Two trained human annotators labelling the sentiment of 20,000 call-centre utterances agreed at a Cohen’s kappa of only 0.605, on a scale where 1.0 is perfect agreement and 0 is chance (ConnexAI researchers, IntechOpen 2026). That ceiling matters because every sentiment score in a speech analytics dashboard is trained and graded against labels like these. Meanwhile, commercial speech recognition systems got 16.5% to 19.2% of words wrong on 50 real call center conversations, versus 10.2% to 11.62% on a standard phone benchmark (Szymanski et al., Findings of EMNLP 2020), and 49.8% of 1,104 companies already run interaction analytics on those transcripts (Metrigy, AI for Business Success 2025-26). We aggregated data from Metrigy, NICE and Verint SEC filings, speech benchmarks published at EMNLP and LREC, by OpenAI and on arXiv, the NBER, the EU AI Act and the Illinois General Assembly, and several other primary sources. For the automation side of the same industry (voicebots and AI agents rather than analytics), see our voice AI contact center automation statistics.
TL;DR
- 49.8% of companies already use interaction analytics and 29.9% more plan to add it (Metrigy, AI for Business Success 2025-26).
- 55% of contact centers still run quality assurance on spreadsheets, down from 65% a year earlier (Scorebuddy and Call Centre Helper, December 2025).
- Commercial ASR word error rate on 50 real call center conversations: 16.5% to 19.2% (Szymanski et al., Findings of EMNLP 2020).
- Whisper large-v2 makes 17.6% word errors on CallHome phone calls but 2.7% on read audiobook speech (OpenAI, Whisper paper 2022).
- On real African-accented call center audio, Whisper-large-v3 scored 24.69% WER and the best model 13.45% (AfriSpeech-MultiBench, arXiv 2025).
- Human annotators agree on call-centre utterance sentiment at a Cohen’s kappa of 0.605 (ConnexAI researchers, IntechOpen 2026).
- Two annotators rating real customer calls agreed at an ICC of only 0.62 on valence and 0.52 on arousal (CusEmo study, arXiv 2023).
- Acoustic-only sentiment models reached 54.23% weighted accuracy on phone calls, against 85.76% for humans (Chen et al., LREC 2020).
- An AI assistant raised support agent productivity by 14% and customer sentiment by 0.18 points on a -1 to 1 scale (NBER Working Paper 31161).
- Inferring employees’ emotions at work from voice has been banned in the EU since February 2, 2025 (EU AI Act, Regulation (EU) 2024/1689).
- Illinois sets damages of $1,000 to $5,000 per violation involving biometric identifiers such as voiceprints (740 ILCS 14).
1. Adoption: Interaction Analytics and Automated Quality Assurance
The adoption question is half settled. 49.8% of 1,104 companies already use interaction analytics (software that transcribes and mines calls, chats and emails for topics, compliance phrases and sentiment), and another 29.9% planned to implement it during the year (Metrigy, AI for Business Success 2025-26). What has changed is the job description. The largest group of adopters, 46.2%, now say the biggest value is comparing how AI agents perform against human agents (Metrigy analysis, May 14, 2025). Speech analytics is turning into the scoreboard for automation, not just a coaching tool.
The quality assurance (QA) process that analytics is meant to replace is still surprisingly manual. A vendor-sponsored survey found that reliance on spreadsheets for QA fell from 65% to 55% year on year, and 44% of contact centers expressed strong interest in automated or real-time scoring (Scorebuddy and Call Centre Helper, What Contact Centers Are Doing Right Now, December 2025). The vendor base is concentrated and large: NICE reported total 2025 revenue of $2,945.4 million, up from $2,735.3 million, with cloud revenue up 12.8%, and it paid a total final consideration of $887.3 million for the conversational AI company Cognigy (NICE Form 20-F for fiscal 2025).
