Back
Designing Leena's Agentic Assistant
Designing Leena's Agentic Assistant
I joined Leena AI in 2019 when the product was a rule-based HR chatbot. Over five years, I led the interface through three distinct evolutions
Chatbot → Generative AI Assistant → Autonomous Agentic Colleague
each one changing what users needed to trust, not just what the product could do. I led the Leena's Agentic Assistant redesign project to make sure
I joined Leena AI in 2019 when the product was a rule-based HR chatbot. Over five years, I led the interface through three distinct evolutions
Chatbot → Generative AI Assistant → Autonomous Agentic Colleague
each one changing what users needed to trust, not just what the product could do. I led the Leena's Agentic Assistant redesign project to make sure
ROLE
Lead Product Designer
DURTION
2 months · Shipped to production
RESULTS
3X adopted across HR, IT, Procurement, Sales, and Finance, on a single design system spanning web, mobile, Slack, and MS Teams.
2X user retention over the product feature
why this mattered
Each stage of AI capability introduced a new kind of user doubt. A rigid chatbot frustrated people with dead ends. A generative assistant impressed people but left them unsure how far to trust it.
An agentic assistant that could act on its own — booking meetings, drafting responses — raised the highest-stakes question yet: how do I trust it to do the right thing without watching it?
Each stage of AI capability introduced a new kind of user doubt. A rigid chatbot frustrated people with dead ends. A generative assistant impressed people but left them unsure how far to trust it.
An agentic assistant that could act on its own — booking meetings, drafting responses — raised the highest-stakes question yet: how do I trust it to do the right thing without watching it?
The Problem
Four things were broken not just one:
1. Chatbot rigid decision trees, repetitive "I didn't understand" dead ends, no discoverability of what it could even do.
2. Generative Assistant hallucination risk, inconsistent tone, unpredictable answer length raising cognitive load.
3. Agentic Assistant fear of losing control ("will it make the wrong decision?"), no visibility into multi-step actions taken on a user's behalf.
Four things were broken not just one:
1. Chatbot rigid decision trees, repetitive "I didn't understand" dead ends, no discoverability of what it could even do.
2. Generative Assistant hallucination risk, inconsistent tone, unpredictable answer length raising cognitive load.
3. Agentic Assistant fear of losing control ("will it make the wrong decision?"), no visibility into multi-step actions taken on a user's behalf.
Research Summary
Contextual inquiries with HR employees, conversation-breakdown mapping, and later trust-perception studies with HR/IT stakeholders showed the same pattern at every stage: capability was never the real bottleneck — legibility was. Users needed to see why the assistant did what it did, not just confirmation that it worked.
Contextual inquiries with HR employees, conversation-breakdown mapping, and later trust-perception studies with HR/IT stakeholders showed the same pattern at every stage: capability was never the real bottleneck — legibility was. Users needed to see why the assistant did what it did, not just confirmation that it worked.
From Insights to Interface
Chatbot → guided prompts, fallback flows instead of dead ends, quick-reply affordances to cut typing fatigue.
Chatbot → guided prompts, fallback flows instead of dead ends, quick-reply affordances to cut typing fatigue.
Generative Assistant → conversation scaffolding (suggested prompts instead of a blank box), progressive disclosure instead of raw errors, lightweight thumbs up/down feedback loops.
Generative Assistant → conversation scaffolding (suggested prompts instead of a blank box), progressive disclosure instead of raw errors, lightweight thumbs up/down feedback loops.
Agentic Assistant → action preview & undo before anything executes, confidence indicators ("85% sure this matches your request"), step-by-step workflow logs so the assistant was never a black box, human-in-the-loop checkpoints throughout.
Agentic Assistant → action preview & undo before anything executes, confidence indicators ("85% sure this matches your request"), step-by-step workflow logs so the assistant was never a black box, human-in-the-loop checkpoints throughout.
Scaling → a modular design system with reusable tokens and conversation patterns, so expanding from HR into IT, Procurement, and Sales didn't mean rebuilding the UX each time.
Scaling → a modular design system with reusable tokens and conversation patterns, so expanding from HR into IT, Procurement, and Sales didn't mean rebuilding the UX each time.
Testing the Assumptions
Usability tests on early Slack integrations, A/B tests on conversation patterns, and heuristic evaluations for accessibility ran at every stage, not just at the end. Drop-off fell after the chatbot fixes; usage shifted from cautious, one-off questions to confident, everyday use once generative scaffolding landed.
Usability tests on early Slack integrations, A/B tests on conversation patterns, and heuristic evaluations for accessibility ran at every stage, not just at the end. Drop-off fell after the chatbot fixes; usage shifted from cautious, one-off questions to confident, everyday use once generative scaffolding landed.
Results & Impact
Full design ownership across three product generations: chatbot MVP → GPT assistant → agentic colleague.
One scalable design system adopted across HR, IT, Procurement, Sales, and Finance.
Seamless cross-platform experience: web, mobile, Slack, MS Teams.
Full design ownership across three product generations: chatbot MVP → GPT assistant → agentic colleague.
One scalable design system adopted across HR, IT, Procurement, Sales, and Finance.
Seamless cross-platform experience: web, mobile, Slack, MS Teams.
Retrospective
The biggest shift in my own thinking: as autonomy increased, the design problem stopped being about conversation and became about accountability. Confidence indicators and action previews weren't polish — they were the actual mechanism that let people hand off real decisions to the assistant. Balancing cutting-edge AI capability with that kind of legibility, at every stage, mattered more than any single interaction pattern.
The biggest shift in my own thinking: as autonomy increased, the design problem stopped being about conversation and became about accountability. Confidence indicators and action previews weren't polish — they were the actual mechanism that let people hand off real decisions to the assistant. Balancing cutting-edge AI capability with that kind of legibility, at every stage, mattered more than any single interaction pattern.
