CLIENT TRIAGE · USE CASE

AI triage for client queries: turning WhatsApp, email and phone into structured cases

AI classifies client queries, finds answer building blocks in the internal knowledge base, and prepares a draft for the case handler.

Researched & fact-checked by: DuneDive LLC · As of: 2026-05

What it is

An average Swiss fiduciary practice with 10-15 staff receives between 80 and 200 client queries per week. Channels are diverse: email to a generic info@ address, direct mails to the personally known case handler, phone calls (often missed because the case handler is in a meeting), client-portal messages, increasingly WhatsApp Business and Threema Work. Topics vary even more: "When will my tax return arrive?", "I received a letter from the Schwyz tax office - what to do?", "My employee is retiring - what do I have to report?", "Do I need a German VAT number for my EU webshop turnover?", "Can I still deduct the renovation of my flat 2026 in the 2025 return?".

AI triage for client queries means: every incoming message is read automatically, mapped to the right topic field and the right case handler, accompanied by a draft answer from the internal knowledge base, and placed in the employee's daily inbox for approval. The machine prepares, the human decides and sends.

Difference from payroll triage and VAT preparation: this pipeline is broader and less specialised. It covers the whole query spectrum of a fiduciary, not a single field.

Why it matters

Three points make client queries the biggest lever in a fiduciary office in 2026.

First: service expectations. Clients - shaped by online banking and platforms like Bexio - expect response times in hours, not two working days. Someone who sends a query on Friday afternoon and has no reply by Tuesday starts looking elsewhere. Treuhandsuisse noted in a 2024 member survey that 38% of mandate losses cited "poor responsiveness".

Second: staff shortage (see also payroll triage). A case handler who manually triages 30-40 queries daily before even starting actual work loses two hours per day to pure sorting and research. Those two hours are the direct ROI source of AI triage.

Third: professional secrecy. Client queries routinely contain particularly sensitive personal data (health, income, family). The AI pipeline must give that data the same protection as a locked filing cabinet. Concretely: no cloud model without a data-processing agreement and EU/CH hosting, pseudonymisation of client IDs before each model call, a local vector database for the client file, and an audit log capturing every model call and every human approval.

Professional duties follow from the codes of Treuhandsuisse and EXPERTsuisse - which do not directly apply Art. 321 SCC (that applies only to physicians, lawyers and a few others) but substantively require a comparable protection level.

How it works

The pipeline has five modules.

Module 1 - multi-channel ingestion: n8n receives emails via IMAP, portal messages via webhook, WhatsApp Business via the official Meta Cloud API, optionally Threema Work via the broadcast API, optionally phone voicemails via Whisper transcription. Each intake gets a case ID and is logged with timestamp and channel.

Module 2 - client identification: The sender email, phone number or portal login is used to find the client number in the CRM. If not uniquely identifiable, the query is marked "intake without client link" and routed to the firm's reception.

Module 3 - classification: A language model (Claude Haiku for standard, Claude Sonnet for hard cases) maps the message to one of about 25 internal topic fields: private tax return, corporate tax return, VAT, payroll, social insurance, day-to-day bookkeeping, annual close, inheritance, pension, real estate, ESG reporting, etc. Urgency is also classified (deadline in text, words like "urgent", "reminder", "blocked").

Module 4 - RAG lookup with two sources: First source is the client file (local Qdrant, confidential): recent correspondence, open cases, agreed fee frame, special agreements. Second source is the firm's practice library (local Qdrant, less sensitive): standard replies, internal guidelines, EXPERTsuisse publications, Treuhandsuisse practice guide, cantonal tax guidelines. Both deliver top passages feeding into the answer prompt.

Module 5 - draft generation and approval: The language model writes a Sie-form draft with citations and deadline notes. On uncertainty it returns: "case handler check required". The draft lands in the responsible employee's inbox (Outlook draft, portal draft, or a dedicated CRM triage inbox). Only after manual approval does the reply leave the office.

For recurring clients the system learns: after 50-100 queries the classifier knows the topic distribution per client and assigns more precisely.

