Quality teams traditionally listen to a very small fraction of calls. A few conversations per agent per month, one or two per cent of total volume. An assessment built on that sample says more about which calls were picked that month than about the agent's actual performance.
Automated analysis removes that constraint. Every conversation can be transcribed, labelled and scored. But full coverage on its own does not amount to better quality management. As coverage expands, it becomes necessary to separate what is reliably measurable from what is contested.
Everything is capped by transcript quality
Every downstream output rests on the transcript. If the transcript is wrong, the summary, the labels and the score are wrong too, and the error propagates silently.
A few concrete factors decide transcript quality on call recordings:
Audio bandwidth. Telephone audio is narrowband and carries considerably less information than a studio recording. The performance a general-purpose speech model shows on clean audio will not hold on a phone call. Benchmark with your own recordings, not with vendor demos.
Organisation-specific terms. Product names, campaign names, branch names, tariff codes and alphanumeric references are where general models struggle most. Introducing a lexicon of those terms delivers far more improvement than changing models. In most organisations a list of one to two hundred terms can be assembled in a day.
Speaker separation. Distinguishing customer from agent is the foundation of any quality score. A "prohibited phrase detected" signal is worthless if you do not know who said it. The most effective solution here is operational rather than technical: if the recording system can capture dual-channel audio, agent and customer are written to separate channels and the diarisation problem largely disappears. Single-channel recordings require algorithmic separation, and error rates climb wherever speech overlaps.
The first question to ask when starting a call analytics project, before choosing a model, is whether the recordings are mono or dual-channel.
Which parts of the quality form can be automated
Quality forms contain two kinds of item, and trying to automate them without separating the two usually fails.
Objective items are those whose presence or absence can be verified from the text:
- Opening and closing announcements delivered
- Identity verification steps completed
- Mandatory disclosures read out
- No prohibited or high-risk phrasing used
- The customer's request summarised back for confirmation
These suit automated checking, and full coverage delivers real gains here. Knowing what percentage of calls skipped a mandatory disclosure from a complete census rather than a sample is a direct compliance benefit.
Subjective items are different: empathy, appropriateness of the resolution, how well the customer was guided. Automated scores can be produced for these, but they should not be used as the decision on their own. Treat the automated score as a prioritisation tool: the system flags which calls a human should listen to, and the human decides.
An uncalibrated score gets abandoned
When automated quality scoring goes live, it has to be compared against human scoring for the first several months. The same calls are assessed by both the system and a quality analyst, and the gap is measured.
Skip that and the sequence is predictable: agents do not trust the scores, they contest them, managers cannot defend them, and within a few months the metric appears in reports that nobody reads.
Alongside calibration you need an appeals mechanism. Agents must be able to challenge a score, the challenge must be reviewed by a person, and the outcome recorded. For the legitimacy of performance management this matters as much as technical accuracy.
Two separate warnings about emotion analysis
Emotion scoring is the most talked-about feature in call analytics. Two different kinds of caution apply: one scientific, one legal.
The scientific side: expression to emotion is not a clean inference
Emotion recognition systems rest on the assumption that particular expressions correspond to particular emotions. That assumption has been challenged extensively in the psychology literature. The comprehensive 2019 review by Barrett and colleagues found that inferring emotion from facial movements is far less reliable than commonly assumed, with the same emotion appearing through different expressions and the same expression arising from different emotions (Barrett et al., Psychological Science in the Public Interest, 2019).
That study concerns facial movements. The core of the critique carries over to voice: inferring emotion from tone independent of context does not form a sound inference chain. Someone speaking loudly may be angry, may have a poor connection, may be in a noisy environment, or may simply speak that way.
A more defensible approach in practice is to use behavioural and content-derived signals instead of "emotion":
- How many times the customer repeats the same request
- How often speech is interrupted within the call
- Occurrence of terms like complaint, cancellation or legal action
- Calls closing without resolution
- A repeat call within a short window
These are both more measurable and easier to explain to a manager. "The customer repeated the same thing three times and the word cancellation came up" produces far more action than "emotion score 0.31".
The legal side: employee emotion is a separate category
This is critical and frequently missed. The EU AI Act prohibits placing on the market or using AI systems to infer emotions of natural persons in the workplace and in education institutions, with an exception only for medical or safety purposes. The prohibitions have applied since 2 February 2025 and breaches sit in the most severe penalty band (analysis of Article 5(1)(f), Future of Privacy Forum).
For a contact centre the implication is that in an EU operation, a system that analyses the agent's emotion falls within the prohibition. Analysing the customer's emotion does not fall under that prohibition, but emotion recognition systems are separately regulated in the Act and people exposed to them must be informed.
For organisations operating only in Turkey with no EU customers, the Act does not apply directly. Under Turkish data protection law, however, proportionality and purpose limitation still govern employee monitoring. It is difficult to argue that inferring emotion from an agent's tone of voice is proportionate to the requirements of the job.
In short: collect signals on the customer side, measure behaviour and process compliance on the agent side, and do not make emotion inference an input to employee assessment.
Connecting analysis to business outcomes
Quality scores have no value on their own. The value is in linking the score to what happens next.
The connections worth building: which topic labels lead to repeat calls, which process steps extend handling time, which missing disclosure turns into a complaint later, which product questions give agents the most difficulty.
With those in place, call analytics stops being a quality tool and becomes an input to process and product improvement. The most valuable output of the recordings is usually not agent scores; it is the topic list that goes to the product and operations teams.
Deployment checklist
- Are recordings mono or dual-channel, and how will speaker separation be handled
- Has an organisation-specific term lexicon been prepared
- Have objective and subjective items on the quality form been separated
- Is the automated score regularly calibrated against human scoring
- Is there an appeals mechanism for agents
- Is emotion inference being used in employee assessment (it should not be)
- Is there an EU operation in scope, and has it been legally assessed
- Do analysis outputs flow regularly to product and process teams
To see how the analysis services connect over API, look at Call Analytics Services and, for free-text sources, Review Analysis Services.
This article is general information and is not legal advice.
References
- Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., Pollak, S. D. (2019). Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements. Psychological Science in the Public Interest, 20(1). journals.sagepub.com
- EU AI Act, Article 5: Prohibited AI Practices. artificialintelligenceact.eu
- Future of Privacy Forum, Red Lines under EU AI Act: emotion recognition in the workplace. fpf.org
- Turkish Data Protection Authority. kvkk.gov.tr