All notes

What Contact Centre QA Actually Is

Three activities share the name: verifying compliance, improving performance, and understanding customers. They need different scorecards.

What Speech Analytics Actually Does

Transcription, then search and classification over the transcript. Everything vendors describe as understanding is pattern matching on text that is itself imperfect.

Why Two Percent Tells You Nothing

The standard practice of scoring a handful of calls per agent per month produces a number with enormous uncertainty, presented to two decimal places.

Anatomy of a Scorecard That Measures Something

Most scorecards accumulate items nobody removes, weight them arbitrarily, and produce a percentage that cannot be traced to any outcome. A shorter one works better.

Manual and Automated Evaluation: What Each Catches

They fail in opposite directions. Automation misses meaning and covers everything; humans understand meaning and see almost nothing.

QA, Coaching and Compliance Are Different Functions

Combining them into one process is the most common structural mistake in contact centre quality, and it makes the coaching adversarial and the compliance unreliable.

What This Data Cannot Tell You

A list of questions routinely asked of quality and speech analytics data that it cannot answer, with what to use for each.

Transcription Accuracy and What Degrades It

Vendor accuracy figures come from clean benchmark audio. Contact centre calls are not that, and the difference is large enough to change what the system can support.

The Accent and Dialect Accuracy Gap

Speech recognition performs measurably worse for some speakers, along lines that map onto race and region. That has consequences.

Categorisation and Topic Detection

The most useful thing speech analytics does, and the thing most often set up badly. Category definitions are the whole product and they need maintenance.

Sentiment Analysis: What It Measures

It classifies language as positive or negative. That is not the same as knowing how the customer felt, and the difference matters.

Emotion Detection and the Science Problem

Products claim to infer emotional state from voice. The scientific foundation is contested and regulators have begun restricting it.

Acoustic Measures: Silence, Talk-Over and Pace

The most reliable outputs in the whole toolkit, because they are measured rather than inferred, and the least discussed.

Redaction, PCI and Sensitive Data in Recordings

Recordings capture card numbers, health data and identifiers. Redaction is a control with known failure modes, and so is pause-and-resume.

Real-Time Analytics and Agent Assist

Prompting during a live call is a different product with different failure modes. When it helps, when it distracts, and the surveillance question it raises.

Language Models in Conversation Analysis

Vendors have replaced rule engines with language models. This genuinely improves some tasks, introduces new failure modes, and makes the system harder to audit.

Keyword Spotting and Full Transcription

Two architectures with different economics and a decisive practical difference: whether you can ask a new question of last year's calls.

Speaker Separation and Why Stereo Matters

Knowing who said what is the foundation of every agent-level measure. Mono recording makes it a guess, and a large number of deployments discover this late.

Defining Quality Before Measuring It

Most scorecards encode assumptions about what good looks like that nobody has tested. Deriving them from outcomes takes a quarter and changes what gets measured.

Calibration: Keeping Evaluators Consistent

Two evaluators scoring the same call disagree more than anyone expects. Measuring the disagreement is the only way to know whether your scores mean anything.

Automated Scoring: Where It Works

Scoring every interaction is the headline promise. It works for a narrow band of items and produces disputes everywhere else.

What Changes When You Analyse Every Interaction

Coverage removes the sampling problem and introduces a volume problem. What genuinely improves, and what organisations discover they were not ready for.

Targets, Gaming and Goodhart's Law

Any QA measure that becomes a target stops measuring what it did. The mechanisms are predictable and some are avoidable.

Rolling Out a QA or Analytics Programme

The technical deployment is the easy part. Agent trust and the first month's findings determine whether the programme survives a year.

Multi-Language and Multi-Site Programmes

Analytics quality varies enormously by language, which means a global programme measures some sites better than others and reports the difference as performance.

Chat, Email and Messaging Quality

Text channels remove the transcription problem entirely and introduce their own. Most programmes apply a voice scorecard to them, which measures the wrong things.

Delivering Feedback That Changes Behaviour

The evaluation is worthless if the conversation produces no change. What the research suggests, and what practice usually does instead.

Disputes and Appeals

A dispute route is not an administrative burden. It is the mechanism that finds broken scorecard items, inconsistent evaluators and transcription failures.

Getting Process Findings Out of QA Data

Much of what QA discovers is not an agent problem. Routing those findings to people who can fix them is the highest-value output.

Evaluator Workload and Score Quality

Score quality degrades measurably with volume and session length. Most programmes set evaluator targets without knowing where that threshold sits.

QA in Outsourced and BPO Environments

When the agents work for someone else, QA becomes a contractual instrument as well as a quality one, and the incentives point in unhelpful directions.

The QA Workflow, End to End

From selecting an interaction to a closed action. Most programmes have the first three steps and stop, which is why the output does not change anything.

What QA Does to the People Being Measured

Contact centre attrition is high and quality programmes contribute to it. The design choices that make monitoring tolerable are known and rarely applied.

The Measures Worth Reporting

A short set that carries information, and the longer set that fills dashboards and drives the wrong behaviour.

Testing Whether QA Scores Predict Anything

The analysis that validates or invalidates an entire quality programme, and it can be run in a week with data most operations already have.

First Contact Resolution, Honestly

The most quoted contact centre metric and the most inconsistently defined. Most published FCR figures are not measuring resolution.

Survey Scores and What They Actually Measure

Response rates are low and non-random, so a survey average describes the people who answered. What that supports and what it does not.

Detecting Scorecard and Model Drift

Scores move for reasons unrelated to performance: evaluator turnover, model updates, call mix. Instrumenting distinguishes them.

Reporting Quality to Leadership

A single quality percentage is what gets asked for and it is the least informative thing you can supply. What to report instead, and how to survive the request.

Call Recording and Consent

Recording rules differ by jurisdiction, by party and by purpose. A single global recording policy will be wrong somewhere, and the exposure is real.

Voiceprints and Biometric Law

Voice authentication creates a biometric identifier, which is regulated more strictly than a recording. Several statutes carry private rights of action.

Monitoring Agents: The Employment Dimension

Continuous analysis of employee speech is employee monitoring, which in many jurisdictions carries notification, consultation and proportionality obligations.

Retention: Recordings, Transcripts and Indexes

Deleting a recording deletes one copy. The transcript, the analytics index, the summary and the backup are separate objects and they are usually forgotten.

Bias in Automated Quality Scoring

Automated scoring inherits the speech recognition accuracy gap, so it can be systematically harsher toward some agents. Measure it.

What to Tell Agents

Agents will find out what the system does. Telling them first is a legal requirement in several jurisdictions and an operational advantage everywhere.

Vendor Due Diligence on Your Recordings

Your recordings leave your estate. Where they go, who hears them, how long they stay and whether they train models are contract questions.

Evaluating a Speech Analytics Platform

Demonstrations run on clean audio with tuned categories. The questions and tests that reveal what the product does on your calls.

Building Speech Analytics Rather Than Buying

Open transcription models have made the technical build tractable. Whether to do it turns on the operational layer, which is most of the product.

Starting a Programme From Nothing

The first ninety days, in order. Almost none of it requires a purchase, and the sequence determines whether the programme is trusted or resisted.

Glossary

Terms used across these notes, defined once, including the several that are used inconsistently across the industry.