Fiduciary firms: AI sorting of invoices, expense reports and VAT anomalies
In brief. A fiduciary firm can reduce the time spent on routine documents by letting an AI model sort each supplier invoice and each expense report into two queues: documents that comply with the rules of the mandate, which receive a proposed entry ready to validate, and documents that show an anomaly (inconsistent VAT, missing mention, duplicate, unusual amount), which are sent back to an accountant with the reason. The model does not set the tax treatment; it applies the firm's rules and flags what deviates from them.
This article describes an illustrative scenario. It is not a named client and we give no quantified results.
The problem: time goes into uneventful documents
In an accounting mandate, some of the documents are ordinary: a subscription, a rent payment, an invoice from a regular supplier, a restaurant receipt. This share varies from one mandate to another and is measured during the pilot. Yet all documents require the same attention, because the anomaly can hide anywhere: a VAT rate that does not match the service, a missing VAT number, a foreign invoice without acquisition tax, an expense report submitted twice.
The accountant therefore checks everything, and it is volume that causes errors of inattention. Sorting aims to reverse the logic: the machine reads everything, the human looks at what deserves it.
The process before and after
Before. Documents arrive by email, by scan or through a portal. A staff member enters them or corrects an automatic recognition, chooses the account and the VAT code, then a manager reviews everything at the end of the period.
After.
- The document is read and its fields are extracted: supplier, date, amounts excluding tax and VAT, rate, VAT number, currency, description.
- Arithmetic and format checks are done by classic rules: adding up the lines, calculating VAT, number format, duplicate detection.
- The model handles what requires reading: nature of the service, consistency between the description and the rate applied, business nature of an expense, comparison with the habits of the mandate.
- If everything is compliant, the document goes to the "proposal" queue with an account, a VAT code and the degree of certainty of each field.
- Otherwise, it goes to the "anomaly" queue with the precise reason.
In both cases, nothing is booked without validation. The difference is that the "proposal" queue is validated in batches and the "anomaly" queue document by document.
The rules belong to the firm and to the mandate
| Check | "Proposal" queue | "Anomaly" queue |
|---|---|---|
| Supplier | Known, with a usual treatment | New, or bank details changed |
| Mentions | Mentions required by the firm present | Missing mention, VAT number absent or malformed |
| VAT | Rate consistent with the service, calculation correct | Unexpected rate, calculation difference, VAT charged by a non-registered party |
| Service from abroad | Treatment provided for by the mandate's rules | Acquisition tax to be examined |
| Amount | Within the supplier's usual range | Marked deviation, beyond the set threshold |
| Duplicate | None | Same number, or same amount and same date |
| Expense report | Receipt present, expense within the client's policy | Receipt missing, possible private expense, ceiling exceeded |
| Readability | Clear document | Illegible scan, partial document, unsupported language |
The rules are set per mandate: what is usual for one client is not for another. Thresholds, known suppliers and expense ceilings are parameters that the firm changes itself. The model settles no tax question: as soon as a treatment is open to debate, the document goes to an accountant.
Human in the loop
- Batch validation of the "proposal" queue by an accountant, who remains responsible for the entry.
- Sample checks, reinforced at the start and then adjusted according to the results.
- Document-by-document review of the "anomaly" queue.
- Memory of corrections: an accountant's correction becomes a rule of the mandate, visible and reversible, rather than hidden learning.
- A mandate owner designated for the rules and thresholds.
Audit trail and retention
Each document keeps its history: extracted fields, checks passed, proposal, person who validated or corrected, date. This is useful internally, and it fits within retention obligations.
Art. 958f of the Swiss Code of Obligations provides that books and accounting records are kept for ten years, on paper, in electronic form or in an equivalent form, as long as the link with the transactions is guaranteed and they remain readable. Art. 70 of the VAT Act (VATA) refers to the principles of commercial law for the keeping of books and requires documents to be retained until the absolute limitation of the tax claim. AI sorting changes nothing about these obligations: the original document remains the reference, and the model's proposal is only an aid.
