Problem
Although companies invest heavily in training, the actual impact is often not as expected. Traditional training sessions and workshops tend to be sporadic and impersonal, and the effectiveness of static materials in preparing participants for real-life situations is difficult to measure. This results in infrequent practice, difficult-to-act-on feedback, and struggle in adapting training to different levels and contexts.
Solution
Roleplay is a gamified training tool based on conversational AI. It allows realistic situations with interlocutors to be recreated in a safe, controlled environment. All owned from an external platform for creating and managing contents and user players.
Role
Following a market status analysis by Business stakeholders, which would yield an initial approach to product requirements, I am tasked with proposing the sitemapping and full functionality to both simulation tool (web app) and management platform. This last one is to be hosted on GL Admin.
What I did
- Qualitative UX research: user interviews and observation. Usability testing. Tools: Google Meet, Docs.
- Conceptualization and ideation (flowcharting + sketching). Tools: Hand drawing, Draw.io
- UX/UI design (wireframing). Tools: Figma + Marvel app
- Prototyping at different levels of detail. Tools: Marvel app
- UX/UI – Dev hand off and iterations follow up for MVP, Beta and releases. Tools: Zeplin, Jira, Confluence, Asana
Why AI
The gap
Soft skills only improve by practising with someone. That someone is usually a trainer or a colleague: sporadic, costly and judging. Practice ends up infrequent, and feedback arrives late and is hard to act on.
What AI makes possible
A virtual interlocutor that replies in real time, adapting to the player's speech, tone and decisions. It is available anytime, a scenario can be repeated as often as needed, and no one is watching. Every session ends with immediate, structured feedback.
Roles
Creator
HR managers, coaches, team leads and PMs with little time for training.
Writes the scenario: the initial case, the interlocutor's interests, the goals, the difficulty and the evaluation criteria.
AI
The interlocutor.
Answers every turn, tracks trust and the interests uncovered, ends the session if the player is disrespectful, and turns the conversation into the report.
Player
Salespeople, support staff and consultants.
Reads the initial case, talks, follows the live signals, gets the report and repeats.
What it had to deliver
For players:
- Conversations that flow naturally.
- Different types of interlocutors.
- Immediate feedback after each session.
- Practice without feeling assessed by others.
- Short sessions that are easy to repeat.
For creators:
- Scenarios adapted to each sector.
- Goals and evaluation criteria per roleplay.
- Adjustable difficulty.
- Feedback aligned with each session's goals.
- Scale while reducing operating costs.
Some good design choices
Conversation as main UX
In the first usability test, the call view had several floating components around the AI interlocutor. Testers:
- Drifted away from the conversation.
- Hesitated before answering.
- Spoke less naturally.
The UI was competing with the interlocutor. The product is the conversation itself, so I applied Hick's Law: the call view keeps the dialogue and only the controls it can't work without.
Combined feedback
The CPO's first idea was highly visual live feedback, flagging hits and mistakes during the call. Before designing it, I ran informal chats with the sales team and future testers:
"I don't know if that's going to ruin the mood."
"I think I'll be thinking more about getting it right than talking."
"It doesn't seem natural to me."
Live corrections would turn practice into an exam. Feedback was split in two phases instead:
- During the call. A non-numerical score that grows as the player uncovers the interlocutor's key information.
- After the call. A downloadable report with KPIs for each conversational aspect.
Avoid group competition
A ranking came out of a brainstorm with the dev team. The CTO approved it for the MVP if it cost under one sprint and added no technical debt.
In the first iteration with testers, players:
- Skipped hard challenges to protect their position.
- Took the shortest route to points, answering with single keywords instead of full sentences.
Competition rewarded beating the score, not practising. The ranking was replaced by a stats dashboard with individual and group averages, so players track progress against themselves.
A catch: the conversational experience
Problem
Active listening, empathy, persuasion and emotional management are all essential skills for professional development in many social environments. However, people lack consistent or reliable opportunities to train these skills.
This can lead to frictions such as fear of judgement, or delayed, not actionable feedbacks.
Solution
A functional proposal for an AI-based conversational simulation that enables users to hone their communication skills in realistic, contextualised scenarios.
The product recreates a videocall environment and generates dynamic dialogue with a virtual interlocutor who responds in real time, adapting to the user's speech, tone and decisions.
After each session, the user receives a structured report combining clear metrics with qualitative feedback to help them understand their strengths and areas for improvement, but without turning the experience into an exam.
