Customer service
Classifying enquiries, preparing replies, making knowledge available: where AI helps in service and where an automatic answer does harm.
What this is about
| Task | Value | Auto-sendable |
|---|---|---|
| Classifying and routing an enquiry | very high | yes |
| Preparing a draft reply | very high | no |
| Detecting sentiment and urgency | high | yes, as prioritisation |
| Providing shipment status | high | yes, if pulled from the system |
| Answering contract questions | medium | no |
| Handling a complaint | low | never automatically |
Introduction
- 01
Knowledge base before tool
A system can only tell people what exists in writing and is current. If your knowledge base is patchy, that is the first task.
- 02
Automate classification, prepare replies
Routing to the right group saves time immediately and carries no risk.
- 03
Enforce citation
Every draft names the knowledge-base passages it rests on. No source, no answer.
- 04
Define escalation paths
When does the system stop and hand to a person? Those rules are the real design.
The metrics that count
| Metric | Why |
|---|---|
| Reopen rate after an automatic reply | Measures whether it was actually resolved |
| Share of drafts sent unchanged | Measures the quality of preparation |
| Escalation rate | Too low means the system answers where it should not |
| Time to first response | The metric customers actually experience |
- Automatic replies only for cases with a checkable data basis.
- A route to a human at any time, visible rather than hidden.
- Disclosure that this is an AI system, at the start of the conversation.
- A log of the information given, for later queries.
The legal frame
| Point | Basis |
|---|---|
| Disclosure duty | Art. 50 AI Act |
| Legal basis for processing | Art. 6(1)(b) GDPR, contract performance |
| Recording conversations | Its own basis, consent or legitimate interest |
| Automated individual decision | Art. 22 GDPR, where legal or significant effect |
| Employee data | Performance monitoring via handling times engages co-determination |
Art. 22 GDPR is the point most often overlooked in service processes. An automatic rejection of a goodwill request can be an automated individual decision with significant effect. Rights then arise to human intervention, to express a point of view and to contest.
Grounding in the knowledge base
- Every answer names its source, and that source is checked automatically against the current version.
- Superseded versions are excluded from retrieval, not merely marked.
- Where nothing supports an answer, the system says so and hands over. The refusal rate is tracked as a metric.
- Permissions are filtered inside retrieval, not afterwards, so no information from other people's matters surfaces.
See RAG in depth.
What belongs in the log
Timestamp, channel, case reference, knowledge versions used, model version, whether sent automatically or by a person, and on escalation the reason. The customer's text in the clear belongs there only where a specific purpose and a retention period exist.
Choosing a task in customer service
FREE ACCOUNT
Selection sheet and templates: customer service
A sheet that picks the first task out of the ones customer service actually has, with two tested templates and the review steps for them.
Worksheet3 items