Elamur Research⌕
☰
home / papers / 2026 / Counterfactual Fairness Evaluation of LLM-Based Contact Center Agent Quality Assurance System

Counterfactual Fairness Evaluation of LLM-Based Contact Center Agent Quality Assurance System

Paper · 2026 · Kawin Mayilvaghanan, Siddhant Gupta and Ayush Kumar

Summary

Contact centers increasingly use LLMs to score agent calls against QA rubrics and to write coaching notes. Because these outputs can affect pay and promotion, the authors ask whether such Auto-QA judges change their verdicts when only irrelevant attributes change. They build a taxonomy of 13 fairness dimensions in three groups: Agent Identity, Contextual Anchoring, and Behavioral Style. For each dimension, they create counterfactual transcript pairs by altering one attribute while keeping the conversation's substance fixed. Examples include swapping names, adding religious phrases, inserting past QA scores, or shifting politeness. Two metrics capture disparity: the Counterfactual Flip Rate (CFR) for binary rubric answers and the Mean Absolute Score Difference (MASD) for confidence and coaching scores. Testing 18 LLMs on 3,000 proprietary contact-center transcripts, they find consistent unfairness. Metadata-based anchoring is the most damaging factor, and implicit linguistic identity cues persist even in otherwise robust models. Larger, more aligned models tend to be fairer, but fairness does not track accuracy. Adding explicit fairness instructions to prompts helps only modestly. The authors argue that fairness audits should be routine before LLMs are used to evaluate workers.

Key points

  • Scale and spread: across the 18 models, CFR ranges from roughly 5.4% to 13.0%. These shifts exceed each model's no-perturbation baseline; for example, claude-4-sonnet scores 5.41% against a 1.33% baseline. This indicates systematic bias rather than random noise.
  • Contextual anchoring is worst: priming with prior coaching notes yields a CFR of 16.44% and an Improvement MASD of 25.61. Past-performance metadata produces a CFR of 9.00%, and customer-profile cues produce 9.64%.
  • Implicit identity cues: when names are paired with cultural or linguistic markers, CFR rises to 9.71% for Ethnicity and 9.24% for Religion.
  • Two failure profiles: some models are consistent but biased (claude-3-5-haiku: CFR 10.57%, baseline 0%). Others are biased and also unstable (deepseek-r1: CFR 10.88%, baseline 9.00%).
  • Prompting and resources: fairness-aware prompting reduces CFR and MASD by about 2–6 points, with llama-4-maverick and nova-pro responding most. For claude-3.5-haiku, CFR slightly worsened (+0.26). Code and a synthetic dataset are available on GitHub at Observeai-Research/FACET, while the real transcripts are not released.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSafety Research · 543 worksBusiness, Management and Accounting · 32 worksOrganizational Behavior and Human Resource Management · 12 works

🧭 Topics 2Research subjects and languages the work is about.Evaluation & benchmarks · 1,220 worksBias, fairness & safety · 517 works

Cite

Cite as Kawin Mayilvaghanan, Siddhant Gupta and Ayush Kumar. 2026. Counterfactual Fairness Evaluation of LLM-Based Contact Center Agent Quality Assurance System. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1890

@inproceedings{mayilvaghanan2026counterfactual,
  title = {{Counterfactual Fairness Evaluation of LLM-Based Contact Center Agent Quality Assurance System}},
  author = {Kawin Mayilvaghanan, Siddhant Gupta and Ayush Kumar},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1890},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1890}
}

