Helpdesk Scaling with AI Pre-Qualification at Reifen24

Starting point

Reifen24 is a brand of Goodwheel GmbH — online tyre retail with B2C and B2B business, classic seasonal trade with massive peaks in spring and autumn when the entire German market simultaneously switches to summer or winter tyres. In those weeks, ticket volume can spike by a factor of five to ten.

When we started at Reifen24, support ran through a shared inbox — a few people worked in parallel, the order was pragmatic (“whoever sees it first”), and at peak times response times sometimes ran into double-digit hours. Escalations were handled verbally, there were no statistics. The managing director knew this didn’t scale — but hiring new staff that’s only needed in two seasonal peaks per year is not economically sensible.

The question wasn’t “how do we grow the team?” but “how do we make the existing team handle the volume?” The answer is AI — but not as a replacement for humans, as a preparation layer.

Discovery & architecture

Discovery quickly showed that the bulk of tickets falls into four categories: delivery status (where is my order?), complaints (wrong tyre delivered / damaged), advice (which tyres fit my car?), and invoice/payment questions. Of these, the first three are standard requests with highly repeated answer structures in over 70 % of cases.

That’s the ideal pattern for AI pre-qualification: incoming ticket → classification → routing to the right person → AI suggests standard answer → human reviews + sends. The team gains time not because the AI answers the ticket, but because the ticket arrives pre-classified and a draft answer is ready.

We deliberately decided against full automation. In tyre retail, you’re liable for wrong advice — if the AI recommends a tyre that doesn’t fit, that’s a real problem. The human stays in the loop and therefore liable, but they don’t do the first share of the work themselves anymore.

Implementation

Phase 1 was the helpdesk build: defining categories, setting up routing rules, defining SLAs per category, building escalation logic. We trained the team intensively for two weeks — everyone now knows which tickets belong to their area and when to escalate. Statistics run automatically since: per category, per team member, per period.

Phase 2 was AI classification. We started with a model on EU data residency that was sharpened over the first two weeks with real tickets. After this tuning, classification accuracy was just above 90 % — good enough to go live.

Phase 3 was the answer suggestions. For each of the four main categories, we worked with the Reifen24 team on answer templates that cover typical standard answers — with clear placeholders for variables (order number, delivery date, tyre model). The AI fills these templates from ticket data, the human checks the variables and sends.

GDPR-wise, the setup is clean: personal data is pseudonymized before classification, the data processing agreement with the AI provider is part of the package, EU data residency is mandatory. Reifen24’s data protection officer signed off on the build before we went live.

Result

Today, Reifen24 scales support without growing the team. In the first seasonal peak after go-live, response times were about 60 % shorter than the previous year — at comparable ticket volume and unchanged staffing. Standard requests are answered pre-qualified, the team focuses on advice and complaint cases where human-to-human is genuinely needed. Statistics give the managing director clear monthly data — which categories are growing, where response times are tipping, where training is needed.

What mattered to us in this case: AI here is not a marketing term, it’s a concrete lever. It doesn’t do the job — it makes the job preparable. That’s the kind of automation that genuinely works in the mid-market.