UX flow
The functional proposal primarily focuses on effectively separating the consultation and conversation parts. Once this has been achieved, the user can immerse in the roleplay experience through a sequence that makes the cognitive process as smooth as possible, enabling focus on responses from the outset.
The main premise is to maximize player user retention and ensure a structured learning experience through modularity. To achieve that, each roleplay is contained within a card (Roleplay 1, 2, 3...), allowing for easy scalability of content.
The videocall is the heart of the prototype, where the interaction logic is tightly controlled to prevent user frustration.
The logic accounts for "Solved," "Unsolved" (Time out), and "Ended" states, all funneling into the reporting engine.
UI design
A progressive training strategy is suggested to keep focus on the main task. During the conversation, semantic tokens communicate success and failure states, enabling intuitive progress within a minimalist learning-focused approach.
Initial case
A modal window provides the player user with relevant information about who will be talking to in advance. This helps to focus the conversation and set expectations. In terms of interaction, functional friction is minimised by providing a single CTA.
Scenario
Throughout the conversation, the UI remains clean and focused on the dialogue, displaying key indicators in real time.
Options
Various contextual features are accessible without overloading the main UI. These features are provided in the form of dialogs, with confirmation required where necessary.
Report
The session closes with a performance report in a constructive tone. It avoids value judgements so the player reflects instead of feeling graded.
- Average score (0–100).
- General indicators (0–100): trust, interests, objections and mastery of information.
- Speech indicators (😞 😐 😊): balance, questions, speech rate and clarity.
- Recommendations (not scored): strengths and weaknesses.
Another catch: the training management
Problem
Alongside identifying the needs of the player user, it is crucial to monitor their progress in a scalable and holistic manner that would allow to accompany the user journey and anticipate any training obstacles that could undermine progress.
Solution
A platform has been defined for content creation, customer management and player users, operating within the GL Admin umbrella but as a completely independent and companion app to the main product. There are two operational levels, depending on the role:
- Customer Success. This is the most in-depth level, aimed at roles responsible for global customer follow-up. It enables the management of games and player users for each customer, as well as the creation of content from scratch and access to the dedicated player database section.
- Client Administrator. We could say this middle level is a lighter version of the CS one, designed so that customers can control their own content and player users, but with no access to others data.
UX flow (CS)
Simplifying the diagram's processes also leads to a flow focused on optimizing the experience. In this way, we see similarities between the customer detail and the Roleplay AI master section, which are translated into views that are very similar in terms of functionality. Although this may seem redundant, it enables the reuse of code and allows the creator-administrator to more easily ensure operational control.
UI design
The aim is for hierarchy to play a key role in relation to data. Different types of grids and context modes minimize operational friction, providing management with full visibility and control of the platform while maintaining the user's focus.
Dashboard
A complete summary of activity, metrics and player user status, filtered by customer, is provided through a series of widgets arranged in a three-column grid. The importance of the information determines the visual weight of each widget. The general idea is to balance between components and spaces for an accurate, error-free understanding of what is happening. Simple animations would be proposed to accompany the evolution of the data showed in graphics.
Roleplays
The Roleplays tab facilitates control and administration of all ongoing sessions from a general perspective, including editing and testing features. Once inside, two views allow to show contents according on administrator preference: Grid, if want to go more visual, and Table, rather practicality-oriented.
User players
Working almost like a microsite itself, it offers a feature that allows users to add, edit and delete accounts, either individually or in bulk, both in games that have already been created and freely for later assignment. This streamlined process eliminates unnecessary steps and interactions that would otherwise slow down the process.
Subscriptions
The visual architecture employs a grid card to organise subscriptions in a modular fashion. This layout makes it easier to have an accurate overview of the general plan status.
Modal overlays, which are activated via context menus, avoid subviews reducing friction while preserving the user's navigation context. Integrating them into both the customer detail and Roleplay AI master section ensures consistent, centralised management of the platform across modules.
Takeaways
Lather, rinse and repeat. Always repeat
When it comes to design for enhancing soft skills, good UX is measured by how often the product is used. Enabling users to practise more often with less friction is key.
User don't give a **** about biz
Designing user-center-based without understanding the expected business impact generates products that may appear correct, but will ultimately be irrelevant.
Happy flow = happy code
Defining a clean and simple flow enables devs to implement it with less ambiguity and detect potential scaling issues in advance.