"We replaced our customer support with AI."
At first that sounds like exactly the kind of statement you see all over LinkedIn and in the headlines right now. More automation, fewer people, lower costs.
But after we developed our own AI agent and tested it in our support, we came to a surprising realization:
👉 The real challenge in customer support was never the people. It was the information.
In this article we show how we automated our support process with AI, what experiences we gained, and why, in our view, most companies start with AI in the wrong place.
Our old support process
Many companies know this scenario:
Support requests come in through a wide variety of channels:
Email
WhatsApp
Phone
Microsoft Teams
Personal messages
Direct calls
It was the same for us. At first glance it worked surprisingly well. Customers could reach us at any time. At second glance, however, a structural problem became apparent.
The problem wasn't availability
The real challenge was that support didn't function as a shared process.
Knowledge was distributed:
Information was in private chats
Answers were in emails
Processes were stored in the heads of individual employees
Solutions often had to be explained multiple times
As a result, we spent a lot of time on:
🔎 Searching for information
🤨 Follow-up questions
↪️ Copy-and-paste work
💬 Internal coordination
And these activities do not create real value for the customer.
What does an ideal support process look like?
If we were to build a support process from scratch today, we wouldn't start with the question:
Which AI should we use?
Rather, with the far more important question:
How should the process work in the first place?
Step 1: A central entry point for all requests
The customer shouldn't have to wonder:
Should I write an email?
Should I use WhatsApp?
Should I call?
Instead, there needs to be a clearly defined entry point.
Step 2: Understand the request
Each request must first be categorized:
Standard question
Specialist question
Problem report
Individual consultation
Only then can it be decided how the request will be handled.
Step 3: Automate standard cases
Recurring questions should be answered immediately.
Examples:
How do I schedule an appointment?
Where can I find specific features?
How do I change settings?
This is where AI is particularly strong.
Step 4: Hand over complex cases to humans
As soon as a request becomes complex, an employee should take over. However, not without context.
A good handover process includes:
the original request
previous responses
the previous chat history
actions already taken
This enables the employee to help immediately, rather than having to research first.
Why not replace all support with AI?
After we defined this ideal process, a provocative question arose:
If standard questions can be automated, why don't we simply replace the entire support with AI?
So we started an experiment.
Our AI agent: More than a traditional chatbot
For the test, we deliberately did not want to use a simple FAQ bot. Instead, we developed our own AI agent. As the test environment we used our software Humafix, which was specifically developed for care providers.
Access to a knowledge base
The AI agent can do more than generate text. It has access to:
Documentation
Processes
Guides
internal knowledge bases
This enables it to answer questions much more precisely.
Access to tools
The key difference: our AI can not only respond; it can also act.
The agent is able to:
retrieve information from the system
read data
to check appointments
to execute actions directly
This turns a classic chatbot into a real AI agent.
The practical test: Does AI really work in customer service?
Once the agent was ready, we tested it with realistic support requests.
Test 1: Answering standard questions
Question: How can I schedule an appointment?
Result: The AI was able to answer the request immediately. No staff intervention was required. AI is excellent for such use cases.
Test 2: Retrieving information from the system
Question: When is employee Lena scheduled next week?
Result: Here the AI had to do more than retrieve knowledge. It had to read information directly from Humafix. That worked as well.
Test 3: Performing actions
Afterwards, the AI should not only provide information but also take action.
For example:
Check availability
Create appointments
Create entries
The agent was also able to successfully handle these tasks.
Test 4: Identifying complex support cases
However, the most interesting test was a different one. What happens when the AI isn't sure of the answer?
An example:
Since today, an employee's assignments are no longer being displayed correctly.
The AI analyzed the problem and tried to identify possible causes. When it couldn't find a confident solution, it didn't make one up. Instead, it forwarded the request to a staff member.
That was one of the most important success factors for us.
The key insight: AI must not become an obstacle
Many companies make the mistake of sending their customers through endless bot loops. We wanted to avoid that. If a customer needs a human, they must be able to reach one.
The best AI support is not the one that answers every request.
The best AI support recognizes when it should stop and hand the request over to a staff member.
Why information is more important than AI
Our most important insight is probably:
Our real problem wasn't support. It was the information.
When knowledge is distributed across:
WhatsApp
emails
phone calls
personal notes
individual employees
then even the best AI only helps to a limited extent. AI can only work with the information that is available and accessible.
That's why we centralized the entire process
The real breakthrough didn't come from the chatbot.
It came from the new process.
Automatic handoff to Microsoft Teams
If a request is escalated, our system automatically creates an entry in the Teams channel.
There the support team receives:
the original customer inquiry
the full chat history
all AI responses
actions taken
relevant context
This allows the responsible employee to take over immediately.
Our 3 most important lessons
1. Don't automate everything just because you can
Automation is not an end in itself. Just because something is technically possible doesn't mean it's worthwhile.
2. Automate routine work
AI should take routine tasks off employees' plates. This frees up time for more valuable activities.
3. Centralize information
Without a central knowledge base, any AI will sooner or later hit its limits. A well-structured process is more important than the newest AI tool.
What does this mean for companies?
When companies come to us today about AI in customer service, we rarely recommend a chatbot right away.
Instead, we start with three questions:
Where do inquiries enter the company?
Where is information lost?
Which tasks really need to be handled by people?
Only after that do we decide together:
which processes can be automated
which systems should be integrated
where AI actually provides added value
Sometimes the right solution is an AI agent. Sometimes a Teams integration. Sometimes a CRM. And sometimes a custom-developed software solution.
Conclusion
After our experiment we no longer believe that AI has to replace humans in support.
On the contrary.
The biggest benefits come when AI takes over routine work and people can focus on the tasks where they are truly needed.
AI should take over the work from people that they really shouldn’t be doing.
Anyone who wants to sustainably improve support processes should therefore not start with AI.
But with the process.
Would you like to find out which processes in your company can be automated?
Watch the full video for the article here!
#AI support#AI customer service#AI customer support#AI chatbot companies#AI agent customer service