Guides

AI in customer service without sending your cases out of the house

A drafted reply saves two minutes per case – that adds up quickly. But the question to settle first is not "how good is the draft" but "who gets to read the case so the draft can exist". Because a ticket rarely contains nothing confidential.

Author: Dr.-Ing. Philipp Schwittek
Reading time: 7 minutes
Updated: 05. September 2026

The key points

Tickets contain names, plant designations, prices and faults – personal and business-critical data at once.
Using an external language model requires a legal basis and a data processing agreement; possible, but it is work.
A model on your own machine does not make the question smaller, but simpler: the data never leaves the house.
The unspectacular tasks pay off most – summarising, sorting, finding again. Not automatic replying.
No draft goes out unread. Anyone who cannot hold that line is saving at the wrong end.

What a ticket actually contains

An average support case contains a person's name and extension, the designation of a plant, often a fault description that allows conclusions about operations, sometimes prices or contract details. That is personal and business-critical at the same time. Before a language model works with it, it has to be clear where that text flows, how long it stays there and whether it is used for training. Those three questions are easy to ask and surprisingly often unanswered.

External with a contract, or in your own house

Both routes are viable. An external provider needs a data processing agreement, a checked legal basis, a statement excluding training use and an entry in the record of processing activities; with providers outside the EU the question of transfer is added. That is manageable, but it is a project. A model on your own hardware shifts the effort from contracts to technology: server capacity, operation, updates. We chose the second route because it answers the question "where does our ticket content go" in one sentence – it goes nowhere.

Where does the data flow, and to which country?
How long is it stored?
Is it used for training – contractually excluded?
Is there a processing agreement and a record entry?
What happens when the provider discontinues the model?

What pays off – and what does not

The greatest benefit lies in the unspectacular tasks. A summary of a long thread saves minutes at every handover. Automatic pre-sorting by topic removes manual distribution. Full-text search across attachments makes findable what arrives as a scanned image. Far less advisable is the fully automatic reply to the customer: it saves little, because it has to be checked anyway, and does great damage when it is wrong once. A draft released by a person is the better compromise – and stays so even as models improve.

Frequently asked questions

Frequently asked questions about AI in customer service

May customer data be put into a language model?

In principle yes, with the usual prerequisites: legal basis, data processing agreement, entry in the record of processing activities, contractual exclusion of training use. With providers outside the EU the transfer assessment is added. If the model runs in your own house, most of this falls away because the data never leaves the business.

Is an in-house model not considerably worse?

For support tasks: no. Summarising, sorting, building a draft reply from an existing thread – smaller models do that reliably. The gap to the large providers shows in free text production, not in these narrowly defined tasks.

Should replies go out automatically?

We advise against it. The time saved is small because it has to be read anyway, and the damage of one wrong automatic reply outweighs the gain from a hundred right ones. Draft yes, release by a person.

Matching services