AI-supported collections without damaging the client relationship

What is AI collections?

AI-supported collections combine the tiered-reminder logic from the ERP (Bexio, Abacus, Banana, Sage, SAP) with language-model text generation that adapts the reminder letter to the specific client and the specific client history. Instead of a uniform standard text, the customer receives a letter that takes account of payment history so far, typical response time, the ongoing relationship and any known special circumstances.

The fiduciary industry knows two tensions in collections. First: reminding too late means lost liquidity and in extreme cases bad debt. Second: too aggressive reminders mean client loss – especially painful in the SME segment where a fiduciary mandate often runs 10 to 30 years. The AI variant tries to address both risks at once: stay reliably on top, without losing the tone.

The use-case is explicitly NOT fully automatic. Before every reminder dispatch the fiduciary (or the responsible clerk in a delegated mandate) decides whether the letter goes out, needs adjustment, or is held back. For very small amounts (under CHF 200) and for long-standing clients an additional loop runs: a personal phone call precedes the reminder letter.

Why it matters

Three reasons. First: Swiss SMEs run an average payment-term overrun of 18 days (BAK Economics 2025 study). That is liquidity unnecessarily caught in the reminder cycle. Whoever reliably sends reminder tier 1 on day 31 – not day 45 – shortens the cash conversion cycle noticeably.

Second: fiduciary offices with 5 to 50 employees often have no dedicated collections staff. The accountant runs reminders alongside daily work – accordingly irregular. An AI-supported pipeline runs daily, proposes overdue cases every morning and only needs confirmation, not research.

Third: deferral-request triage. When a client writes after tier 2 "please grant a deferral due to liquidity bottleneck", the manual answer is a 15-minute job per case. The AI classifies the request (standard 30-day deferral, standard instalment plan, escalated liquidity emergency), proposes an answer and references the rules from the Treuhand|Suisse collections guide. The fiduciary reviews and sends – processing drops to 3 minutes.

Fourth, more psychological: the personalised reminder tone. A standard reminder feels anonymous and confrontational. A letter that contains "we know you have been our client since 2012" and "we understand the current financial year is putting pressure on you" empirically achieves higher payment rates – and lower client churn. This is not manipulation; it is relationship instead of template.

How it works

Six components work together.

ERP integration: The pipeline pulls open items via the ERP interface. Bexio delivers via REST (endpoint /2.0/invoice), Abacus via REST (from version 2023), Banana via XML export, SAP Business One via Service Layer, Sage 50 via ODBC. We recommend read-only access for extraction and a separate write-only access to push the reminder status back.

Tier logic: Per open receivable the due date is checked against the defined tiers (default: tier 1 after 30 days, tier 2 after 50 days, tier 3 after 80 days, debt-collection prep after 100 days). Configurable per client. Long-standing clients often get a 15-day grace before each tier.

RAG over client history: A vector database holds, per client, the relevant history: prior payment behaviour, past reminder disputes, correspondence, special agreements. For each reminder, the context is pulled – "this client typically pays 12 days after tier 1", "last deferral due to industry crisis 2023".

LLM letter generator: The language model writes the reminder letter with appropriate building blocks. Tier 1 is friendly-reminding, tier 2 factual-firm, tier 3 formally legal (with notice of further legal steps). The letter contains client-specific hints without entering private detail. We recommend Claude Sonnet or GPT-4.1 for DE/FR/IT, or Mistral Large under strict data protection.

Deferral-request classifier: Incoming emails or letters are parsed and classified (no deferral request, standard deferral, instalment plan, escalated emergency). A reply template per class. The fiduciary decides, sends, or escalates.

Human decision and audit trail: Every letter must be released by an authorised person before leaving the outbox. Every release is logged in the audit trail (model version, prompt, sent text, timestamp, person). Retention 10 years per Art. 957a CO.

NOT fully automatic: no letter goes out without human release. No debt-collection step is triggered automatically.

Collections workflow in 6 steps

1. 01Set up ERP integration: read-only pull of open items from Bexio/Abacus/Banana, write-only channel for reminder status.
2. 02Tier configuration per client: default 30/50/80 days, long-standing clients with 15-day grace, define collections threshold.
3. 03Index client history: load correspondence, special agreements, payment behaviour into vector database.
4. 04Run the daily pipeline: every morning the pipeline proposes the reminders due today, with draft letter, client context and recommendation.
5. 05Human release: the fiduciary or responsible clerk reviews, adjusts if needed, releases or holds.
6. 06Store the audit trail: model version, prompt, sent text, timestamp, releasing person – 10 years per Art. 957a CO.

When to use

The method suits fiduciary offices with a stable client base (50+ mandates) and regular reminder volume (at least 30 open items per month). Below that threshold the pipeline effort rarely pays off.

Particularly useful for mandates with high recurrence (monthly or quarterly billing) – for example fee mandates with a flat hourly rate, VAT flat preparation, payroll. Here the client history pays off quickly because the pipeline learns per-client patterns.

In a multilingual context (DE/FR/IT) AI collections is especially attractive. Instead of three template sets for three language regions the language model delivers the letter in the client language with the right politeness conventions – TGV-style for Romandie, more formal in Italian-Swiss, direct in German-Swiss.

For SME mandates that have outsourced their accounting to the fiduciary office, the pipeline can be applied twice: once for client-fee reminders, once for the client's own customer reminders. The latter is an additional service offering with clear value.

When not to use

DO NOT use for clients in acute financial distress (restructuring talks, imminent bankruptcy). These need personal contact and individual arrangement – an AI-generated tier reminder escalates the situation and damages the relationship.

DO NOT use without a human release loop. The temptation is strong to say "tier 1 is a standard text, that can go out automatically" – but even at tier 1 there are cases (recently deceased client, known illness, holidaying owner) where an automatic letter damages the relationship. Human release per letter is NOT negotiable.

DO NOT use when the language model has third-country hosting and the client history contains sensitive data (account balance, personal life situation, information relevant to professional secrecy). EU or Swiss hosting (Mistral, local Ollama) is required.

DO NOT use for debt-collection steps. The transition from collections to debt enforcement (debt-enforcement requests, court proceedings) is a legal decision and belongs in the hands of the fiduciary expert or counsel. The AI may compile the debt-collection prep dossier – the step itself must be triggered by a human.

DO NOT use without a data-protection notice. If client history sits in a vector database, that must be mentioned in the fiduciary office data-protection declaration. revDSG requires transparency about the processing.

Trade-offs

FAQ

Related topics

Sources

1. Treuhand|Suisse – Inkasso- und Mahn-Leitfaden für KMU-Treuhand · 2026-03
2. Bexio API V2 – REST-Dokumentation (Endpoints invoice, contact, payment) · 2026-04
3. Abacus AbaConnect REST – API-Referenz · 2026-02
4. Banana Buchhaltung – XML-Schnittstelle Doku · 2026-01
5. BAK Economics – Zahlungsmoral Schweizer KMU 2025 · 2025-11