Case study · Fiduciary & administrative services · French-speaking Switzerland
A real engagement, client anonymised. Figures measured with the client and rounded; not audited by a third party.
A 35-person fiduciary firm now invoices in 3 days instead of 12
An AI assistant connected to the accounting software, mailbox and documents prepares entries, invoices and reminders; a person approves every outgoing action. Results measured over 90 days.
Scope and limitations
Rounded figures, measured on the trial scope and then the extended scope. A real engagement: the client’s name and any detail that would identify it have been removed.
- Company
- Fiduciary and administrative services, around 35 people, 3 legal entities
- Scope
- Around 450 active client files: documents, invoicing, reminders and approvals
- Software in place
- Swiss accounting software, Microsoft 365, partial document filing, email used as the to-do list
- Starting point
- No AI tool in use; individual use of ChatGPT with no rules
Starting situation
The back office was absorbing the growth: every new client file meant more data entry, filing, reminders and cross-checks between the three entities. Process knowledge sat with two people, and invoices regularly went out late.
Management had received several vendor offers promising turnkey “AI automation”. None answered the practical questions: where does our clients’ data go, who approves an accounting entry before it is posted, what happens when the AI gets it wrong?
The question asked
Buy yet another piece of software, or add an AI assistant on top of the software already in place, with a person approving every outgoing action?
Constraints
Sensitive fiduciary data
Financial and personal data of around 450 clients, covered by the Swiss data protection act and professional secrecy. No data could be used to train a vendor’s models.
No in-house IT person
The system had to run without anyone technical on site: simple handling, clear alerts, a written procedure for outages.
An SME budget
An investment the partners could defend, committed in stages, each one approved on measured results.
French and German
Documents and correspondence in both languages, often mixed within a single file.
What we put in place
We replaced no software. The assistant reads incoming email, documents and accounting entries, prepares actions and acts only within written limits: role-based access rights, human approval before anything is sent, a log of every operation.
The building blocks
Stages of the engagement
Analysis and starting measurement 3 weeks
Mapping of the real processes, measurement of handling times and invoicing delays, sorting of sensitive data, choice of the first scope.
Set-up 6 weeks
Access rights, read-only connections, a test environment with anonymised files, operations log.
Supervised trial 8 weeks
The assistant works for one of the three entities; every action is approved by a person; weekly error review and threshold tuning.
Extension 6 weeks
Roll-out to all three entities, sample checks for low-risk actions, team training.
Security review and handover 3 weeks
Operating procedures, alerts, security review; management takes over the monthly check-in.
- Nothing sent without human approval during the first 3 months
- Data kept separate by entity and by client file
- Models used with no retention and no training on client data
- When in doubt, the assistant holds back and passes the file to a person
- Tamper-proof operations log, reviewed with the external audit firm
- Monthly check-in with management: errors, costs, decisions to extend
Measured results
The first three weeks were used to measure the starting situation. The figures below compare the 90 days before the trial with the last 90 days of the engagement, on the same scope.
Rounded figures, measured on the trial scope and then the extended scope. A real engagement: the client’s name and any detail that would identify it have been removed.
What we would do differently
Every study also says what could have been done better.
- 01
Measure from the first week. We consolidated the starting measurement during the analysis; six weeks of finer data would have made the final comparison stronger.
- 02
Bring in the external auditor at set-up. Their review of the log during the trial validated the approach; earlier, it would have saved two rounds of rework on the log format.
- 03
Build fewer connections at the start. Four were delivered; two were enough for the trial. The other two should have waited for proof of use.