| Metric | Value | Source |
|---|---|---|
| Companies already using interaction analytics | 49.8% of 1,104 companies | Metrigy, AI for Business Success 2025-26 |
| Companies planning to implement interaction analytics during 2025 | 29.9% | Metrigy, AI for Business Success 2025-26 |
| Analytics adopters whose biggest value is comparing AI agent effectiveness with human agent effectiveness | 46.2% (the largest group) | Metrigy analysis, May 2025 |
| Contact centers using spreadsheets for quality assurance, this year vs last year | 55% vs 65% | Scorebuddy and Call Centre Helper survey, December 2025 (vendor-sponsored) |
| Contact centers with strong interest in automated or real-time quality scoring | 44% | Scorebuddy and Call Centre Helper survey, December 2025 (vendor-sponsored) |
| NICE customer organizations and Fortune 100 customers | More than 25,000 organizations in over 150 countries, including 85 of the Fortune 100 | NICE Form 20-F, fiscal 2025 |
| Verint Fortune 100 customers and countries served | Over 80 of the Fortune 100, in over 175 countries | Verint Form 10-K, fiscal year ended January 31, 2025 |
| Verint revenue from its subscription business | Approximately 80% | Verint Form 10-K, fiscal year ended January 31, 2025 |
Context note: figures claiming that manual QA teams review only “1% to 3%” or “fewer than 5%” of calls circulate widely, but every version we found traced back to vendor blogs or unverifiable citations, so we excluded them. No independent public survey in 2025 or 2026 reports the share of calls that are human-reviewed versus machine-scored. The Verint Form 10-K also reports that about 72% of its revenue came from the Americas.
2. Transcription Accuracy on Real Call Audio
Every text-based sentiment score starts with a transcript, and narrowband (8 kHz), spontaneous call audio is among the hardest everyday audio a recognizer faces. Word error rate (WER) is the share of words the system substitutes, deletes or inserts compared with a human transcript. Three commercial systems scored 16.5% to 19.2% WER on 50 real call center conversations, against 10.2% to 11.62% on the Switchboard phone benchmark (Szymanski et al., WER we are and WER we think we are, Findings of EMNLP 2020). The authors note that the best published Switchboard results are about 5%, so real calls ran at roughly three to four times the error rate that headline benchmarks suggest (16.5 / 5 = 3.3 and 19.2 / 5 = 3.84). Most recent available data for commercial systems on private call center audio: Szymanski et al., 2020; we found no public repeat of this test on real call recordings with current commercial engines.
Newer open models show the same gap. Whisper large-v2 scored 2.7% WER on LibriSpeech test-clean (read audiobooks) but 17.6% on CallHome and 13.8% on Switchboard, both telephone conversation sets (OpenAI, Robust Speech Recognition via Large-Scale Weak Supervision, 2022). Accent widens it further: on a small private set of real African-accented agent-customer calls, results ranged from 13.45% for the best model to 88.71% for the worst (AfriSpeech-MultiBench, arXiv November 2025). Our accent bias and word error rate statistics cover that dimension in depth.
| Metric | Value | Source |
|---|---|---|
| WER of three commercial ASR systems on 50 real call center conversations (8.5 hours of audio, 2.2 hours of speech) | 16.5% to 19.2% | Szymanski et al., Findings of EMNLP 2020 |
| WER of the same three systems on the HUB’05 Switchboard telephone benchmark | 10.2% to 11.62% | Szymanski et al., Findings of EMNLP 2020 |
| Call center WER by business domain for the best system | 13.73% (insurance) to 20.95% (travel booking) | Szymanski et al., Findings of EMNLP 2020 |
| WER on customer speech vs agent speech, for the weakest of the three systems | 20.99% vs 17.83% | Szymanski et al., Findings of EMNLP 2020 |
| Whisper large-v2 WER on read audiobook speech (LibriSpeech test-clean) vs CallHome phone calls vs Switchboard phone calls | 2.7% vs 17.6% vs 13.8% | OpenAI, Whisper paper 2022 |
| NVIDIA Conformer-CTC large vs wav2vec2-large-960h WER on CallHome phone calls | 11.3% vs 34.8% | OpenAI, Whisper paper 2022 (comparison table) |
| WER on real African-accented call center audio: best model vs Whisper-large-v3 vs GPT-4o Transcribe | 13.45% vs 24.69% vs 23.20% | AfriSpeech-MultiBench, arXiv November 2025 |
| Share of an ASR-transcribed call-centre evaluation set (618 utterances) containing transcription errors | 42.6% (263 samples) | ConnexAI researchers, IntechOpen 2026 |
Outlier note: the agent-versus-customer gap is not universal. For the best of the three commercial systems, customer and agent WER were nearly identical (16.48% vs 16.49%), so the customer side is harder for some engines, not all. Errors also differ in kind from benchmark errors; our speech recognition hallucination statistics cover invented words, which a WER score alone does not separate out.
