AI in Customer Service: Response Times Reduced by More Than 50%

Customer service is a critical area for any company with a large customer base: fast, accurate, and customer-friendly communication has a direct impact on satisfaction, brand perception, and operational efficiency. In recent years, the explosive growth of artificial intelligence (AI) has created new opportunities for optimizing customer service processes.

In which sectors is this relevant? Anywhere where customer service activities are carried out. Typical examples include the banking, insurance, telecommunications, and utility sectors—or even public administration.

Now, using the example of a large energy company, let us demonstrate how AI can support customer service.

AI in Customer Service: Getting Started

With the advent of large language models such as ChatGPT, this company also had the opportunity to try out these tools in a large enterprise customer service environment to make work more efficient.

Within customer service, complaint handling was selected, a priority area for two reasons:

• On the one hand, because of its significant impact on customer satisfaction – after all, dissatisfied customers pose a threat to the company’s reputation.
• On the other hand, due to the large number of complaints – for example, in the banking sector, up to 8-10 thousand complaints may be received each month, which must be handled within a period specified by law.

It is important to note that customer complaints are very difficult to standardize: no two complaints are the same, each one is unique in some way.

The first step was to assess the complaint handling process, which typically looks like this:

1. The customer submits a complaint through one of several channels (by phone, post, email, or online), which is then forwarded to the customer service team specifically responsible for handling complaints.
2. The complaint is forwarded to the appropriate department if customer service is unable to provide an immediate response. In many cases, customer service is unable to respond immediately because a professional response is required that is beyond their competence. (For example, if there was a power outage in a given area, customer service cannot explain why there was a power outage, so they have to ask the department responsible for the network.)
3. The specialist department formulates a professional statement – naturally in technical language that is appropriate from both a legal and a business policy perspective – and sends it back to customer service.
4. Customer service translates the response into customer-friendly language and sends it back to the customer. It is important that the customer also understands why there was a power outage at their place of residence – because if they receive a response written in the technical jargon used by companies on a daily basis (whether in the banking sector, energy or telecommunications), there is a good chance that they will not understand the response.

The objective was to ensure that responses were not only accurate but also quick. So when a customer makes a complaint, it is very important to respond to their inquiry as quickly as possible – this is usually regulated by legal deadlines.

AI-supported Solution for Complaint Handling

The AI-based solution developed in-house is capable of quickly generating (preliminary) responses for customer service representatives, taking into account legal and business regulations. At the same time, AI does not replace human decision making, but rather assists customer service staff in their work by:

• reducing response times
• ensuring professional and customer-friendly communication
• reducing the workload of customer service representatives

It is important to note that AI makes suggestions for the response letter but does not send it to the customer – this gives the customer service rep the opportunity to modify the text before sending it, thus maintaining human control.

Technology Background: How the System Was Developed

The solution was built in the Microsoft Azure cloud, with a JavaScript-based backend and an HTML-based internal user interface for customer service staff.

RAG (Retrieved Augmented Generation) technology ensured that the AI only used the predefined documentation as a basis for formulating responses, without “adding” from other sources or “hallucinating.”

Access was restricted to the appropriate group (i.e., complaint handling agents) using the Microsoft Entra ID service.

Risk management and compliance were important considerations, as was the isolated operation of the system: the AI does not have direct access to critical systems, minimizing errors and data security risks.

Finally, it was also necessary to ensure that the system was not biased and did not use inappropriate language.

The system was developed in an agile manner: the first working version was released for testing to a few agents, and the system was finetuned based on their feedback—for example, the terminology used by the system was modified.

Compared to an average enterprise development project, development and implementation took place extremely quickly, within three to four months: essentially, a language model had to be installed from the Azure cloud, and an internal user interface had to be developed, then tested and fine-tuned.

Metrics and Results

The effectiveness of an AI application can be measured using several KPIs:

• Cost savings: lower operating costs (OPEX)
• Efficiency: faster response times to customers
• Productivity: handling more cases in a single day
• Satisfaction: NPS (Net Promoter Score) measurement based on customer feedback
• Return on Investment (ROI): quick return on resource investment – the goal is to introduce a financially sustainable application

Thanks to the AI-supported solution, the response process, which previously took an average of 8–10 days, has been reduced to 3–5 days.

What’s more, the use of AI not only increased speed but also improved customer satisfaction: NPS increased in cases where AI wrote the response letter.

Finally, customers themselves did not notice that they received (at least partially) AI-generated responses to their inquiries.

Lessons Learned

After the deployment of the system, the following conclusions can be drawn:

• AI in customer service does not replace humans but rather supports them effectively. Human control, or “human-in-the-loop,” remains important.
• Pilot testing and continuous fine-tuning are important, with the incorporation of user feedback.
• In an enterprise environment, special attention must be given to legal, ethical, and compliance issues.
• Even before implementation, it is important to assess the impact an AI tool may have on the operation of other systems—for example, to eliminate risks, this AI assistant operated in isolation and was not directly connected to other systems.

Do you think there’s room for efficiency improvements in the customer service of your organization? Why not discuss this over a cup of coffee?