A breach is not a mere compliance error; it can have criminal consequences.
The problem with AI tools and the cloud
Almost all common AI services process input in the cloud, mostly with one of the large US providers. If an audit or fiduciary tool uses AI, the decisive question is therefore not “cloud or not”, but where the engagement data goes, in what form, and who processes it.
Public chat AI is generally unsuitable for engagement data: input may be used for training and lies outside any data processing agreement. Data protected by secrecy needs a service with a data processing agreement, purpose limitation and a controlled location – and data minimisation: what the AI does not receive, it cannot disclose.
The US CLOUD Act: why a Swiss data centre alone is not enough
Many providers advertise that “data stays in Switzerland”. That only concerns the physical location, not legal access. The US CLOUD Act obliges US companies to hand over data on official order, regardless of where it is stored. Access attaches to the parent company, not the server location. This applies to all US hyperscalers, even with a Swiss region.
For data under secrecy this is central: the Data Protection Commissioner of the Canton of Zurich holds that data under professional or official secrecy may not be processed in Microsoft 365/Copilot because of the CLOUD Act, unless technical measures rule out unilateral access by the provider.
Seven questions for every AI solution handling engagement data
| Question | What matters | In pruvia |
|---|---|---|
| Where are the raw documents? | Originals (PDF, scans, photos) should not leave Switzerland. | Vouchers are kept in Swiss object storage (Exoscale, Zurich); the original never goes to the AI. |
| What does the AI receive? | As little as possible: text instead of a document, pseudonymised instead of plain. | Only pseudonymised text. Scans and images are first converted to text on our own infrastructure in Switzerland (text recognition, OCR). |
| Is there pseudonymisation? | Replace known names, companies, addresses and patterns (OASI, IBAN, e-mail, phone) before sending. | Yes, fail-closed: without successful pseudonymisation nothing is sent. If no text can be extracted from a voucher, the AI analysis is skipped. |
| Where does the AI compute? | Controlled region, endpoint in Switzerland or the EU. | AWS Bedrock, Zurich endpoint, model inference within the EU; other regions are technically refused. |
| No training? | Contractually excluded, no permanent storage at the AI service. | No use for training, no storage beyond the request. |
| Is there a data processing agreement? | Agreement with confidentiality, purpose limitation and listed sub-processors. | Data processing agreement with an annex on technical and organisational measures; text recognition runs in-house and is not an additional sub-processor. |
| Who decides? | AI results are suggestions; adopting them into the working papers is a deliberate step. | Every AI suggestion needs confirmation and sign-off; the function can be switched off per company. |
Vouchers without the original: how pruvia does it
Vouchers are the most delicate case: they contain names, addresses and amounts in plain text – and a scan cannot simply be “anonymised”. pruvia solves this in three steps, all before anything is sent to the AI:
- Use the text layer: if a PDF has a text layer (e-invoice, accounting export, digitally generated document), only that text is used, never the file.
- Recognise scans and photos locally: pure scans and images are converted to text by our own text recognition on pruvia servers in the Zurich data centre. The raw image stays in Switzerland; it goes to no third-party or US service.
- Pseudonymise, then send: the extracted text is pseudonymised like any text input and only then transmitted to the AI. The original remains unchanged in the file store.
The PDF import of trial balance, journal and bank statements is different: files with a text layer are read locally and deterministically; only a scan without a text layer is read there by the AI in the original – not pseudonymised, but exclusively within the EU (AWS Bedrock, Zurich endpoint) and without training. To avoid this, import PDFs with a text layer or Excel files.
During recognition the audit area shows the notice “pruvia is pseudonymising the documents”. Voucher checking then runs as usual: is the balance on the voucher, and if not, what is missing? Answers are suggestions; the auditor decides.
How to use AI in line with professional secrecy
AI is not prohibited – on the contrary. According to doctrine and case law (BGE 145 II 229), cloud providers count as auxiliaries; engaging them does not breach professional secrecy, provided the provider is chosen carefully, confidentiality is contractually secured and purpose limitation is observed. The cleanest basis is the client’s consent, for instance as a clause in the engagement letter.
Technically, the principle of data minimisation applies: pass on as little as possible. The most robust approach: keep raw documents in Switzerland and give the AI only pseudonymised text, processed in a controlled region, without training and without permanent storage at the AI service. That is exactly how pruvia is built.
Frequently asked questions
May auditors and fiduciaries use AI at all?
Yes. Engaging AI and cloud services as auxiliaries is permissible if confidentiality is contractually secured, the data remains purpose-bound and clients are informed.
Do vouchers go to the AI in pruvia?
Not in the original. Vouchers are converted to text – directly where a text layer exists, scans and photos via our own text recognition on infrastructure in Switzerland – and transmitted as pseudonymised text. If no text can be extracted, the AI analysis is skipped.
Is Microsoft Copilot subject to the US CLOUD Act?
Yes. Microsoft is a US group; the CLOUD Act also applies to Swiss data centres and “Microsoft Switzerland”.
Is a Swiss data centre enough?
Not on its own. The physical location does not protect against access via the US parent company. What matters is the provider’s jurisdiction, the form of the data (pseudonymised text rather than the original document) and complementary technical measures.
May I enter engagement documents into a public AI?
For data under secrecy, generally no: no data processing agreement, possible use for training.
How do I stay on the safe side?
Keep raw data in Switzerland, give the AI only pseudonymised text, conclude a data processing agreement, exclude training and obtain the clients’ consent.
Try pruvia
The system tour shows pruvia with sample data, without registration, including AI suggestions for an audit procedure. For questions about data protection and professional secrecy in your audit firm, arrange an online demo.
Further Knowledge pages
Audit software Switzerland
What audit software for Swiss audit firms must be able to do – point by point with the current state in pruvia.
Journal entry analysisJournal entry testing: 14 test patterns
Which patterns pruvia looks for in the journal, what they find and how findings feed into the risk assessment.
Type of auditLimited statutory examination
Workflow under the SER, negative assurance, audit programme with leadsheets and auditor’s report – and what pruvia takes care of.
Type of auditOrdinary audit and ICS
Risk assessment, ICS existence check, tests of controls and comprehensive report to the board of directors – how pruvia maps them.
CollaborationDocument list and client portal
Request documents and questions with a deadline, control reminders and file incoming items as evidence in the audit area.
Quality managementISQM-CH 1 in the audit firm
Eight components, evidence from the audit files, monitoring, annual evaluation, as well as persons, regulations and insurance.
TermsAudit, Revision, examination: terms
What audit, “Revision”, statutory auditor, audit tool and audit software mean in Switzerland and how they relate.
Change of softwareTransition year and prior years
Mark years audited in the legacy system, file old audit files and start the first year in pruvia with previous-year figures.