AI Platform & Chat
Building a conversational AI platform that connected Calendly’s meeting, scheduling, and relationship data into one intelligent experience.
Role
Staff Product Designer + Interim Product Manager
Team
Tiger team of myself + 4 engineers
My Role
Led product strategy and design from concept through Beta launch
Opportunity
Meeting context was fragmented across multiple tools
Customers were using multiple tools across scheduling, CRM, meeting notes, and AI to manage their client relationships—creating fragmented workflows, context switching, and missed follow-ups.
Calendly's AI Notetaker was already helping users capture individual meetings, but users couldn't search across their meeting history or act directly on what the AI surfaced.
Research
Three recurring needs across user types
Customer research and AI Notetaker feedback surfaced three recurring needs:
Preparing for meetings
Acting on follow-ups,
Understanding patterns across meetings and relationships.
Strategy
Create a reusable AI platform for Chat and embedded AI workflows
I partnered with engineering to scope the Horizontal AI Platform—a unified agentic data layer connecting Calendly's relational data with an OpenAI backend. By establishing standardized tool definitions (Send email, CRUD meetings, CRUD contacts), we ensured the engine could power both the immediate conversational assistant and future embedded AI workflows.
AI Behavior
Defining how the assistant handled ambiguity and uncertainty
I defined and iterated the response rules supplied to the agent, shaping how it handled ambiguity, missing data, unsupported requests, feature discovery, and tone. Testing exposed where responses felt inaccurate, overly restrictive, or unhelpful, and I used those findings to refine the rules.
Response boundaries
Keep answers grounded in supported meeting and follow-up workflows.
Feature discovery
Surface relevant Calendly capabilities when they help complete the user's goal.
Ambiguous requests
Ask clarifying questions instead of guessing
Unsupported requests
Explain the limitation and provide a useful fallback.
Missing data
Be transparent about what was searched and what wasn't found.
Progressive loading
Show task-specific progress cues while the model works to make latency clearer.
Beta Release
Launching Beta to 10,000 users in 90 days
We launched AI Chat Beta to 10,000 users to validate the assistant’s value, discoverability, and interaction model.
Beta showed strong activation—but friction after the first query
30% of AI Assistant visitors submitted a first query, while the lower repeat-query rate and 3+ step path to core actions revealed opportunities to make capabilities clearer and actions easier to reach.
18%
Discovery Rate
Users who visited the AI Assistant tab
30%
Activation Rate
Users who submitted an initial query
11%
Engagement Rate
Users who submitted a second query
3+
Steps to action
Required to complete a core action
UX Enhancements
Redesigning around discovery, activation, engagement, and action
Based on Beta behavior and usability feedback, I designed the next iteration around four opportunities: discovery, activation, continued engagement, and faster access to actions. These updates were in development when I left Calendly.
Updated Experience
Connecting insights to follow-up actions
The next iteration connected personalized starting points, contextual recommendations, interactive Calendly objects, and actions into one continuous conversational workflow.
Impact
Launching AI Chat, then scaling the AI Platform
We launched AI Chat Beta in 3 months and reached 30% adoption among AI Assistant visitors in the first month. The Horizontal AI Platform later expanded to power two in-app experiences, an external Gmail agent, and a Calendly MCP server integrated with Claude and ChatGPT.
3
Months to initial beta launch
30%
Adoption among AI Assistant visitors
4
Experiences powered by the AI Platform
Takeaways
What Beta changed about our AI strategy
Capabilities need visible starting points: Open-ended AI still needs examples that communicate what the system can do.
Response behavior is part of the product experience: Clarification, fallbacks, boundaries, and tone directly affect usefulness and trust.
Context can sometimes beat prompting: Some needs were predictable enough to surface directly inside existing workflows.