Most service teams that try AI get stuck in the same place. A pilot answers a few simple emails, and then nobody is sure how to let it handle real volume without risking the customer relationship. Beurer, a German family company with more than 500 health and wellbeing products, has moved past that point. At Customer Impact Summit 2026 in Cologne, Heike Hocks-Hamdorf, Director Customer Care at Beurer, and I explained how. Short version: test the AI hard before any customer sees it, widen its scope step by step, and treat it as a system that keeps learning rather than a project that ends.
The talk is in German. The write-up below is in English.
Why AI in Beurer's service had to earn trust first
Beurer's service team handles between 7,000 and 13,000 cases a month, depending on the season. Customers ask about more than 500 products, and some of them are medical devices such as blood pressure monitors and thermometers. That brings regulatory limits that apply to customer service as well.
Most customers still write by email. Heike put it plainly on stage: someone who buys a heated blanket usually doesn't want to chat with a bot. So the question was not whether to use AI, but how to use it on the channels customers already use, without losing the human side of the service and without lowering the quality customers expect from the brand.
How Beurer went live: four steps, each with a clear bar
Beurer did not switch the AI on and hope for the best. Each step had to pass before the next one started.
- Train and test on real emails. Heike and a colleague read through thousands of past emails. They used them to teach the AI what it may do and what it may not, and refined the rules until the AI handled cases the way their own team would. The bar: Beurer would not go live until the AI matched its team's handling in at least 95% of cases.
- Reviewed drafts. Once live, the AI only wrote drafts. Dedicated team members checked every single one, for content and for tone, asking themselves whether they would send it exactly like that in their own name. Beurer moved on when acceptance was above 90%.
- Autonomous in selected areas. The AI then answered emails on its own, but only in specific, chosen topics.
- Keep widening. The team keeps checking what the AI can already do well and extends its scope from there. This step has no end date.
The AI got a name along the way: KAI, short for Kundenservice + AI. Customers now write thank-you notes to KAI the way they would to a person on the team.
What happens when a customer writes in
Under the hood, Beurer's setup is modular. Every email or web form goes to an AI supervisor first. The supervisor reads the request and decides whether one of the AI specialists can handle it. If not, it hands the case to a person straight away. If yes, it activates that specialist, for example the one for heating products.
Each specialist has tools. Some look things up: product manuals, product data, or the customer's history in Salesforce. Others take action, such as updating records in Beurer's systems.
I walked through a real case. A customer had bought an HD 75 heated blanket and asked for a replacement switch because the old one was broken. The heating-products specialist:
- looked up the customer in Salesforce,
- checked the warranty policy,
- read the product data,
- updated the customer record in the CRM,
- and replied: this device has to be inspected as a whole, so Beurer doesn't send single replacement switches. To start a warranty check, please send these three pieces of information.
The customer had that reply within about a minute and could answer straight away.
Why the AI answered that way
The behavior comes from the specialist's instructions, written in plain language. In this case three of them did the work: always maintain the CRM record, ask for all missing information in one email, and clarify the product, the type of service and the customer's intent before suggesting anything.
These instructions grow over time. Heike said Beurer's first specialists had about half a page of instructions each. Within six months, the team noticed it kept writing the same instructions for different products. So it split them into reusable blocks. Today Beurer works with 13 blocks, for example:
- CRM check: who is this customer, have they bought from us or contacted us before, is this a follow-up?
- Follow-up questions: ask everything relevant at once, the way a person on the team would write an email. Heike admitted this one was learned the hard way: email is not chat, and customers don't want ping-pong.
- Knowledge: which sources the AI may use, and which source wins when they disagree. Beurer has FAQs, manuals, web pages and Salesforce Knowledge, and not all of them are always equally up to date.
With these blocks, a new specialist mostly needs its product-specific part. That makes building new specialists faster.
The knowledge behind the AI grew the same way. Beurer started at the end of 2025 with five specialists, a few self-written warranty documents and some manuals. The team soon saw that wasn't enough for real volume. Over the following six months it connected Salesforce Knowledge and, through an MCP server, its product information system (PIM).
The result: at the start of 2026, the AI fully resolved 28% of the cases routed to it. Today it is 49%. Heike credits the growth and sharpening of the specialists for the difference.
The Beurer customer story tracks the operation with other measures:
- 70% of all customer inquiries are routed to AI agents.
- Cases the AI handles on its own reach a resolution rate above 90%.
- For tickets that still need a person, handling time dropped by half, from 12 to 6 minutes, because the AI prepares the case first.
Instructions guide the AI. Hard gates stop it.
My main point was what we call authority: what the AI may decide, what it may say and what it may do.
Instructions work in layers. General guidelines apply everywhere, channel guidelines apply to one channel, and specialist instructions apply to one topic. But a language model interprets instructions, and you can't be fully sure how. It's like a team member: you can tell people never to get angry with a customer, and it can still happen.
That's why instructions alone aren't enough. Typewise adds hard gates that the platform enforces, whatever the AI decides:
- Draft mode on a channel: the AI can't send an email. A person has to press send.
- Approval on an action: an action such as issuing a refund can be set to require approval. The system then won't run it until a person approves.
Combining interpreted instructions with enforced gates lets a team set the AI's authority precisely. End-to-end tests then check that the whole setup still behaves as intended, which matters each time a specialist changes or a new model comes out.
How the system keeps getting better
Improvement doesn't stop at go-live. At Beurer, people keep reviewing the AI's work through ratings and audits. One useful view is the handoff reasons: why did the AI pass a case to a person, and was that intended? If not, was data inconsistent, was knowledge missing, or do the instructions need work? Beurer also flags conversations where a customer asks for a human, to spot acceptance problems early.
Doing all of this by hand takes a lot of time. On stage, I showed how we approach that at Typewise with Nova, an AI operator with access to the service data. A team can ask it, for example, why a certain type of request suddenly increased, and it can propose a change to a specialist and test it. People still decide which changes go live.
The lessons Beurer shared
The host asked us where things went wrong. Heike told one story. A customer's device was out of warranty, so the AI told the customer it couldn't be repaired and should be disposed of. The customer politely wrote back that they had thrown it away. In fact, a new cable would have fixed it. Beurer apologized, sent a voucher for its online shop and added a rule to its general guidelines: never give customers disposal advice. "We didn't think of it the first time," Heike said (translated from German). It's a good example of why the guidelines layer keeps growing.
My lesson was about preparation. At the start, everything sounds simple. Then a complex routing logic turns up mid-project, or an integration doesn't allow what the use case needs, and the work stalls. We now run a discovery phase first and writes a project charter: which use cases, which systems they need, and who in IT grants access and API keys. That surfaces blockers before the build starts.
We both ended on the same point. AI in service isn't a project you finish. Models change constantly, and new capabilities follow. Beurer's 13-block specialists would not have fit into a model's context a year ago. Teams have to keep growing with the system without losing sight of the bigger picture.
Talk to us
If you're working out how to take AI in your own service from pilot to real volume, we're happy to compare notes. Get in touch. You can also read the Beurer customer story.
