Skip to main content
Once a simulation run completes, Bluejay generates a detailed results report. This page explains the metrics, call-level data, and best practices for interpreting those results so you can turn raw data into actionable improvements.

Results Overview

Every completed simulation run produces an interactive performance dashboard. The dashboard aggregates data across all calls in the run and surfaces both high-level summary metrics and granular, per-call details.

Summary Metrics

The top of the results dashboard presents key performance indicators for the entire run:
  • Redundancy and Hallucination Checks
    • Flags instances where your agent repeated unnecessary information or generated inaccurate, unsupported responses. A clean check means the agent stayed concise and factually grounded.
  • Average Latency
    • The mean response time across all calls in the run. Lower latency generally correlates with a more natural conversational experience.
  • Goal Outcomes (Pass/Fail)
    • Shows the percentage of calls where your agent achieved the predefined goals. This is one of the most direct indicators of whether your agent is performing as intended.
  • Call Statuses
    • A breakdown of how individual calls concluded — for example, successful, escalated, or dropped. Use this to quickly spot patterns like a high escalation rate.
  • Agent Speaking Percentage
    • The proportion of conversation time your agent spoke relative to the digital human. A healthy balance depends on the scenario — support calls may skew toward the agent, while intake calls should let the customer speak more.

Call-Level Details

Below the summary, you can drill into each individual call for deeper analysis:
  • Transcripts
    • Full text records of each conversation. Use transcripts to trace exactly where the agent succeeded or went off-track.
  • Call Logs
    • Step-by-step logs of every interaction event, including tool calls, transfers, and hold times. Useful for diagnosing process-level issues.
  • Recordings
    • Audio playback of each call. Recordings capture tone, pacing, and naturalness that transcripts alone cannot convey.
  • Call-Specific Metadata
    • Contextual data points such as timestamps, detected customer emotion, response accuracy scores, and scenario-specific tags. These help you filter and compare calls within a run.

Custom Metrics

If you have custom metrics configured for your agent, their scores will appear alongside the built-in metrics for each call. Custom metrics let you measure domain-specific quality signals — such as compliance adherence, upsell success, or empathy — that the default metrics do not cover.

Interpreting Results

When reviewing simulation results, consider these guiding questions:
  • Are goals being met? Start with the pass/fail rate. If it is below your target, drill into the failing calls to identify common failure modes.
  • Is latency acceptable? High average latency may point to slow LLM responses, integration bottlenecks, or overly complex prompts.
  • Is the agent hallucinating? Even a small number of hallucination incidents warrants investigation — check the transcripts of flagged calls to understand what triggered the inaccurate output.
  • How do call statuses distribute? A spike in escalations or dropped calls can signal issues with your agent’s ability to handle certain scenarios end-to-end.
  • What do the recordings reveal? Quantitative metrics tell you what happened; recordings tell you how it felt. Listen to a sample of calls to catch issues like unnatural phrasing or awkward pauses.

Comparing Across Runs

One of the most powerful ways to use simulation results is to compare them over time. By running the same simulation repeatedly — especially after making changes to your agent’s prompt, tools, or configuration — you can:
  • Track whether goal pass rates trend upward.
  • Confirm that latency stays within acceptable bounds after updates.
  • Verify that fixes for hallucination or redundancy actually hold.

Best Practices

  • Review results promptly — Analyze results soon after a run completes so the context behind any changes is still fresh.
  • Focus on failing calls first — Start with calls that failed their goals or were flagged for hallucination, then work outward.
  • Use recordings for qualitative review — Numbers reveal trends, but listening to calls uncovers nuance.
  • Leverage custom metrics — Define metrics that match your business-critical success criteria so the dashboard highlights what matters most to you.
  • Track trends, not just snapshots — A single run is a data point; comparing runs over time shows whether your agent is genuinely improving.

Example Results Dashboard

Simulation Run: "Billing Support Stress Test #12"

Summary Metrics:
- Average Latency: 2.5 seconds
- Redundancy Checks: Passed
- Hallucination Checks: 0 incidents
- Goal Outcomes: 85% Pass rate
- Call Statuses: 12 Successful, 2 Escalated, 1 Dropped
- Agent Speaking Percentage: 55%

Individual Call Data:
- Call Transcript available
- Call Recording available
- Detailed call log with timestamps and agent/customer exchanges