Call analytics glossary

Call analytics

Also called: call tracking, call reporting

The practice of capturing data from phone conversations and turning it into something a business can act on.

The term covers two quite different things, which is the source of most confusion. The older meaning is metadata: how many calls, how long, how many abandoned, which marketing campaign produced them. The newer meaning is analysis of what was actually said inside the call. A vendor can truthfully claim "call analytics" while only doing the first, so it is worth asking which one you are being sold.

Metadata analytics tells you a customer rang three times this fortnight. Content analytics tells you it was the same unresolved booking each time. Only the second is actionable without someone listening to the recordings.

Conversation intelligence

Software that analyses the content of recorded conversations to surface coaching points, risks or commercial signals.

Most conversation intelligence products were built for sales teams: did the rep mention the competitor, ask discovery questions, talk more than the prospect. The underlying technology transfers to customer service, but the questions being asked do not. A sales-tuned tool scores technique. A service-tuned tool needs to answer a different question — did the business keep its promise to this person?

If you are evaluating a conversation intelligence product for service quality, check what it was trained and configured to look for. Talk-ratio and monologue-length metrics are strong signals in a discovery call and nearly meaningless in a customer complaint.

Speech analytics

Also called: keyword spotting, phrase detection

Analysis applied to call audio or its transcript, historically keyword and acoustic based, to detect defined words, phrases or vocal characteristics at scale.

Classic speech analytics works from a list you supply: flag any call containing "cancel", "complaint", "ombudsman", "speak to your manager". It is fast, cheap and completely literal. It finds the words you thought of, and nothing else. A customer who says "look, I might just go elsewhere next year" has told you they are leaving without using a single one of your keywords.

Some speech analytics also measures acoustic properties — volume, pitch, talk-over, silence. That detects a raised voice. It does not detect a quiet, courteous conversation in which something went badly wrong.

Sentiment analysis

Also called: sentiment scoring, tone analysis

Scoring how positive or negative a conversation sounds — a measure of tone and emotional intensity, not of outcome.

Sentiment is the most widely deployed and most widely misunderstood signal in call monitoring. It answers "did this feel bad?" and is routinely bought to answer "did this go bad?" — which is a different question with a different answer more often than most buyers expect.

The gap is measurable. Across a 30-day review of 2,349 calls at an Australian corporate travel agency, the calls that genuinely warranted a manager's attention had a median sentiment of 82 out of 100 — positive. Only three of forty-five scored 40 or below. A capable agent stays calm and warm while the business fails the customer, and every signal sentiment measures says the call is going fine.

The inverse costs money too: a customer swearing cheerfully about an airline strike, being helped brilliantly, scores badly and generates an alert nobody needed. The full evidence is here →

Call QA sampling

Also called: quality assurance, call monitoring, call calibration

The traditional method in which a supervisor reviews a small sample of calls against a scorecard.

Typical coverage is one to two per cent of calls. The review itself is usually good — a human hears nuance no model does. The problem is arithmetic: if you review 2% of calls, you see 2% of the problems, and which 2% is decided by who had time on Thursday.

Sampling also drifts. Two reviewers grade the same call differently; the same reviewer grades differently in the last hour of the week. That is not a criticism of the people, it is what sampling by humans is. Automated review does not make human QA redundant — it tells the human which calls are worth their attention.

Call scoring

Assigning a numeric rating to a call against defined criteria, manually or automatically.

Scores are useful for tracking a team over time and poor at triggering action. A score of 62 does not tell a manager what to do on Monday. It also invites an unhelpful conversation with the agent about the number rather than about the customer.

The alternative is a reason: this call is being raised because a callback was promised on 12 August and never made, and the customer has now rung three times. That is actionable in thirty seconds and cannot be argued with.

Speaker diarisation

Also spelled: diarization

Separating a recording into who spoke when, so a transcript reads as a labelled conversation rather than one block of text.

Diarisation matters more than it sounds. Without it, "I'll call you back this afternoon" is a floating sentence — you cannot tell whether the customer said it or your agent promised it, and those mean opposite things. Any analysis of accountability depends on knowing who owned the commitment.

Quality varies with how the audio was captured. A stereo recording that keeps each party on its own channel diarises reliably; a mono mix has to be separated by inference and can collapse to a single speaker on calls where one party dominates.

Escalation detection

Identifying calls where a customer is asking for a manager, threatening to leave, or signalling that the matter has outgrown the person handling it.

The naive implementation is keyword matching on "manager" and "supervisor". In practice most Australian customers escalate without ever using those words — "is there someone else I can talk to", "how far up does this go", or simply going quiet and asking how to close the account.

Useful escalation detection also has to distinguish a customer escalating at your business from one venting about a third party while your agent helps them. Only the first needs a manager.

Repeat contact rate

Related: first contact resolution (FCR)

The share of customer matters that need more than one contact to resolve.

This is one of the few call metrics that is a direct measure of failure regardless of tone. If someone rings three times about the same booking, the number of calls tells the story on its own — each individual call may have been handled pleasantly.

Measuring it properly requires linking calls to a matter rather than to a phone number or an agent, because the customer may ring from a different number and reach a different person each time. Grouping repeat contacts on one issue is also what stops a single unresolved problem generating five separate alerts.

PII redaction

Also called: data masking, sensitive data scrubbing

Automatically removing personally identifying or payment data from call records.

Ask precisely what is redacted. Redaction applied to a transcript removes the card number from the text; the spoken digits still exist in the original audio file, exactly as they do in your phone system today. Audio redaction is a separate and much harder capability, and a vendor claiming "card numbers are removed" without saying which layer they mean is worth pressing.

For most businesses the transcript is the artefact that gets copied into emails, tickets and dashboards, so redacting it removes the majority of the practical exposure — but it is not the same as the recording being clean. How WiseSentry handles this →