The smart colleague – AI agent supports internal sales at VAUDE

VAUDE develops durable and functional products according to the “Design for Life” principle. The focus is on recyclability, reparability, and the lowest possible use of resources. In this way, VAUDE combines a passion for mountain sports with responsibility for people and the environment.

In addition to distributing VAUDE products to retailers, the company, founded in 1974, now operates 22 franchise stores in the DACH region. The planning and manufacture of customized products is also possible in cooperation with VAUDE: a separate business unit focuses on equipping companies with custom-made or specially branded products.

To support this broad range of services, the outdoor and sports equipment supplier has been relying on BE-terna Fashion based on Microsoft Dynamics 365 Finance and Operations since 2021. “With the new solution, we have replaced our old and highly customized system and taken a big step toward making our IT landscape future-proof with the new cloud technology,” emphasizes Daniel Köhler, IT manager at VAUDE. "Despite all the improvements, there are also processes that have remained the same with the new solution. The focus was on processes at the ‘human-machine interface’, i.e. interactions with the system in which users transfer or read data from other sources (e.g. emails) into the system. We wanted to automate these processes."

Compact AI project with tangible benefits

As a long-standing software and consulting partner, BE-terna knows VAUDE's IT inside out – and at the same time has extensive expertise in the field of state-of-the-art AI. This technology appeared to be an attractive option in the context of the objective, not least because VAUDE already has a very good foundation for fast and cost-effective implementation of corresponding scenarios through its use of the Microsoft ecosystem: functions such as document analysis, the use of AI language models, and scaling options for Azure services are already available there.

"When we took a closer look at the possibilities together with BE-terna, we had the impression that there are many processes that have the potential to be automated through the use of AI agents. However, we needed an approach that could be implemented even with limited resources,“ recalls Daniel Köhler. ”So we decided to set up a pilot project (MVP) and develop an agent in the internal sales department to take some of the pressure off."

At that time, the daily work for the employees there was characterized by a high degree of manual effort: They had to process and respond to dozens of email inquiries every day, enter corresponding orders into the system, compile backorder lists, and handle customer return inquiries. Against this backdrop, there was often no time for further consultation or upselling opportunities.

The AI colleague reports for duty

The planned AI agent was intended to support the internal sales team in two specific scenarios: it was to enter incoming orders into the ERP system and, if necessary, generate backorder lists for forwarding to customers. The project was estimated to take 12 weeks – a period during which the AI experts at BE-terna not only designed, developed, and implemented the AI agent. Testing was also carried out within the three-month period: VAUDE employees repeatedly fed the agent with test emails and checked whether it was processing the tasks correctly.

In mid-September, right on schedule, the AI agent went live and has since been supporting internal sales staff by doing valuable preliminary work on incoming inquiries: to date, 71 percent of orders have been correctly identified by the AI agent and entered into the system correctly. For the remaining orders, the AI agent correctly recognized that it did not have the ability or permission to process them and forwarded them to an employee accordingly. Inquiries about backorder lists have so far been handled entirely by the AI agent.

Step-by-step approach of AI

The inquiries that the AI agent is confronted with in practice could not be more diverse. From image attachments of products photographed from the VAUDE web shop with instructions such as “This bag in blue, please!” to clearly formulated orders with the complete item number, order quantity, and customer data via attached Excel lists in customer-specific structures – the AI agent must be able to handle all these scenarios, and it can. To do this, the AI relies on a multi-step approach and uses several tools at its disposal:

When a new email arrives in the central mailbox monitored by AI, it first triggers the agent, who then takes action. The agent checks the content and decides whether the email contains a task that they can and are allowed to take on. To do this, it is important to understand the intention correctly. “If the content of an email reads, for example, ‘I have a question about my order for item XY,’ this does not mean that this email is the actual order,” explains Daniel Köhler. The agent must therefore not only respond to specific keywords such as “order,” but also understand the context of a conversation and draw the right conclusions from it.

Once a task has been identified that the agent can take on, the relevant item, order, and customer information must be extracted from the email. Identifying and compiling the item number in particular presented a challenge. At VAUDE, an item number always consists of the product number, color, and size. However, it is not uncommon for this information to be described only in words in an email, for example, “I would like product no. 459 in blue and size 38.” The agent must compile the correct item number from text like this. Finally, order information such as desired quantities or delivery dates must be extracted from the email and assigned to the correct customer in the customer database.

Once the agent has successfully completed all these tasks, it imports the order into the ERP system. To do this, it uses the EDI import mechanism of the BE-terna Fashion industry module, as this runs a wide variety of check routines (Does the customer exist? Does the desired item exist?) in the background as standard. Finally, the AI agent formulates an email response for the customer.

Transparency and control are paramount

The AI agent handles the entire process with maximum transparency: In a specially created interface, it writes a short feedback message to the employees and describes in detail how it specifically proceeded to process the respective email. Which decisions were made and on what basis? From which sentences in the email was data such as article information extracted? What actions were taken? It also shows which identified tasks it was able to perform itself and which, if any, still remain for the employees. In doing so, it only ever takes on tasks that it is actually allowed to perform within its defined limits (so-called “boundary conditions”).

As “humans in the loop,” employees review the processing and approve the proposed implementation for the agent. “From a purely technical standpoint, it would of course be possible to automate the process without human confirmation,” explains Köhler. “But we want to gain experience here. Even with the intermediate step of approval, the agent provides us with enormous relief: instead of having to process every email manually, our colleagues receive a completely finished processing proposal that they simply have to confirm with a click. Humans remain in control, while the agent takes over the time-consuming work.”

Expansion possibilities yes, independent learning no

In order to complete the individual components of its tasks, the AI agent has a portfolio of various “tools” at its disposal that enable it to extract data from text, analyze attachments, or formulate email responses, for example. The agent's range of capabilities could be expanded accordingly by providing new tools. VAUDE and BE-terna are currently working on further improving the existing tools and creating specific tools for special cases.

“However, our agent does not learn independently,” emphasizes the IT manager. “This may come as a surprise, because at first glance it seems very practical for the agent to be able to adapt to new circumstances on its own. But you have to remember that the agent works in our production system. Incorrect entries or actions could have fatal consequences. We therefore want tasks to be processed in the intended manner and to retain control over them at all times.”

Looking ahead to the next stages

“The first stage was a complete success, but our AI journey is far from over,” concludes Daniel Köhler. Thanks to the successful use of the AI agent in the internal sales department, other departments at VAUDE have now also become aware of the advantages of AI support. Accordingly, Daniel Köhler and his team are working closely with the experts at BE-terna to develop new scenarios.

For example, it would be conceivable to support end-user communication on repairs or general questions about sustainability with another AI agent. In this context, the tool approach could play to its strengths in the future, as agents could share individual tools or pass on partial steps to each other for processing (“You can write emails, please reply to my customer here”). Accordingly, skills do not have to be developed individually for each agent. To expand the use of AI, VAUDE plans to provide specific training to employees in the future so that they can support the development, training, and use of agents in their respective areas. 

“The basic idea behind our AI efforts is automation. We no longer think ‘only’ in terms of efficiency gains, but want to relieve employees of repetitive tasks and have them performed by AI,” emphasizes Daniel Köhler.