DESIGNING AI SUPPORT FOR RETAIL STAFF

AI is becoming increasingly common in workplace environments, and more companies are exploring how it can support internal workflows. In this project, we collaborated with Mio, who wanted to understand how an AI chatbot could support retail staff in their daily work.

The retail environment is fast-paced and often unpredictable. Staff constantly switch between tasks, help customers, and search for information on the spot. When they don’t have an answer, they usually rely on colleagues, suppliers, or internal systems. This often takes time and interrupts the flow of work.

What became clear early on was that access to information is not just important, it is critical. When answers are delayed, it directly affects both the employee’s workflow and the customer experience.

AI has clear potential to support retail staff by making information more accessible and reducing the need for manual searching. At the same time, introducing AI into a work environment brings new challenges.

From the research, it became clear that it is not enough for the system to work well technically. How it is perceived by users is just as important. Staff raised questions about whether they could trust the answers, where the information comes from, and what happens to the data they enter.

This creates a tension where the same system can feel either helpful or uncomfortable depending on how it is designed.

To understand these needs, we conducted seven semi-structured interviews with retail staff and complemented this with a survey. The goal was to capture how they currently work, where friction occurs, and how they feel about introducing AI into their workflow.

The material was analyzed using thematic analysis to identify recurring patterns in both behavior and perception.

From the research, four key themes emerged that shaped how the AI needed to be designed in this context.

While participants were generally positive towards AI, their expectations were very specific. They did not want a system that tried to replace their role, but one that clearly supported their work. Especially in situations that required judgment, it was important that the user remained in control.

At the same time, the retail environment placed strong demands on speed and clarity. Interactions needed to be quick, direct, and easy to process, without unnecessary complexity or conversational elements that could slow users down.

Another recurring theme was trust. While participants were open to using AI, this trust depended on the system being transparent and reliable. Being able to understand where information comes from, and to recognize when something might be incorrect, was critical.

Finally, the value of the system was closely tied to its ability to reduce workload. If the AI added friction or required additional effort, it risked being ignored entirely.

How might we design an AI chatbot that supports retail staff in fast-paced situations while still maintaining trust, clarity, and a sense of control?

The solution needed to be easy to understand and quick to use. It should support staff without getting in their way, reduce cognitive load during busy moments, and make relevant information available exactly when it is needed. Just as importantly, it needed to feel trustworthy and transparent.

To guide the design, we translated both the research insights and existing AI guidelines into a set of principles.

The system should be transparent, so users can understand what it does and where information comes from. It should give users control over the interaction, including the ability to guide and correct it. Communication should be clear and easy to scan, especially in time-sensitive situations. Finally, the AI should be positioned as a tool rather than something that mimics human behavior.

In the ideation phase, we explored how an AI chatbot could fit naturally into the existing workflow. The focus was not on adding new features, but on reducing friction in tasks that were already part of the workday.

We looked at how the system could provide answers quickly, support information search in real time, and avoid interrupting ongoing interactions with customers.

The concept resulted in an internal AI chatbot called Maia. It is designed to give retail staff fast access to relevant information through a conversational interface.

Maia already exissted as an early prototype, presented below. Technically, the system combines a language model with retrieval-based methods, allowing it to generate responses based on internal company data. This makes it possible to provide more accurate and context-specific answers.

The project resulted in a set of adapted design guidelines for AI in retail, along with a concept for an internal chatbot.

It also highlighted several UX risks related to trust, transparency, and user control, which are critical to address when introducing AI into a work environment.

This project showed me that designing AI for a workplace is not only about making tasks faster. In a retail environment, speed can be valuable, but it also raises a bigger question: am I reducing friction for the user, or am I reinforcing an already stressful workflow?

A key learning was that trust is shaped by more than the interface. Features like memory, personalization, and automation can be useful, but in a workplace they also connect to power, privacy, and organizational culture. Even if a system is technically transparent, users still need to feel that it works for them, not as something that monitors them.

I also learned that AI guidelines cannot simply be applied as fixed rules. They need to be interpreted through the specific context, the users’ expectations, and the ethical tensions that appear in real use. In this project, that meant designing Maia as a clear tool under human control, rather than a system that feels too intelligent, social, or autonomous.

The next step would be to test Maia in a real retail environment. Since the hi-fi prototype was not tested in actual use, there are still assumptions around how well the interaction works during customer conversations, time pressure, and divided attention.

I would also want to involve retail staff in more iterative test cycles. Observations, follow-up interviews, and feedback sessions could reveal needs that did not appear in the initial research.

Another important direction would be to further explore transparency, privacy, and explainability. This includes how sources are shown, how uncertainty is communicated, and how the system explains what is stored or not stored.

Finally, future work should look more closely at accessibility and inclusion, such as screen reader support, color contrast, keyboard navigation, and how the chatbot works for users with different levels of technical confidence.