Karmameju is a Danish skincare brand with its own webshop in several markets and a customer service function that is a central part of the brand experience. Growth in the number of customers brought a matching growth in the number of enquiries.
Challenge
Karmameju aims to deliver customer service that matches the brand: personal, technically accurate and fast. That ambition comes under pressure when the number of enquiries grows faster than the team.
The pressure rarely hits quality straight away. It hits response time first, then thoroughness, and finally tone. At the same time, the writing itself is not the expensive part of a reply. Looking things up is: finding the order, checking the product’s properties, seeing what the customer has bought before, and judging whether something similar has been answered previously.
The question was therefore not whether to hire more people. It was how much of a reply could genuinely be prepared automatically, without the customer noticing any difference — other than the answer arriving faster.
Solution
The first step was an analysis, not a build. Every enquiry was pulled into one combined dataset and categorised with AI, to map what customers actually write about, how often each type recurs, and how much of the answer could be derived from data the company already held. That analysis became the requirement specification.
On that basis, a solution was built that prepares the reply before anyone opens it:
The enquiry is pulled automatically the moment it lands in the support system, together with the context that comes with the customer.
Previous replies are pulled in as reference, so the answer builds on the company’s own practice rather than on generic knowledge.
Product data is pulled in, so answers about use, ingredients and variants are factually correct.
Customer and loyalty club data is pulled in, so the answer takes membership, history and relationship into account.
Tone and style are written into the model, not something corrected afterwards. That is the difference between a draft that saves time and a draft that costs time.
A second leg was built alongside it: every enquiry is pulled back down together with the reply that was actually sent to the customer. The model therefore learns not from its own suggestions, but from what a person judged good enough to send. The distance between draft and sent reply is the real measure — and that distance keeps falling.
The final step is deliberately human. Customer service reads, judges and sends. Nothing goes out automatically. In theory that costs a little efficiency, but it ensures that no wrong promises are made, that responsibility sits in one place, and that the team experiences the tool as help rather than replacement.
Results
of all replies involve AI, fully or in part.
Quality has not dropped. On the contrary, the answers have become more consistent, because the context is the same every time.
- Noticeably less time spent looking things up, so the team spends its time on judgement and on the complex cases
- High satisfaction in the customer service team, which is decisive for whether a solution actually gets used
- A solution that improves by itself, because every sent reply feeds back as training data
- Customer service no longer scales linearly with customer growth