Errors: which ones, and who bears them
- Undetected anomaly. A faulty document goes into the "proposal" queue. It can still be stopped at batch validation and at the end-of-period review. If it gets through, the consequence is that of an ordinary accounting error, and responsibility remains that of the firm towards its client. This is why human validation does not disappear.
- False anomaly. A correct document is sent back to the accountant. The cost is time. Too many false alerts end up no longer being read, however: this rate is monitored.
- Extraction error. An amount or a date misread. Arithmetic checks catch some of these; displaying the document next to the proposal lets you see the rest.
Confidentiality and data protection
A few reference points, to be validated according to your situation:
- Confidentiality. The fiduciary firm is bound to confidentiality towards its clients by its mandate. Auditors are also covered by art. 321 of the Swiss Criminal Code, which mentions auditors bound by professional secrecy under the Code of Obligations, as well as their auxiliaries.
- Personal data. Expense reports and payroll documents contain personal data, sometimes sensitive (a medical certificate attached to a report, for example). The Federal Act on Data Protection (FADP) applies.
- Subcontracting. Art. 9 FADP allows processing to be entrusted to a processor if a contract or the law provides for it, if only processing that the controller could carry out itself is performed, and if no legal or contractual duty of confidentiality prohibits it. The confidentiality clauses of your mandates should therefore be re-read before sending documents to an external service.
- Disclosure abroad. Art. 16 FADP makes it subject to conditions.
Why an open-source model hosted in Switzerland suits this case
Three reasons. First, volume: a fiduciary firm processes documents continuously, and a fixed cost quickly becomes more attractive than per-page billing. Second, confidentiality: a client's accounts say almost everything about them, and a model that runs on your infrastructure or in a data centre in Switzerland avoids sending these documents to a vendor. Third, the task: extracting fields and checking rules is within reach of medium-sized open-source models (Llama, Mistral, Qwen or Gemma families), to be verified on your own documents, in particular for handwritten or poorly scanned invoices.
If the firm already works entirely in Microsoft 365 and its mandates allow it, a solution in that environment may also be suitable. Our decision guide compares the two options.
How a pilot runs and is measured
- Test on a closed period. The model processes the documents of an already booked quarter for two or three mandates, and its proposals are compared with the entries made.
- Parallel operation on the current period.
- Going live on the pilot mandates, with reinforced sample checks.
| Indicator | What it measures |
|---|---|
| Share of documents in the "proposal" queue | The real potential |
| Proposals validated without correction | The quality of the account and VAT code proposed |
| Real anomalies among the documents flagged | The relevance of the alerts |
| Anomalies found at end-of-period review and not flagged | The residual risk |
| Processing time per document | The gain for the team |
| Corrections when checking the VAT return | The quality at the end of the chain |
Limits: when not to automate
- Tax questions that require analysis: mixed supplies, property transactions, complex international situations.
- Mandates too small or too irregular to show patterns.
- Poor-quality documents, as long as scanning is not improved at the source.
- Closing and year-end adjusting entries, which depend on the accountant's judgement.
Frequently asked questions
Does the model book entries on its own? No. It prepares a proposal. An accountant validates, in batches for compliant documents, one by one for anomalies.
Do we need to change accounting software? No. The sorting sits upstream and passes its proposals to the existing software, by import or through an interface.
What happens when VAT rules change? Rates and rules are parameters maintained by the firm, not knowledge frozen in the model. They are updated without retraining.
Do our clients have to be informed? This is to be assessed according to your mandates and your information duties. Hosting in Switzerland without disclosure to a third party simplifies the answer.
Where should we start? With two or three high-volume mandates with standardised documents, on a period that is already closed.
Going further
See also our Ark Fiduciaire client case, our solutions for invoice processing and accounting automation, as well as the same scheme applied to small insurance claims.
Our AI consulting starts with a discovery workshop. Contact us to talk about it.