From 100M minutes of calls to a focused MVP.
The AI Sales Agent MVP used real call data to solve key sales problems and improve results. We analyzed over 100 million minutes of sales calls from thousands of representatives, identified where reps fumbled, shipped a working agent in three weeks, and let the market pick the vertical.
We started with the data
// COMPETITIVE ADVANTAGEWe had access to over 100 million minutes of recorded sales calls from thousands of representatives. I worked closely with the data team to analyze what was actually happening in these conversations: where reps got stuck, where prospects disengaged, and what triggered common objections. Rather than relying solely on dashboard metrics, we focused on understanding the actual conversation dynamics.
We used Whisper for transcription and OpenAI models to identify initial patterns in the conversations. Claude Opus, while expensive, proved invaluable for this analysis. Our goal wasn’t perfection in the first iteration, but rather gathering sufficient signal to build a working prototype.
Where reps consistently lost the call
- Talking too much before qualifying
- Missing tonal cues when buyers hesitated
- Mishandling price objections
This analysis provided the foundation for our training approach. We translated these insights into specific tasks and developed our initial agent.
Kixie’s data foundation
What we had to work with.
Built the MVP with constraints
// NOT HYPOTHETICALSThe first version accomplished its core objectives: answering calls, following structured scripts, and managing basic objections. Rather than attempting to solve every possible use case, we concentrated on high-volume, lower-risk scenarios like initial lead capture and first-touch interactions.
Following a week of research, I collaborated with engineering to define clear requirements, removed features that didn’t support our primary use case, and launched the MVP for internal testing within 3 weeks total.
- Answer incoming calls
- Follow structured scripts
- Handle basic objections
- Capture lead information
- High-volume scenarios
- Low-stakes conversations
- Top-of-funnel interactions
- First-touch experiences
Research
One week analyzing 100M+ minutes of call data with the data team. Built the failure-point taxonomy that informed everything downstream.
Requirements
Worked with engineering to translate insights into a narrow, shippable scope. Cut anything that didn’t serve top-of-funnel calls.
Development
Built the agent with clear boundaries on what it would and wouldn’t handle. Structured scripts, basic objections, lead capture.
Dogfooding
Internal testing within three weeks. Every call monitored, every failure logged, every fix shipped same week.
Let the market tell us where to go
// VERTICAL-FIRST GTMWe ran structured surveys across five verticals, sent demo calls, and watched who leaned in. Hard Money Lending stood out immediately. They had compliance needs, heavy phone usage, and lean teams that couldn’t scale reps.
I helped design the survey, ran working sessions with sales and CS, and pushed for a vertical-first go-to-market instead of a generic pitch. We focused. Built objections that mattered to lenders. Added disclaimers. Tuned the tone.
Engagement multiplier vs SMB Sales baseline
Improved based on real usage
// FOUR WEEKS OF ITERATIONThe initial calls revealed significant challenges: awkward pauses, unnatural phrasing, and missed conversational cues. The team put in long hours to address these issues, driven by a shared commitment to shipping a working product. We monitored every call, identified failure points, and implemented fixes quickly.
I ran the feedback loop directly. No middle layers. PM to engineer to call review to fix.
Three metrics · paired endpoints
Made hard calls when it helped the product
// THE PIVOTOur initial approach attempted to serve five different verticals simultaneously, which proved ineffective. The scripts became generic, the model performance degraded, and the product lacked clear focus.
I advocated for a strategic pivot and worked to align our go-to-market team and leadership on this direction. We decided to concentrate exclusively on lending, redesigned our conversation flows based on actual lender calls, and trained the model with more targeted data. The improvement was immediately apparent.
- Generic scripts for 5 verticals
- Confused AI model
- Diluted value proposition
- Low customer engagement
- Specialized lending scripts
- Focused AI training data
- Clear value proposition
- High customer adoption
Poor results
Five verticals, generic scripts, degraded model performance. The data was honest about the cost of trying to do everything at once.
Stakeholder buy-in
Aligned GTM and leadership on a lending-only bet. Made the case with vertical engagement data, not opinion.
Immediate impact
Redesigned flows on real lender calls, retrained on focused data. Adoption rose, error rates fell, customers engaged.
Why it worked
// OUTCOMEWe transformed repetitive, error-prone first-touch calls into consistent, reliable interactions. By analyzing over 100 million minutes of call data, we identified a focused use case and built features that directly improved key performance metrics.
We shipped the MVP in 3 weeks and established a weekly iteration cycle based on actual call performance. Call completion improved from 60% to 85%, objection handling success increased from 10% to 40%, and sentiment detection accuracy improved by 3.2×. The results came from focused execution, data-driven decisions, and rapid feedback loops.
Closed alpha with select customers.
Built at Kixie
// CONTEXTPowered by 100M+ minutes of real sales conversations. Learn more at kixie.com.
Ship an AI MVP that lifts a metric.
One narrow use case. Real usage. Weekly improvement. If you have volume and a clear problem, I’ll help you scope, ship in ~3 weeks, and iterate from calls.