Case Study // Vessa.ai
Designing a Character-Led Conversational Concierge for a Text-Only Interface
Vessa.ai set out to build a restaurant booking concierge delivered entirely through WhatsApp. The product needed to provide high-quality, personalised recommendations while maintaining a distinct, brand-aligned personality.
1. Context & Problem Framing
Vessa.ai set out to build a restaurant booking concierge delivered entirely through WhatsApp. The product needed to provide high-quality, personalised recommendations while maintaining a distinct, brand-aligned personality.
This created a set of overlapping challenges:
Without a clear behavioural system, the AI risked becoming inconsistent, overly verbose, or unreliable, effectively improvising the brand in real time.
2. Role & Responsibility
I worked directly with the founder to design the behavioural and interaction layer of the product.
3. Approach & Reasoning
Key Insight
In a text-only environment, there is no UI to structure the experience. Clarity, pacing, and trust are entirely carried through language.
1. ICP Interviews (PAs supporting HNW clients)
Insight: Users prioritise speed and confidence over exploration.
Decision: Shift from conversational discovery to a recommendation-first model.
2. Service Philosophy (e.g. "Unreasonable Hospitality")
Insight: Great service anticipates needs, not just responds.
Decision: Embed proactive suggestions such as drinks, transport, and proximity.
3. Technical Constraints (WhatsApp, APIs, memory)
Insight: The system cannot reliably answer all queries.
Decision: Design for explicit limitation handling, not silent failure.
Lead with strong, confident suggestions before asking follow-up questions.
Design around what the system does not know, not just what it does.
Use phrasing, formatting, and pacing to replace missing UI elements.
Reduce unnecessary back-and-forth to reach value quickly.
Prioritising speed and clarity reduced opportunities for open-ended exploration, but significantly improved trust and usability in a messaging context.
4. System Design & Execution
Rather than writing a single monolithic prompt, I designed a modular character and interaction system.
A
Character DNA
Defined core, immutable traits: service orientation, confidence level, communication style, and boundaries and limitations.
Character Bible (Living System)
Centralised all behavioural logic: tone and voice rules, interaction patterns, refusal and repair strategies, emoji usage and constraints, and self-awareness and scope boundaries.
This acted as the single source of truth for all outputs.
B
Designed a consistent response structure:
This replaced typical chatbot behaviour: question-heavy, low-confidence responses.
C
Created structured examples for key interaction types:
These reduced variability and improved reliability across edge cases.
D
Developed a constrained emoji framework:
E
Due to incomplete API coverage, I built a heuristic mapping system linking dietary needs to cuisine likelihood.
This enabled the system to make informed recommendations under uncertainty while reducing hallucination and improving user trust.
F
Trained a GPT on the full character system to act as a real-time reference tool for the team, provide consistent answers to edge cases, and enable rapid content creation in Vessa's voice.
This functioned as an operational layer for maintaining consistency, not just a writing tool.
5. Outcomes & Evidence
6. Reflection & Transferable Insight
Transferable Insight
In constrained conversational environments, behavioural design replaces interface design.
Without a defined system, AI defaults to generic outputs, brand becomes inconsistent, and trust erodes quickly.
With a structured approach, interaction becomes predictable and scalable, character reinforces product value, and the system can extend across channels such as voice and avatars without redesign.