How to scale customer service with AI, without losing quality.
Seven out of ten service requests are routine. The other three need humans. How to cleanly separate the two, without losing staff or frustrating customers.
By Florian Wessling
A typical mid-market service day: three staff members answer seventy emails. About fifty of them are routine, delivery dates, invoice addresses, warranty scope, manual lookups. The other twenty are real problems that need humans.
The unhealthy ratio: two thirds of the time goes into routine, one third into actual value creation. This is exactly where the Service Machine starts.
What the solution does at its core
The idea is simple. We feed an AI with your own knowledge sources: FAQs, help articles, contracts, manuals, internal documentation.
When a customer asks a question, the system pulls the relevant document passages and sends them, together with the question, to the AI. The AI formulates the answer based on these sources, with source citations, so anyone can check where the information came from.
The key point: the AI does not invent anything. It works exclusively with your content. If the content is good, the answers are good. If the content is wrong, fix the content, you don’t have to retrain the AI.
Why this fits the mid-market
Three reasons why this kind of service solution is a better fit for mid-market setups than generic chatbots or full custom development.
You stay in control of the source. When you change a document, the AI’s answer changes too. No multi-week training runs, no black-box model trained on something you’ve never seen.
You can start small. A Service Machine works with a hundred documents just as well as with ten thousand. You start with your most-used FAQs and expand when you see the need.
Escalation to humans stays clean. As soon as the AI realises it doesn’t have a confident answer, it hands off to a human, with full conversation history. Nobody has to re-read the last ten minutes of bot dialogue.
What you should clarify before setup
Before we build a Service Machine for a mid-market client, we clarify four points with you.
What’s routine today? Three to five days of ticket transcripts, categorised. Without this baseline you’re building blind.
Where do your sources live? Confluence, SharePoint, Google Drive, a PDF folder on a network drive. We work with anything, but we want to know what shape the documents are in.
What is the AI allowed to decide on its own? For an industrial client the AI might not be allowed to extend warranties or grant discretionary refunds. For a software vendor it might be allowed to log feature requests in the backlog.
Who handles escalation? A specific person, a specific interface, a specific response-time expectation.
What runs after 90 days
From our pilot setups, as a realistic range:
- Seventy to eighty percent of routine requests are answered autonomously, with full source attribution.
- Response times for these routine cases drop from hours to seconds.
- Complex cases land with full context on your team’s desks. Handling time per complex ticket drops by about thirty percent because the briefing is already prepared.
- Your service staff spend more time on the third that actually matters.
What does not happen: layoffs. We say this explicitly in every service setup, the AI does not replace people. It takes over the routine so people can focus on the cases that need experience and judgement.
Where the typical traps are
Bad source data. If your FAQs haven’t been updated in three years, the AI will answer wrong too. Before every setup we run a source check with you and prioritise what should be updated first.
Escalation thresholds set too loose. If the AI tries to answer everything itself, it hands off nothing. We set thresholds conservatively. Better to escalate too early than too late.
No tonality guidance. A service AI that sounds like a generic help desk feels off-brand to your customer. We train the AI on your brand voice, with examples, corrections, and a member of our team as the responsible owner.
When the solution actually pays off
Rule of thumb: from about two hundred service tickets per month upwards. Below that, the cost of a RAG setup is often higher than the value, because the maintenance work doesn’t disappear, it just moves.
Above two hundred tickets, the case becomes interesting quickly. At a thousand tickets per month, the setup typically pays back within three to four months. With more, faster.
If you’re wondering whether this makes sense in your service setup, we’re reachable by phone (+49 40 46 89 67 68 0) or via the contact form. The first call has no sales section. We listen, frame the situation, tell you honestly whether it’s worth it.
