What Is AI Call Centre Analytics and Sentiment Analysis?

Last updated: 9 October 2026. First published 8 March 2025 as "AI Speech Analytics", and rewritten in full. Sources were read on 9 October 2026 and are linked inline.

AI call centre analytics turns recorded calls into structured data (transcripts, topics, extracted details, quality scores) so every call can be measured. Sentiment analysis estimates how the caller felt, from their words, voice, or both.

Two things matter before you buy. Emotion detection on real calls is far less accurate than marketing suggests: 45.6% across four emotions in a peer-reviewed study of real emergency calls. And in the EU, tracking call-centre employees' emotions from their voice is prohibited.

AI speech analytics turns every call into data you can count

AI speech analytics, also called call centre or conversation analytics, applies speech recognition and language models to call audio. Speech analytics focuses on what was said, voice analytics on how it was said (pitch, pace, pauses), and sentiment analysis on an estimate of positive, neutral or negative feeling.

The same technology is used in two very different settings:

• Analysing a human team: software listens to calls between customers and human agents, to coach the agents, check compliance and spot recurring problems.

• An AI agent analysing its own calls: an AI voice agent also writes a summary, tags and a score for each call it handled. Here the system grades its own work.

How it works: from a recording to a number

Most products follow the same sequence:

1. Capture the audio. Calls are recorded, or the audio is streamed for live analysis.

2. Transcribe. Speech recognition turns audio into text.

3. Separate the speakers, so a cheerful agent doesn't mask an angry caller.

4. Classify, extract and summarise: tag the topic, pull out details such as a booking date, and write a summary.

5. Score each call for sentiment and quality.

6. Aggregate and alert: dashboards show trends, and rules flag calls for review.

Every step depends on step 2: a transcript that drops "not" flips the sentiment, and errors are uneven. Across five commercial speech recognition systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers (Koenecke et al., PNAS, 23 Mar 2020; US English, 2019 systems, recorded interviews rather than phone calls). For how transcription accuracy is measured and priced, see AI call transcription.

The components, and the question each one answers

Brilo's documentation draws a useful line between two of these. A summary "is prose, for a person to read", while extraction pulls out "named values ... so you can filter, count, and send them to another system". It adds: "You cannot filter prose" (Personalised call summaries, read 9 Oct 2026).

Sentiment analysis reads words, voice, or both

• Lexical (text) sentiment reads the transcript: "this is going in circles" scores negative. It is only as good as the transcript, and misses sarcasm or a flat, tired "fine".

• Acoustic sentiment, often sold as emotion detection, reads the audio: pitch, energy, tempo and pauses. It can catch the clipped "fine", but is harder on compressed phone audio and treats the voice as personal data.

Most products combine the two.

The split is also legal. The European Commission's guidelines say the AI Act's "emotion recognition" covers only inference "based on biometric data", and they name "characteristics of a person's voice" as an example. Inferring emotions "from written text (content/sentiment analyses)" falls outside the prohibition (Commission guidelines on prohibited AI practices, 4 Feb 2025, paras 250–251).

How accurate emotion detection is on real calls

We found no independently verified accuracy figures for commercial emotion detection. The clearest public evidence is research on real call recordings.

Deschamps-Berger, Lamel and Devillers tested a speech emotion recognition model on CEMO, a corpus of real French emergency-call recordings: 440 dialogues from 485 speakers (ACII 2021, arXiv 2110.14957).

• On IEMOCAP, an acted-emotion research dataset: 63% unweighted accuracy across four emotions.

• On the real calls: 45.6% across four emotions (anger, fear, positive, neutral).

• Anger versus neutral only: 76.9% on real calls, against 81.1% on acted data.

The authors' conclusion: "Real-life emotions are clearly more complex than acted ones."

Caveats: emergency calls, French, audio only, 2021 models, and almost 80% of segments were neutral. A service line differs, but the direction holds: models do worse on real callers than on acted data, and "is this caller angry?" is far easier than "which emotion is this?".

What sentiment measurement showed in a field study

The strongest published use of sentiment analysis in customer service is as a research yardstick. In the 2023 NBER working paper version of a study of 5,179 support agents at one firm, researchers scored chats with SiEBERT, a sentiment model, from −1 (negative) to 1 (positive) (Brynjolfsson, Li and Raymond, NBER working paper w31161, revised November 2023). All figures below are from that working paper version.

• Average customer sentiment: 0.14 (mildly positive).

• Average agent sentiment: 0.89. Agents were "unfailingly positive".

• After agents got a generative AI assistant, customer sentiment rose by 0.18 points, about half a standard deviation. Agent sentiment barely moved (+0.02).

Caveats: text chat, not voice; the AI suggested replies to human agents rather than taking calls; one company's 2023 tooling.

Sentiment worked as an aggregate measure across thousands of conversations. That says nothing about the accuracy of any single call's score. We found no independent figures for other claimed benefits, such as CSAT gains or cost savings, so none are quoted here.

See what post-call analysis shows on your own calls. Brilo's Free plan includes 10 minutes a month. See Brilo pricing.

Analytics for a human team vs an AI agent's own call analysis

| | Speech analytics on a human team | Built-in analysis on an AI voice agent | Post-call surveys | |---|---|---| | What is analysed | Calls between customers and your human agents | Calls the AI agent itself handled | The customers who reply | | Coverage | Every recorded call | Every call the agent takes, apart from those too short to score | A self-selected minority | | Independence | Separate from the agent being judged | The same system scores its own work | The customer's own view | | Best use | Coaching, compliance, root causes | Sorting calls for review and finding gaps | The customer's own rating |

A built-in score is the system marking its own homework: useful for triage (listen to the worst calls first), weak as evidence the deployment works. Our guide to monitoring an AI voice agent covers independent checks. For QA programmes on human teams, see AI call centre quality assurance. Live sentiment cues for human agents belong to AI agent assist.