3. Text Sentiment on Call Transcripts: Human Agreement and Machine Accuracy
The uncomfortable number in sentiment analytics is the human baseline. Two annotators labelling 20,000 call-centre utterances (broadband, insurance and other industries) agreed at a Cohen’s kappa of 0.605 (ConnexAI researchers, Enhancing Call Centre Sentiment Analysis with Large Language Models, IntechOpen 2026). Kappa corrects raw agreement for chance: 1.0 is perfect agreement and 0 is what chance alone would produce. Any vendor figure for “sentiment accuracy” is graded against labels that humans themselves often dispute.
Within that ceiling, models do well on clean text and stumble on the class that matters most. With full annotation instructions, a 32-billion-parameter open LLM matched the human gold standard 91% of the time on three classes, but its macro F1 (the average of the per-class F1 scores, which punishes ignoring rare classes) was 0.79. The real stress test is ASR output: on production calls transcribed by machine, the same approach reached a macro F1 of 0.74 from single utterances and 0.86 when it saw a six-turn window. Context did more than size: in the same tests, a 4-billion-parameter Qwen3 model scored 0.78 macro F1 while its 32-billion-parameter sibling scored 0.56. On the Switchboard phone corpus, the label mix shows why “neutral” dominates dashboards: 52.6% of segments were neutral (Chen et al., A Large Scale Speech Sentiment Corpus, LREC 2020).
| Metric | Value | Source |
|---|---|---|
| Agreement between two human annotators on the sentiment of 20,000 call-centre utterances (Cohen’s kappa) | 0.605 | ConnexAI researchers, IntechOpen 2026 |
| Open LLM (Qwen2.5-32B-Instruct) with full instructions vs human gold standard, 3 sentiment classes, 8,750 samples | 0.91 accuracy, 0.79 macro F1 | ConnexAI researchers, IntechOpen 2026 |
| Fine-tuned DistilBERT vs SVM classifier, macro F1 on human-transcribed calls | 0.90 vs 0.82 | ConnexAI researchers, IntechOpen 2026 |
| LLM on machine (ASR) transcripts of production calls: single utterance vs six-turn context window, macro F1 | 0.74 vs 0.86 | ConnexAI researchers, IntechOpen 2026 |
| Average accuracy of individual human annotators against the majority label, Switchboard phone calls (49,500 segments, 140 hours) | Around 85% | Chen et al. (Google), LREC 2020 |
Context note: in the same chapter, the negative class was the weak spot even for the best classifier (fine-tuned DistilBERT reached an F1 of 0.78 on negative sentiment vs 0.94 on positive), and the SVM collapsed on machine transcripts, scoring a macro F1 of 0.25 on agent speech and 0.46 on customer speech. Machine-transcribed production calls were also skewed: 61% neutral, 30% positive and 9% negative. A model that labels everything neutral looks accurate on such data while missing the angry 9%.
4. Acoustic Emotion Detection: What the Voice Adds
Tone of voice is the selling point of “emotion AI”, yet in the Switchboard phone-call sentiment benchmark it was the weakest single signal. An acoustic-only model reached 54.23% weighted accuracy on three-class sentiment, against 85.76% for humans, and adding text lifted it to 65.65% (Chen et al., LREC 2020, reporting Lu et al. 2019 baselines). Weighted accuracy counts every segment equally; unweighted accuracy averages the three classes, so the acoustic model’s 39.63% unweighted score sits only about 6 points above the 33.3% that random guessing would score (39.63 - 33.33 = 6.3). Most recent available data for a public phone-call sentiment benchmark comparing all three input types: Switchboard sentiment corpus, LREC 2020.
