From Prototype to Production: How Perk Built a Voice AI Agent That Makes 10,000 Calls a Week
When hotels failed to record virtual credit card payments, Perk's customer care team was making 10,000 proactive verification calls weekly. Discover how their engineering team transformed this manual process into an AI-powered solution that improved success rates while reducing workload. Learn their prototyping approach using Make.com and the key insights that helped them scale to production.
The $1M Problem Hidden in Hotel Payments
Perk's business model revolutionized corporate travel by eliminating expense reports - travelers book through their platform, and hotels get paid directly via virtual credit cards. But when hotels occasionally failed to record these payments, it created catastrophic guest experiences. Imagine arriving after a 14-hour international flight to discover your room wasn't paid for and might be unavailable.
Their solution? A customer care team making 10,000 proactive verification calls weekly to hotels worldwide. While effective, this manual process was unsustainable. "We were leaving hundreds of calls unmade each day due to volume constraints," explains Steve, Perk's Product Manager.
Key Insight: The team realized voice AI could handle the repetitive verification calls while freeing human agents for complex cases. Their challenge was creating an agent that matched human success rates in natural conversations across languages.
From Hackathon to Functional Prototype in 1 Week
The breakthrough came during an internal AI workshop that turned into a hackathon. Using Make.com, the team connected their Zendesk ticketing system to Twilio's telephony API and 11 Labs' voice generation.
"We avoided backend changes by parsing hotel data from existing support tickets," explains Gabby, Senior Engineering Manager. This let them test the concept without engineering resources. Their first prototype could:
- Extract hotel phone numbers and guest details from tickets
- Initiate calls via Twilio
- Handle basic conversation flows
- Record outcomes back to their system
Initial tests revealed unexpected challenges - from IVR systems (45% of failures) to hotels asking if they were speaking with AI. One receptionist even said, "I actually prefer talking to an AI!"
Breaking Down the Perfect Hotel Call
The team discovered that successful calls followed a predictable structure they could engineer into the AI agent:
- IVR Navigation: Handling "Press 1 for front desk" systems
- Booking Confirmation: Verifying correct guest and reservation
- Payment Verification: Ensuring virtual card was charged
- Graceful Exit: Ending calls politely after completion
Critical Lesson: A single monolithic prompt failed unpredictably. Breaking conversations into phases with separate specialized prompts increased reliability by 62% in testing.
They also implemented text-to-speech markdown for proper number pronunciation (saying "five" instead of "5") and added pauses for natural cadence - small details that significantly improved comprehension.
The Prompt Engineering Breakthrough
Early prompts gave the AI too much freedom, leading to verbose, confusing conversations. The team learned to:
- Define narrow conversation boundaries
- Provide explicit off-ramps for dead ends
- Limit response length
- Structure information delivery
Philip, Senior Software Engineer, notes: "We initially gave the AI an 'experienced travel agent' role - big mistake. It started trying to modify bookings! Now we specify exact scripts with minimal variation allowed."
Their evaluation system caught regressions - like when a prompt change caused repetitive phrasing that annoyed hotel staff. Continuous monitoring ensures quality as they expand to new languages.
Building an Evaluation System That Scales
With thousands of weekly calls, manual monitoring became impossible. Their solution was a two-layer evaluation system:
1. Success Classification: LLMs analyze transcripts to determine if hotels agreed to charge the virtual card, with strict criteria to avoid false positives that could leave travelers stranded.
2. Behavioral Quality: Separate evaluations measure politeness, clarity, and conversation flow across 8 dimensions. This catches issues like abrupt endings or excessive repetition.
The team still manually reviews a sample of calls, using findings to refine their evaluation criteria and prompt engineering. "It's an ongoing process," Gabby explains. "Every new failure mode teaches us something."
From 5 Calls/Day to 10,000 Calls/Week
Perk's rollout strategy ensured reliability at scale:
- Shadow Testing: AI made calls in parallel with humans to compare outcomes
- Phased Volume Increase: Starting with 5 calls/day, then 50, then 500
- Market Expansion: First English, then German, with Spanish coming next
Steve recalls: "We'd sit in a room watching call transcripts appear in real-time during early tests. Now it handles more calls before lunch than our whole team could do in a week."
The system's success spawned requests from other departments facing high-call-volume challenges - unexpected validation of their approach.
