Core Concepts

Before diving into setup or daily use, it helps to understand the key ideas behind Agent QA.

1.1 How Agent QA Works

Agent QA follows a multi-stage pipeline:

  1. Audio uploaded via the ingestion API
  2. Transcribed -- speech-to-text with PII masking and speaker diarization
  3. GWE workflow evaluates the transcript against your QA parameters
  4. Callback processes results -- extracts rule outcomes, contact reasons, sentiment, and insights
  5. QA score computed -- weighted average with critical-rule auto-fail
  6. Score and Scorecard saved and visible on the Audit tab

Every call goes through this pipeline automatically. The entire flow -- from audio upload to a completed scorecard -- typically takes 2 to 5 minutes.

1.2 Key Terms

TermWhat It Means
AI EmployeeA configured workflow on the Ema platform. Agent QA is a type of AI Employee. You may also see this called a "persona" in technical contexts.
Parameter (also called a "rule")A single quality criterion used to evaluate agent performance. Examples: "Greeting," "Empathy," "Issue Resolution." Each parameter has pass/fail criteria and a weight.
ScorecardThe detailed breakdown of an interaction's evaluation -- one result per parameter, with rationale and transcript evidence.
QA ScoreA number from 0 to 100 representing overall quality. Calculated from weighted parameter results.
Auto-LearningThe system that generates and refines your QA parameter instructions using a set of human-graded example calls (your "golden dataset").
Continuous LearningThe ongoing process of improving evaluation accuracy through feedback submitted by your review team.
Golden DatasetA collection of call recordings with verified human evaluations (Pass/Fail/N/A per parameter) used to calibrate Auto-Learning. Also referred to as "label data" in the setup interface.
Critical ParameterA parameter where failure automatically sets the entire interaction score to 0, regardless of other results. Used for non-negotiable standards like regulatory compliance.
N/A (Not Applicable)When a parameter does not apply to a particular call. N/A results are excluded from scoring entirely -- they do not help or hurt the score.
Contact ReasonA three-level categorization of why customers are calling: L1 (broad category), L2 (subcategory), L3/Call Driver (specific reason).
UCIDUnique Call Identifier -- the unique ID for each call recording.
VerbatimA direct quote from the transcript used as evidence for an evaluation decision.

1.3 Evaluation Mechanisms

Each parameter uses one of three mechanisms that determine what data is consulted during evaluation:

MechanismWhat It ChecksWhen to Use
Transcript OnlyThe call transcript aloneGreeting, empathy, communication skills -- anything observable from the conversation
Knowledge VerifiedTranscript + your knowledge base documentsInformation accuracy, policy compliance -- anything requiring verification against reference materials
Action BasedTranscript + actions in external systems (CRM, ticketing)Case creation, ticket updates -- anything requiring verification of what the agent did in their tools

1.4 The Daily QA Loop

Agent QA delivers the most value when used as a continuous cycle:

  1. Audit -- review interactions
  2. Metrics -- spot trends
  3. Insights -- identify coaching priorities
  4. Feedback -- correct evaluations
  5. Improve -- refine rules
  6. Back to Audit

Each stage feeds into the next. Reviewing interactions surfaces evaluation disagreements. Metrics reveal patterns. Insights pinpoint coaching priorities. Feedback corrects individual evaluations. Improvements refine the underlying rules.

1.5 Languages Supported

Agent QA supports transcription and evaluation for a wide range of languages including: Arabic, Belarusian, Bengali, Bosnian, Bulgarian, Catalan, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, Flemish, French, German, Greek, Hebrew, Hindi, Hungarian, Indonesian, Italian, Japanese, Kannada, Korean, Latvian, Lithuanian, Macedonian, Malay, Marathi, Norwegian, Persian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swedish, Tagalog, Tamil, Telugu, Turkish, Ukrainian, Urdu, and Vietnamese.

Last updated: Jul 3, 2026