Real customer service calls are even harder to label. In a French business-travel call center (the CusEmo study), two annotators rating customers’ emotion continuously reached an intraclass correlation (ICC, an agreement score from 0 to 1) of 0.62 for valence (how positive or negative) and 0.52 for arousal (how activated or calm) (End-to-End Continuous Speech Emotion Recognition in Real-life Customer Service Call Center Conversations, arXiv 2023). The best model’s concordance with those ratings (CCC, where 1 is perfect) was 0.699 for valence and 0.508 for arousal, after it was told how empathetic the agent was. In other words, the customer’s mood is easier to read when the model also hears the agent. For clinical uses of the same technology, see our speech emotion recognition accuracy statistics.
| Metric | Value | Source |
|---|---|---|
| Weighted accuracy, 3-class sentiment on Switchboard phone calls: acoustic-only model vs acoustic plus text model vs model using speech recognition features vs human listeners | 54.23% vs 65.65% vs 70.10% vs 85.76% | Chen et al. (Google), LREC 2020 |
| Unweighted (class-averaged) accuracy, same four approaches in the same order | 39.63% vs 54.59% vs 62.39% vs 84.61% | Chen et al. (Google), LREC 2020 |
| CusEmo source collection from a French business-travel call center | 5,548 conversations, 375.39 hours | CusEmo study, arXiv 2023 |
| CusEmo conversations with context annotation; customer audio rated for emotion | 469 conversations; 13.15 hours | CusEmo study, arXiv 2023 |
| Agreement between two annotators rating customer emotion (ICC): valence vs arousal | 0.62 vs 0.52 | CusEmo study, arXiv 2023 |
| Best model concordance with human ratings (CCC) when agent empathy is added: valence vs arousal | 0.699 vs 0.508 | CusEmo study, arXiv 2023 |
Context note: the CusEmo corpus is for internal use only and has not been published, and only 469 of the 5,548 conversations received context annotation. Real call-center emotion data is scarce precisely because recordings carry personal data, which is one reason most published accuracy figures come from podcasts, acted speech or decades-old phone corpora.
5. Measured Business Outcomes
Analytics vendors sell outcomes, and the best-controlled evidence comes from text, not voice. In a study of 5,179 customer support agents and about 3 million chats, an AI conversation assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled agents (Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161). The study measured customer mood with an LLM-based sentiment classifier (SiEBERT, scoring text from -1 to 1), which makes it one of the few large field studies that uses automated sentiment as an outcome rather than a feature.
The softer metrics moved too: customer sentiment rose by 0.18 points (half a standard deviation), requests to speak to a manager fell by almost 25%, and handle time dropped by 3.8 minutes. Resolution barely changed. Buyers seem to believe the data belongs at the top: nearly 84% of companies, and 96.7% of Metrigy’s success group, say customer interaction data should be part of executive performance dashboards (Metrigy analysis, May 2025).
| Metric | Value | Source |
|---|---|---|
| Productivity gain (issues resolved per hour) with an AI conversation assistant, 5,179 agents | 14% on average | NBER Working Paper 31161, Brynjolfsson, Li and Raymond |
| Productivity gain for novice and low-skilled agents | 34% | NBER Working Paper 31161 |
| Change in mean customer sentiment (SiEBERT score from -1 to 1) | +0.18 points, about half a standard deviation | NBER Working Paper 31161 |
| Change in customer requests to speak to a manager (baseline about 6 percentage points) | Almost 25% lower | NBER Working Paper 31161 |
| Change in average handle time | 3.8 minutes shorter, a 9% decline | NBER Working Paper 31161 |
| Change in resolution rate (baseline 82%) | +1.3 percentage points, significant only at the 10% level | NBER Working Paper 31161 |
| Companies that want interaction data on executive dashboards: all companies vs the most successful group | Nearly 84% vs 96.7% | Metrigy analysis, May 2025 |
Outlier note: agent sentiment in the NBER study rose by only 0.02 points, about 1% of a standard deviation, so the gain showed up in how customers wrote, not how agents wrote. The study covers text chat at one company; we found no comparably large controlled study of voice-call analytics outcomes.
6. Regulation and Privacy: Emotion Recognition Rules and Voiceprint Law
The legal line runs between whose emotions are read and from what. Since February 2, 2025, the EU AI Act has prohibited AI systems that infer the emotions of a person in the workplace, except for medical or safety reasons (Regulation (EU) 2024/1689, Articles 5(1)(f) and 113). The Act defines emotion recognition as inference “on the basis of their biometric data”, which is why the Commission’s guidelines treat a call center that uses voice recognition to track employees’ anger as prohibited, but a call center tracking customers’ emotions as outside the ban, and inference from written text as outside it too (European Commission guidelines on prohibited AI practices, February 4, 2025, examples as summarized by CMS, February 10, 2025). In practice, transcript-based sentiment on customers is the safest design in Europe, and voice-tone scoring of agents is the riskiest.
