Why most companies fail at automating with AI
Many companies only know the promises of AI
Most companies come to AI through talks, demos and posts on LinkedIn. There you see agents that are supposed to empty inboxes, write quotes and replace entire departments. Anyone who is not deep enough into the technology themselves can hardly judge which of this works in their own business. So AI becomes the answer to every question and the project starts with the tool instead of the process.
What gets forgotten is that a lot could be automated long before AI. Forms, webhooks, filter rules in the inbox and interfaces between programs have been around for years. They work reliably, cost almost nothing per run and always deliver the same result for the same input. If AI is used at these points anyway, the automation becomes more expensive and slower. It also becomes less reliable, because the AI does not treat the same input the same way every time.
A ticket does not need AI
A customer reports a problem through the form on the website. This should automatically create a ticket in the support system with name, customer number, affected product and description. The form already knows all of these fields. A webhook sends them to Zendesk, Jira or an in-house system and the ticket is there within a second.
At this point an AI would only read the same fields once more and pass them on. Every call uses up tokens and takes a few seconds. With a few hundred reports a month the costs hardly show, but the benefit is zero. On top of that comes the risk that the AI assigns a field wrongly or mixes up the customer number.
A rule recognises fixed patterns in emails
Many companies want incoming emails to be sorted automatically. Part of them follow fixed patterns that never change. The electricity provider's invoice always comes from the same address and always carries the same subject with a changing invoice number. The parcel service's shipping confirmation looks the same every time. Applications arrive through the careers portal with a fixed tag in the subject.
With emails like these, input and output are always the same. A rule checks sender and subject and files the email in the right folder or forwards it to accounting. Every inbox can do this out of the box and it needs no AI.
When the content matters, AI helps
It gets harder with emails whose content has to be read first. A customer writes that last week's delivery was incomplete. In the same paragraph he asks for a quote for a larger quantity and mentions his new billing address in passing. Sender and subject reveal none of this. A fixed rule would sort this email wrongly or not at all.
This is where AI helps, because it understands the context. So that it sorts the emails the way your team would, you store your company's categories in the system prompt. This includes a few typical examples for each category and the rules for special cases. One such rule could be that every complaint over 500 euros goes to the team lead. The AI then does not answer with free text but with a category from a fixed list. The next system can process this category without asking back.
Both approaches work best together. The rules first sort out everything that follows a fixed pattern. Only the remaining emails go to the AI. This lowers the costs and the AI only gets the cases where it really contributes something.
Customer support with AI responds to the customer
Customer support that responds to what the customer actually needs is a good use of AI. A template with fixed text blocks only fits as long as the request matches the expected case exactly. The AI, in contrast, reads what the customer really wants.
If a customer asks why her order has not arrived yet and whether she can move the delivery date to Monday, the AI recognises both requests. It fetches the shipping status from the inventory system and checks the knowledge base to see whether a later date is possible. From this it writes a reply that answers both questions. For simple requests the reply goes out directly. For complaints, cancellations or goodwill cases it lands as a draft with an employee who checks and approves it.
Where AI reaches its limits
An AI is only as good as the information it receives. If the knowledge base is outdated or access to the order data is missing, it writes friendly replies with wrong content. When in doubt it would rather invent a plausible answer than admit that it does not know something. That is why the system prompt also includes the instruction to hand unclear cases over to a human.
The costs grow with the volume. With a few hundred emails a month the tokens hardly matter. With several thousand a day it pays to check exactly which of them the AI needs to see at all. Data protection also needs attention, because customer data goes to the provider of the language model. A data processing agreement and a model hosted in the EU are therefore mandatory.
Break down the process first and then choose the tool
Before we build anything, we break the process down into individual steps together with your team. For each step we check whether input and output are always the same or whether someone has to understand the content. In the first case a rule is enough, in the second we use AI. Alexander Röse maps the processes with you and makes sure the result stays easy to use in everyday work. Jan Umbach is responsible for the architecture and for connecting the AI to your existing systems.
Do you want to know where AI really makes a difference in your business? Write to us. We go through your processes with you and also tell you where a simple rule is the better choice.
