Building a travel spend dashboard with AI

Creating the user interview guide

To deepen my understanding of the Travel Manager’s expectations, I crafted a structured interview guide. This guide followed UX research best practices and was divided into key sections: user profile, company context, dashboard usage, and pain points.

The questions explored how Travel Managers currently monitor travel data, what information they value most, and which features would make their daily work easier. By generating and refining this guide with AI, I ensured a comprehensive and unbiased framework ready to simulate realistic user interviews.

Extracting user insights

Using the interview guide, I simulated five realistic user interviews with Travel Managers of different ages, company sizes, and levels of tech proficiency.

Each persona provided unique perspectives on their challenges, workflows, and expectations regarding travel data management.

Once the interviews were completed, I analyzed and clustered the insights by theme, such as data accessibility, reporting frequency, and desired automation features. This synthesis allowed me to clearly identify recurring needs and prioritize the most impactful opportunities for the dashboard design.

Key takeaways for design

• Build an AI-powered, modular dashboard that adapts to different user profiles (tech-savvy vs traditional)

• Focus on simplicity, personalization, and automation rather than feature overload

• Include natural language interactions and automated reporting for time-pressed users

• Combine business KPIs + sustainability metrics to align with modern corporate goals

Creating a realistic dataset

To design a dashboard that truly reflects real-world usage and edge cases, I needed a believable dataset that would make my design decisions more grounded and relevant.

To generate this realistic dataset, I asked ChatGPT to:

• Define the key data categories relevant to a Travel Manager’s dashboard (employees, departments, destinations, carbon emissions...)

• Generate a fake dataset containing diverse travel records across multiple departments and global destinations

• Ensure data variability by including spend amounts, trip dates, anomalies, and sustainability metrics (CO₂ emissions)

• Expand the dataset to 200 unique entries to ensure statistical relevance and realistic visualization potential

• Export the final dataset into a CSV file format, ready to be used for dashboard design and testing

AI-assisted exploration

I first asked ChatGPT to generate a detailed prompt that I could then use with other LLM-based design tools. Using this prompt, I created wireframes with Lovable and V0, comparing how each tool interpreted the brief and translated user needs into design solutions. I asked each tool to explore four different layout possibilities.

Iterating on the wireframes

After reviewing the wireframes, my preferred concept was the split-screen layout, for several reasons:

• It introduced a disruptive and modern approach compared to traditional dashboard designs

• It integrated the AI Assistant directly within the interface, encouraging users to interact with it naturally

• It provided more space for AI-driven features, such as smart suggestions, clearly demonstrating the assistant’s added value

Building on this concept, I asked Lovable to refine the layout by incorporating the following improvements:

• Allow users to close, reduce, or reopen the AI Assistant panel (positioned on the right)

• Keep the AI panel open by default, while maintaining the flexibility to close it

• Ensure the left panel is responsive, expanding to full width when the assistant is closed

• Add an “Employee” filter, complementing the existing Department and Date range options

As a final adjustment, I asked Lovable to reposition the collapse button so it sits centrally between the two sections, rather than being placed entirely on the right side of the interface.

I chose not to focus too much on small details at this stage, as it was only a first draft. It allowed me to get a clearer vision of the dashboard’s overall structure and interactions. In my view, it effectively captured the intent of creating a dynamic, AI-assisted data visualization, which led me to progress toward the high-fidelity prototyping phase.

Build interactive prototypes and iterate with AI feedback

For the prototyping generation phase, I returned to ChatGPT to create a detailed and structured prompt specifically designed for Lovable. The goal was to move from wireframes to a high-fidelity prototype that visually captured the final look and feel of the dashboard.

In the prompt, I included the brand color palette to define the visual identity and ensure consistency, as well as the CSV file containing realistic travel data, which allowed Lovable to populate the dashboard with credible and data-driven content.

After generating the first version of the prototype, I decided to run several iterations to refine both the visual balance and usability of the interface.

Here are the main improvements I applied:

• Adjusted the background color to a light grey tone for better readability and contrast.

• Fixed alignment issues within the Smart Suggestions section

• Updated the color palette to achieve a more professional and cohesive look

• Increased the contrast between the left and right panels, making the split-screen structure more visually distinct

• In the conversation section:

  • Added a “New Conversation” button at the top right corner of the card for quick access.

  • Included a message input field below the chat messages, allowing users to easily continue their conversations with the AI assistant

Iterating with AI feedback

To refine my prototype, I shared a screenshot of the dashboard with ChatGPT and asked for a detailed UX and accessibility review. Based on the feedback, I then requested a structured prompt specifically tailored for Lovable to apply the recommended improvements. After generating the updated version in Lovable, I made a few final adjustments myself to polish the visual design.

Simulated testing

To test my prototype's usability, I provided ChatGPT with the following hypotheses:

• The double scroll/navigation might be confusing

• The number of graphs on the page might not be enough

• We might need more filters

• The "anomalies and alerts" section might be too highlighted

• We might need sub-sections in the dashboard, with a navigation bar

I then asked it to generate a detailed usability testing plan to validate or invalidate each of them. I asked for the following items to be included in the test plan: the test objectives, the typology of participants, the method, the tasks, and the success metrics.

ChatGPT generated the test plan, and I ran the test simulation.

I then shared the key recommendations from the usability report with Lovable and asked the tool to apply the improvements directly to the prototype.