Recital 44 of the Act justifies the ban by pointing to serious concerns about the scientific basis of emotion recognition, naming limited reliability, lack of specificity and limited generalisability, the same weaknesses the agreement scores in sections 3 and 4 measure. In the United States, the sharpest exposure is biometric: Illinois counts a voiceprint as a biometric identifier, so any system that builds speaker voiceprints (for example, for caller authentication) falls under its rules and its per-violation damages (Illinois Biometric Information Privacy Act, 740 ILCS 14).
| Metric | Value | Source |
|---|---|---|
| Date the EU ban on inferring emotions in the workplace and in education (Article 5(1)(f)) applies | February 2, 2025 | EU AI Act, Regulation (EU) 2024/1689, Article 113 |
| Maximum fine for using a prohibited AI practice | EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher | EU AI Act, Article 99(3) |
| Status of emotion recognition systems that are not prohibited (for example on customers) | Listed as high-risk AI systems | EU AI Act, Annex III, point 1(c) |
| General application date of the EU AI Act as published | August 2, 2026 | EU AI Act, Article 113 |
| Call center using voice recognition to track employees’ emotions such as anger | Prohibited | European Commission guidelines, February 2025 |
| Call center using voice recognition to track customers’ emotions | Outside the workplace prohibition | European Commission guidelines, February 2025 |
| Illinois liquidated damages per violation involving biometric identifiers such as voiceprints: negligent vs intentional or reckless | $1,000 vs $5,000 | Illinois BIPA, 740 ILCS 14/20 |
| Illinois BIPA effective date and latest amendment | October 3, 2008; amended by Public Act 103-769, effective August 2, 2024 | Illinois General Assembly |
Context note: the Commission’s guidelines are non-binding, and authoritative interpretation rests with the Court of Justice of the EU. NICE’s own annual report warns that the AI Act is “gradually applicable through August 2026” and that fines for non-compliance “are significant” (NICE Form 20-F for fiscal 2025).
Summary: Call Center Speech Analytics Sentiment Detection Accuracy by the Numbers
| Metric | Value | Source |
|---|---|---|
| Companies already using interaction analytics vs planning to implement it in 2025 | 49.8% vs 29.9% (of 1,104) | Metrigy, AI for Business Success 2025-26 |
| Contact centers still using spreadsheets for quality assurance | 55% (down from 65%) | Scorebuddy and Call Centre Helper, December 2025 |
| NICE 2025 total revenue | $2,945.4 million | NICE Form 20-F, fiscal 2025 |
| Commercial ASR WER on 50 real call center conversations | 16.5% to 19.2% | Szymanski et al., Findings of EMNLP 2020 |
| Same systems on the Switchboard phone benchmark | 10.2% to 11.62% | Szymanski et al., Findings of EMNLP 2020 |
| Whisper large-v2 WER: read audiobook speech vs CallHome phone calls | 2.7% vs 17.6% | OpenAI, Whisper paper 2022 |
| WER on African-accented call center audio: best model vs Whisper-large-v3 | 13.45% vs 24.69% | AfriSpeech-MultiBench, arXiv 2025 |
| Human-human agreement on call-centre utterance sentiment (Cohen’s kappa) | 0.605 | ConnexAI researchers, IntechOpen 2026 |
| LLM sentiment macro F1 on ASR transcripts: single utterance vs six-turn context | 0.74 vs 0.86 | ConnexAI researchers, IntechOpen 2026 |
| Weighted accuracy on phone-call sentiment: acoustic-only vs acoustic plus text vs human | 54.23% vs 65.65% vs 85.76% | Chen et al. (Google), LREC 2020 |
| Annotator agreement on customer valence vs arousal in real calls (ICC) | 0.62 vs 0.52 | CusEmo study, arXiv 2023 |
| Productivity gain from AI conversation assistant, all agents vs novices | 14% vs 34% | NBER Working Paper 31161 |
| Customer sentiment change (-1 to 1 scale) and requests to speak to a manager with AI assistance | +0.18 points; almost 25% fewer requests | NBER Working Paper 31161 |
| EU ban on workplace emotion inference applies from | February 2, 2025 | EU AI Act, Article 113 |
| Call center using voice recognition to track employees’ emotions vs customers’ emotions | Prohibited vs outside the workplace ban | European Commission guidelines, February 2025 |
| Maximum EU fine for prohibited AI practices | EUR 35,000,000 or 7% of worldwide turnover | EU AI Act, Article 99(3) |
| Illinois damages per biometric violation, negligent vs intentional | $1,000 vs $5,000 | Illinois BIPA, 740 ILCS 14 |
Methodology and Sources
- Metrigy, The Value of Interaction Analytics (AI for Business Success 2025-26, 1,104 companies) and Metrigy analysis of the same study, May 14, 2025: adoption, planned adoption, top use cases, dashboard attitudes.
