How Lindy is shipping reliable Voice AI Employee with automated quality engineering

The Challenge

Building a voice agent that feels human requires more than just clear audio. Lindy's agents handle different business processes - such as refund eligibility and appointment scheduling where missing a specific step in a workflow results in a failed customer experience.

The team faced several hurdles:

The "Interrupter" Problem

Real customers don't wait for an AI to finish its sentence. Lindy needed to ensure their agents could handle "Interrupter" personalities without losing context or stuttering, requiring a pause-and-resume capability that feels natural.

Workflow Regression

Every time a prompt is updated to improve a specific flow, there is a risk of breaking other flow. Lindy needed a way to verify that agents consistently met "Expected Outcomes" (e.g., AI Agent provided the correct 5-7 day refund window) across thousands of iterations.

Latency and Conversational Flow

To maintain a natural cadence, the team needed to monitor average latency and ensure a balanced "Talk Ratio." If an agent speaks too much or responds too slowly, the human-like "AI Employee" experience is broken.

The Solution

Lindy integrated Cekura into their development workflow to move from reactive testing to a rigorous, data-driven framework. Cekura provides the Lindy team with:

Infrastructure Testing

Lindy utilizes Cekura's "Interrupter" and "Normal Male" personalities to simulate varied human behaviors. They specifically monitor "Stop Time After User Interruption" to track how long agent takes after interruption.

Workflow Testing

Using Cekura's "Expected Outcome" metric, Lindy automatically verifies if the agent completed all necessary steps of a scenario. For example, in a refund inquiry, Cekura flags if the agent failed to confirm the delivery address or the specific processing timeline.

Latency & WPM Benchmarking

The team tracks Words Per Minute (WPM) and Latency in milliseconds to ensure agents are efficient. By keeping agents under 200 WPM range with a Talk Ratio of under 0.8, Lindy ensures their agents are helpful without being overbearing.

"With Cekura, we can stress-test our agents against different personalities and complex customer journeys with total confidence. It has turned our quality assurance from a manual bottleneck into a competitive advantage."

— Flo Crivello, Founder & CEO at Lindy

The Results

Lindy's voice agents now deliver a high quality experience that rivals human performance.

By leveraging Cekura, they have moved to a proactive quality engineering model:

• 100% Outcome Verification in Testing: Cekura successfully flags any call where a specific business rule is missed (e.g., failing to inform a customer they are eligible for a refund). This allows Lindy to tune prompts for specific scenarios like "Past Order Details Confirmation" with total precision.
• Superior Interruption Handling: Through rigorous testing of the "Interrupter" personality, Lindy optimized their "Stop Time" to under a second in many cases, ensuring the AI never "talks over" the customer.
• Security & Protocol Adherence: Simulations proved that Lindy agents successfully handle malicious calls - refusing to share system prompts or owner emails, while maintaining a professional, helpful tone.