What Brilo measures on every call

Brilo is an AI voice agent, not a speech-analytics platform for human teams. It analyses its own agents' calls (sources read 9 Oct 2026).

Pulse treats Resolution as "the most important factor". A smooth transfer to a person counts as good resolution.

Where Brilo's analysis stops

Brilo's documentation states these limits:

• The scoring rules are fixed. Pulse criteria and weights cannot be customised yet ("on our roadmap").

• Voice only. Pulse scores voice calls. Chat, SMS and email are "coming soon".

• Some calls are not scored, such as very short calls.

• No live sentiment is documented. Every analysis feature above runs after the call ends.

• No analysis of human-agent calls.

• Recording and retention are per agent. With recording off, the summary and extracted details are still kept. The Retention Period setting "is not enabled on every account" (Recordings, transcripts and retention).

Not stated in the documentation: how accurate Pulse is or how it was validated; whether Sentiment uses the caller's voice or only the transcript ("Positive, relieved, or happy tone" is listed as a signal, without saying how tone is detected); and whether Pulse can be exported (the webhook payload has no Pulse field). Treat Pulse as a sorting tool, and check it by listening.

The EU AI Act draws the line at employees' emotions

The EU AI Act treats emotion recognition differently depending on whose emotions are read. This section follows the Commission's guidelines on prohibited practices (4 February 2025), which are non-binding; authoritative interpretation rests with the Court of Justice.

• Employees: prohibited. Article 5(1)(f), which applies from 2 February 2025, bans AI that infers emotions in the workplace, except for medical or safety reasons. The guidelines' own example: "Using webcams and voice recognition systems by a call centre to track their employee's emotions, such as anger, is prohibited." Personal-training use is allowed only if results aren't shared with HR and can't affect assessment or promotion.

• Customers: allowed, but regulated. "Using voice recognition systems by a call centre to track their customers emotions, such as anger or impatience, is not prohibited by Article 5(1)(f)." Such systems are high-risk, and data protection and consumer law still apply.

• Disclosure. Under Article 50(3), deployers of emotion recognition systems "must inform natural persons who are exposed to those systems" (Commission FAQ on Article 50, last updated 24 Jul 2026).

• Text-only sentiment is outside the definition (no biometric data).

For Brilo, one question stays open: whether Pulse's Sentiment area uses voice features (biometric) or only the transcript. The documentation does not say, so this page does not decide which rules apply. We could not confirm from these sources when the high-risk obligations take effect.

This section explains the Commission's published guidance. It is not legal advice.

Common misconceptions about call analytics

"We analyse 100% of calls, so we understand them." Coverage is not accuracy. Every score inherits the transcript's uneven errors.

"Sentiment is our CSAT." Sentiment is a model's estimate; CSAT is what customers chose to tell you. NBER's agents averaged 0.89 because politeness is the job, not because every call went well.

"More tags and fields mean more insight." Brilo's guidance is "Start with four, not fourteen": did they get what they wanted, what did they want, do they need a person, and the one detail your team always chases.

"The total tells you how we're doing." Brilo's guidance: "Trends beat totals", and "A hundred is where the charts start meaning something".

For where analytics fits among the other AI tools in a call centre, see the AI call centre guide.

How to start: a four-week analytics routine

A workable first month, using Brilo's documented features:

1. Day 1: switch on topic tagging and four fields. Turn on AI Call Tags. Add the four extraction fields above. Make choice fields wherever you can list the options: "free text gives you thirty spellings of the same thing".

2. Weeks 1–4: review tags weekly, sort the call log by lowest Pulse Score, and listen to the bottom calls.

3. Fix the cause, not the score: usually missing knowledge, a missing action or a missing rule.

4. Day 30: compare trends. Look at how outcomes and tags moved, not the totals.

5. Then connect the data: send transcripts, tags and fields to your CRM with the post-call webhook.

For the wider rollout, including success criteria and testing, see how to implement AI in your call centre.

AI call centre analytics and sentiment analysis: frequently asked questions

What is AI speech analytics? AI speech analytics uses speech recognition and language models to turn call audio into searchable, countable data: transcripts, topics, extracted details, sentiment and quality scores. It shows what customers ring about without listening to every recording.

How accurate is call centre sentiment analysis? It depends on the transcript, the audio and the number of categories. In a peer-reviewed study of real emergency calls, a speech emotion model scored 45.6% across four emotions and 76.9% on anger versus neutral (ACII 2021). Vendor accuracy figures are rarely independently verified.

Is emotion recognition legal in EU call centres? Tracking employees' emotions from their voice is prohibited under Article 5(1)(f). Tracking customers' emotions is not prohibited, but is high-risk and people must be informed. Not legal advice.

Which AI call centre solutions automate appointment calls and also report on call quality? Look for one that books during the call and scores every call afterwards. A Brilo agent can book into Google Calendar while the caller is on the line (Book appointments), then adds a Pulse Score, tags, extracted fields and a summary to each call. With any vendor, ask how its quality score is validated.

Does Brilo do sentiment analysis? Yes, as one of the four Pulse Score areas, rated after each call. Live sentiment, human-agent calls, and whether it uses voice tone are not documented.

To see what post-call analysis shows on your own calls, start on the Free plan: see Brilo pricing.