Pipeline in 6 steps

01Multi-channel intake: email, portal, WhatsApp Business, Threema Work, voicemail transcripts land via n8n in a central triage inbox.
02Client match: CRM lookup by email, phone or portal login. Pseudonymisation of the client ID before any LLM call.
03Classification: language model maps to one of 25 topic fields, sets urgency, detects escalation cases (reminder, insolvency, deadline < 5 days).
04RAG lookup: two sources in parallel - client file (confidential, local) and practice library (guidelines, EXPERTsuisse publications).
05Draft generation: Sie-form proposal, with citations, deadline note, optional fee note (e.g. "this advice falls outside the flat fee").
06Approval: draft sits in the responsible employee's inbox. Only after manual approval is it sent. Audit log records model call, sources, draft, approver.

When to use

Useful from about 50 client queries per week, clearly economic from 100. Concretely: a St. Gallen fiduciary with 14 staff and 280 active mandates averaging 160 queries per week; a Geneva boutique with French- and German-speaking clients; a Zug practice with a high share of international clients (English communication, many treaty constellations).

Especially economic when the firm (a) runs a client portal and wants to grow usage, (b) has introduced or plans WhatsApp Business, (c) handles a high volume of "when will my tax assessment arrive" standard queries.

When not to use

Not for very small firms (< 30 queries per week). Setup overhead exceeds the gain.

Not without data-protection groundwork. Client queries routinely contain health data, family circumstances and financial position - all particularly sensitive personal data under Art. 5 lit. c revDSG. Preconditions: written DPA with the LLM provider, EU or CH hosting (Mistral La Plateforme EU, Anthropic via AWS eu-central-1 with zero retention, local Ollama), pseudonymisation of client IDs before model calls, documented data-protection impact assessment.

Not without a clear escalation protocol. If the AI handles a very urgent query (insolvency filing, tax-office reminder with seizure threat) on the standard triage path, time is lost. The classifier must define an "escalation track" pushing such cases to management within 15 minutes.

Not for queries containing legal advice in the narrow sense. Fiduciaries may not offer such advice under the Lawyers Act. The classifier must detect such queries and route them to an external lawyer - not answer substantively itself.

Trade-offs

STRENGTHS

Response time pressed from 2-3 working days to 4-8 hours - direct competitive advantage
Case handlers save 1.5-2 hours per day of sorting and research
Recurring standard queries ("where is my tax return") are 80% covered by drafts
Audit log meets revDSG documentation duty for automated processing

WEAKNESSES

Privacy groundwork (DPA, DPIA, pseudonymisation) is substantial - 2-4 weeks before first pipeline goes live
Classifier needs 4-6 weeks of training with case-handler feedback
Client expectations can rise ("I now expect a fast reply always") - the firm must communicate the SLA consciously
Escalation track must be cleanly defined, otherwise urgent cases drown in the triage flow

FAQ

How is professional secrecy concretely protected?

Three layers. Layer one: pseudonymisation. Before any LLM call, client name, address and all personally identifying data are replaced with placeholders ("client_4711", "address_4711"). The mapping stays local. Layer two: hosting. Models run in EU or CH (Mistral La Plateforme EU, Anthropic AWS eu-central-1 with zero retention, local Ollama). No US data transfer without a TIA. Layer three: audit log. Every model call is recorded with timestamp, caller, input hash and result hash. On request (FDPIC enquiry, client complaint), it is provable what was stored and what was deleted.

What happens with a WhatsApp query from an unknown client?

The client-match step fails (number not in CRM). The pipeline marks "new contact", skips the confidential RAG step, and either answers only generically or routes to reception, who calls back. Privacy principle: never reply with client data without knowing the recipient.

How many wrong classifications per day are normal?

In rollout 5-10%. After 4-6 weeks of training and feedback by case handlers, the error rate drops to 1-3%. Important: errors are collected in the audit log; a small feedback UI ("classification correct? wrong -> right category") lets case handlers tune the classifier. Every error is a training data point.

Sources

Treuhand|Suisse - Praxisleitfaden für Treuhandunternehmen · 2025-09
EXPERTsuisse - Fachpublikationen Treuhand und Wirtschaftsprüfung · 2025-12
EDÖB - Erläuterungen zur DSFA-Pflicht im KMU-Umfeld · 2025-06
Meta - WhatsApp Business Cloud API Dokumentation · 2026-04
Treuhandsuisse - Mitgliederbefragung 2024 zu Mandats-Bindung · 2024-10

FITS YOUR STACK?

What this looks like in your business – a 30-minute intro call.

Book a call