Leena AI
From Chatbot to Agentic Colleague
I led the interface evolution behind Leena AI's assistant, from a rule-based FAQ bot to a full agentic AI colleague, across five product generations: rule-based chatbot, NLP-based chatbot, generative assistant, and finally agentic assistant.
Role
Lead Product Designer
Timeline
Three product generations · Shipped to production
team
1 Designers & 1 PM
Scope
Mobile, Agentic AI, Conversational UX
Final Results
3x adoption across HR, IT, Procurement, Sales, and Finance
2x user retention over the previous product experience
2x user retention over the previous product experience
Why this mattered
Two problems hit at the same time.
Trust: once the assistant started proposing autonomous actions, people didn't trust it enough to let it act.
Scope: the product grew from a handful of FAQ use cases to fifteen, spanning HR, IT, Procurement, Sales, and Finance. A single misread request could now break someone's actual workflow, not just annoy them.
The problem, by stage
Stage | What was broken |
|---|---|
Rule-based chatbot | Decision trees, dead ends, no discoverability |
NLP-based chatbot | Couldn't hold context, users lost their place mid-conversation |
Generative assistant | Hallucination risk, inconsistent tone, unclear visibility into answers |
Agentic assistant | No visibility into multi-step actions, users afraid of losing control |
What research surfaced
24 interviews, 1,000+ conversation logs, and 65 usability sessions built out 4 personas and pointed to four recurring themes:
Discoverability: people didn't know what the assistant could actually do
Conversations didn't scale: context got lost the moment a request spanned more than a couple of turns
Trust needed transparency: confidence in an answer meant nothing without a reason behind it
Autonomy needed control: the more the assistant could act on its own, the more people wanted a say before it did
The decision that shaped everything after
Once the assistant could act, not just answer, conversation design stopped being enough. The interface had to account for itself: what it did, and why, every time.
That single shift, from designing for conversation to designing for accountability, is what the next five changes were built around.
Guided chatbot
Replaced dead-end intent matching with guided prompts, fallback flows, and structured forms, so the product didn't depend on getting every phrasing right.
A persistent workspace
Chat couldn't carry fifteen use cases on its own. A modular workspace became the real anchor, with chat as one entry point among several.
A system that scaled across departments
Built modular enough that patterns proven in HR could extend into IT, Procurement, Sales, and Finance without a rebuild each time.
Calibrated trust for the generative assistant
Scaffolded prompts, progressive disclosure, and lightweight feedback let people build trust turn by turn instead of all at once.
Accountability for the agentic assistant
Action previews, per-action confidence indicators, step-by-step logs, and human-in-the-loop confirmation before anything irreversible happened.
Testing before trusting the design
Every major decision went through usability testing and A/B experiments rather than shipping on instinct. Testing confirmed that better discoverability cut down on repeated fallback conversations, that structured request flows improved task completion over open-ended chat, and that visible reasoning increased how often people let the assistant act autonomously.
Results
3x adoption across HR, IT, Procurement, Sales, and Finance
One design system scaled across 5 departments with no rebuild
Fallback-triggered conversations dropped once guided flows replaced dead-end intent matching
Structured tasks completed faster once they had a defined path outside open chat
Trust in autonomous actions grew as the assistant's reasoning became visible step by step
Looking back
As the assistant's autonomy grew, the real design problem shifted from conversation design to accountability design. Confidence indicators and action previews weren't polish, they were what let people actually hand off real decisions to the assistant instead of double-checking everything it did.