- Scorebuddy and Call Centre Helper, What Contact Centers Are Doing Right Now, December 11, 2025: QA tooling survey, vendor-sponsored; sample size not disclosed on the release.
- NICE Ltd, Form 20-F for the fiscal year ended December 31, 2025 (SEC EDGAR): revenue, cloud growth, customer base, Cognigy consideration, AI Act risk factor.
- Verint Systems, Form 10-K for the fiscal year ended January 31, 2025 (SEC EDGAR): customer base, countries, subscription share, regional revenue.
- Szymanski et al., WER we are and WER we think we are, Findings of EMNLP 2020: commercial ASR on 50 real call center conversations.
- Radford et al. (OpenAI), Robust Speech Recognition via Large-Scale Weak Supervision, 2022: Whisper WER on LibriSpeech, CallHome and Switchboard.
- Ashungafac et al. (Intron Health), AfriSpeech-MultiBench, arXiv November 2025: call center domain WER for 18 ASR systems.
- Jordan, Fonseka, Stacey and Sitvaev (ConnexAI), Enhancing Call Centre Sentiment Analysis with Large Language Models, IntechOpen, June 20, 2026: human agreement, LLM and classifier sentiment accuracy, ASR robustness. The authors work for a contact center AI vendor.
- Chen et al. (Google), A Large Scale Speech Sentiment Corpus, LREC 2020 (LDC2020T14): Switchboard sentiment labels, human accuracy, acoustic and text baselines.
- End-to-End Continuous Speech Emotion Recognition in Real-life Customer Service Call Center Conversations (CusEmo), arXiv 2023: real call center emotion corpus and model results.
- Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161: productivity, sentiment, escalation and attrition outcomes.
- Regulation (EU) 2024/1689 (EU AI Act), EUR-Lex: Article 3(39), Article 5(1)(f), Article 99(3), Article 113, Annex III, Recital 44.
- European Commission, Guidelines on prohibited AI practices, February 4, 2025 and CMS summary of the call center examples, February 10, 2025.
- Illinois General Assembly, Biometric Information Privacy Act, 740 ILCS 14: voiceprint definition and liquidated damages.
- Data watch: the commercial call center WER study (2020), the Whisper paper (2022), the Switchboard sentiment baselines (2020) and the CusEmo study (2023) are the most recent available public data for their measurements; we found no newer public test of commercial engines on real call recordings, and the call center datasets used in these papers are private. The IntechOpen chapter is internally inconsistent in places (its abstract cites 3,000 sentiment observations while the body uses 8,750, and it gives an ASR F1 of “around 0.76” in the abstract versus a macro F1 of 0.74 and accuracy of 0.76 in the body); we used the body tables. The CusEmo paper’s text reports 0.657 and 0.434 for one setting that its table lists as .667 and .499. AfriSpeech-MultiBench’s call center set has only 16 samples (0.80 hours), and the paper says both 18 and 19 systems were evaluated. The NBER outcomes come from text chat at a single firm, not voice calls. Metrigy describes the same adoption result as 49.8% in one place and “only half” in another. Widely repeated claims that manual QA covers 1% to 3% of calls could not be traced to an independent survey and were excluded. The EU AI Act timeline is quoted as published in the Official Journal; check for later amendments before relying on the August 2, 2026 date for high-risk obligations.
Last updated: October 27, 2026. We update this roundup quarterly, and the next refresh is expected when Metrigy publishes its next AI for Business Success study and NICE files its next annual Form 20-F.