Results That Surpassed Human Performance
Today, Perk's voice AI agent:
- Makes 10,000+ calls weekly across multiple languages
- Achieves better success rates than human agents
- Reduced instances of unpaid rooms at check-in by 73%
- Provides detailed analytics on payment failure reasons
Perhaps most importantly, it solved an impossible scaling problem. "We could never have manually made all the verification calls needed globally," Steve notes. "The AI lets us be proactive at scale."
Key Takeaway: By starting small with Make.com and relentlessly focusing on measurable outcomes, Perk transformed a prototype into a production system that handles more calls with better consistency than humans could achieve.
Watch the Full Case Study
See the complete interview with Perk's engineering team discussing their prototyping challenges and key breakthroughs (jump to 12:30 for the Make.com workflow details).
Frequently Asked Questions
Common questions about voice AI automation
Perk faced a critical issue where hotels sometimes failed to record virtual credit card payments for business travelers. This created scenarios where exhausted travelers arriving at hotels might find their rooms unavailable despite having booked through Perk's system.
Their manual solution involved making 10,000 proactive verification calls per week through customer care teams. While effective, this process created significant operational strain and couldn't scale to cover all necessary verifications globally.
- Problem affected 5-7% of hotel bookings
- Manual calls took 3-5 minutes each
- Customer care team couldn't keep up with demand
The team used Make.com to create an initial prototype that integrated with their existing systems without requiring backend changes. This allowed them to test the concept rapidly using available tools.
Their prototype connected to Zendesk to pull hotel information, used Twilio for calling infrastructure, and experimented with 11 Labs for voice generation. This approach let them validate the solution in just one week without committing engineering resources.
- Built functional prototype in 7 days
- Used existing ticket data for hotel information
- Tested with real hotels within 2 weeks
Major challenges included handling IVR systems, natural number pronunciation, conversation flow control, and multilingual support. IVR systems alone accounted for 45% of initial failures.
The team solved these by breaking conversations into structured phases (IVR, booking confirmation, payment request) with specialized prompts for each stage. They also implemented rigorous evaluation systems to monitor performance across languages.
- IVR navigation was biggest initial hurdle
- Number pronunciation required special formatting
- Multilingual support needed native-language prompts
They track two key metrics: success rate (whether hotels confirm they'll charge the virtual card) and behavioral quality (politeness, clarity). Success classification uses LLMs to analyze call transcripts against predefined criteria.
For behavior, they evaluate 8 dimensions including repetition frequency, abrupt endings, and proper introductions. This dual approach ensures both functional success and positive user experiences.
- Daily monitoring of 10+ quality metrics
- Manual review of 5% of calls for calibration
- Automated alerts for performance deviations
The system now handles over 10,000 calls weekly across multiple languages with better success rates than human agents achieved. More importantly, it reduced instances of travelers arriving at hotels with unpaid rooms.
By being proactive at scale - something impossible with manual processes due to volume constraints - the AI solution improved both operational efficiency and customer experience simultaneously.
- 73% reduction in unpaid room incidents
- 10,000+ calls handled weekly
- Higher success rate than human agents
While starting with Make.com, Perk migrated to a custom solution using Twilio for telephony infrastructure and LLMs for conversation handling and classification. The system integrates with their existing CRM and support ticket platforms.
The architecture separates conversation phases into distinct modules with evaluation systems monitoring each component's performance daily. This modular approach allows continuous improvement of individual conversation segments.
- Twilio for telephony infrastructure
- LLMs for conversation handling
- Custom evaluation and monitoring systems
The team implemented a rigorous four-phase validation process: internal testing with roleplaying, shadow testing comparing AI vs human outcomes, gradual production scaling with manual review, and full deployment only after matching human performance.
This careful approach ensured reliability before scaling, catching issues like overly verbose responses or IVR navigation problems early when they were easier to correct.
- 4-phase validation process
- Started with 5 calls/day
- Manual review of early calls
GrowwStacks specializes in building custom voice AI solutions like Perk's for businesses needing high-volume, natural conversation automation. We combine telephony expertise with advanced prompt engineering and evaluation frameworks.
Our team can design, implement and optimize complete voice agent systems tailored to your specific use case - whether for customer support, verification calls, appointment scheduling, or other conversation-intensive processes.
- End-to-end voice AI implementation
- Custom conversation design
- Performance monitoring systems
Automate Your High-Volume Calls With AI
Manual call processes create bottlenecks and limit scalability. GrowwStacks can design a custom voice AI solution that handles your call volume with consistent quality.