References

58 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2019. Fairness and Machine Learning. find
  2. Su Lin et al. Blodgett. 2020. Language (technology) is power: A critical survey of "bias" in nlp. In ACL.
  3. In Conference on Empirical Methods in Natural Language Processing. find
  4. Zirong Chen, Ziyan An, Jennifer Reynolds, Kristin Mullen, Stephen Martini, and Meiyi Ma. 2025. Logidebrief: A signal-temporal logic based automated debriefing approach with large language models integration. Preprint, arXiv:2505.03985. — [arXiv:2505.03985](https://arxiv.org/abs/2505.03985) find
  5. Intersoft Consulting. 2024. Art. 22 gdpr automated individual decision-making, including profiling. find
  6. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. Deepseek-r1: Incentivizing reasoning capability in ll…
  7. Yang Deng, Xuan Zhang, Wenxuan Zhang, Yifei Yuan, See-Kiong Ng, and Tat-Seng Chua. 2024. On the multi-turn instruction following for conversational web agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8795–8812, Bangkok… find
  8. Ananya et al. Deshpande. 2023. Measuring and mitigating toxicity bias in large language models. In EMNLP. find
  9. Emily et al. Dinan. 2020. Multi-dimensional gender bias classification in dialogue agents. In EMNLP. find
  10. Reem Gody, Mahmoud Goudy, and Ahmed Y. Tawfik. 2025. Convogen: Enhancing conversational ai with synthetic data: A multi-agent approach. Preprint, arXiv:2503.17460. — [arXiv:2503.17460](https://arxiv.org/abs/2503.17460) find
  11. Bryce Goodman and Seth Flaxman. 2017. European union regulations on algorithmic decision making and a “right to explanation”. AI Magazine, 38(3):50–57. find
  12. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  13. Shengyue Guan, Haoyi Xiong, Jindong Wang, Jiang Bian, Bin Zhu, and Jian guang Lou. 2025. Evaluating llm-based agents for multi-turn conversations: A survey. Preprint, arXiv:2503.22458. — [arXiv:2503.22458](https://arxiv.org/abs/2503.22458) find
  14. Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2025. Comprehensive framework for evaluating conversational ai chatbots. Preprint, arXiv:2502.06105. — [arXiv:2502.06105](https://arxiv.org/abs/2502.06105) find
  15. Digvijay Ingle, Aashraya Sachdeva, Surya Prakash Sahu, Mayank Sati, Cijo George, and Jithendra Vepa. 2024. Probing the depths of language models’ contact-center knowledge for quality assurance. find
  16. Hayate Iso, Pouya Pezeshkpour, Nikita Bhutani, and Estevam Hruschka. 2025. Evaluating bias in LLMs for job-resume matching: Gender, race, and education. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Techn… find
  17. Matt J. Kusner, Joshua R. Loftus, Chris Russell, and Ricardo Silva. 2018. Counterfactual fairness. Preprint, arXiv:1703.06856. — [arXiv:1703.06856](https://arxiv.org/abs/1703.06856) find
  18. Md Tahmid Rahman Laskar, Cheng Chen, Xue-Yong Fu, Mahsa Azizi, Shashi Bhushan, and Simon Corston-Oliver. 2023. Ai coach assist: An automated approach for call recommendation in contact centers for agent coaching. Preprint, arXiv:2305.17619. find
  19. Ninareh et al. Mehrabi. 2021. A survey on bias and fairness in machine learning. ACM Computing Surveys.
  20. Eric et al. Mitchell. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In NeurIPS.
  21. Moin et al. Nadeem. 2021. Stereoset: Measuring stereotypical bias in pretrained language models. In ACL.
  22. Nikita et al. Nangia. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In EMNLP.
  23. Ihor Pysmennyi, Roman Kyslyi, and Kyrylo Kleshch. 2025. Ai-driven tools in modern software quality assurance: an assessment of benefits, challenges, and future directions. Technology audit and production reserves, 3(2(83)):44–54. find
  24. Pooja S. B. Rao, Laxminarayen Nagarajan Venkatesan, Mauro Cherubini, and Dinesh Babu Jayagopi. 2025. Invisible filters: Cultural bias in hiring evaluations using large language models. Preprint, arXiv:2508.16673. — [arXiv:2508.16673](https://arxiv.org/abs/2508.16673) find
  25. Shourya Roy, Ragunathan Mariappan, Sandipan Dandapat, Saurabh Srivastava, Sainyam Galhotra, and Balaji Peddamuthu. 2016. Qart: A system for real-time holistic quality assurance for contact center dialogues. Proceedings of the AAAI Conference on Artificial Intelligence, 30. find
  26. John et al. Smith. 2022. Impartial: Evaluating fairness of pretrained language models via masked counterfactuals. In ACL. find
  27. Kaisong Song, Lidong Bing, Wei Gao, Jun Lin, Lujun Zhao, Jiancheng Wang, Changlong Sun, Xiaozhong Liu, and Qiong Zhang. 2019. Using customer service dialogues for satisfaction analysis with context-assisted multiple instance learning. In Proceedings of the 2019 Conference on Empirical Methods in Nat… find
  28. U.S. Equal Employment Opportunity Commission. 2024. Regulations and guidelines. find
  29. Ze Wang, Zekun Wu, Xin Guan, Michael Thaler, Adriano Koshiyama, Skylar Lu, Sachin Beepath, Ediz Ertekin, and Maria Perez-Ortiz. 2024. JobFair: A framework for benchmarking gender hiring bias in large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 322… find
  30. Yihao Wu, Tianrui Wang, Yizhou Peng, Yi-Wen Chao, Xuyi Zhuang, Xinsheng Wang, Shunshun Yin, and Ziyang Ma. 2025. Evaluating bias in spoken dialogue llms for real-world decisions and recommendations. Preprint, arXiv:2510.02352. \#\# References Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2019. Fa… — [arXiv:2510.02352](https://arxiv.org/abs/2510.02352) find
  31. Su Lin et al. Blodgett. 2020. Language (technology) is power: A critical survey of "bias" in nlp. In ACL.
  32. Zirong Chen, Ziyan An, Jennifer Reynolds, Kristin Mullen, Stephen Martini, and Meiyi Ma. 2025. Logidebrief: A signal-temporal logic based automated debriefing approach with large language models integration. Preprint, arXiv:2505.03985. — [arXiv:2505.03985](https://arxiv.org/abs/2505.03985) find
  33. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. Deepseek-r1: Incentivizing reasoning capability in ll…
  34. Yang Deng, Xuan Zhang, Wenxuan Zhang, Yifei Yuan, See-Kiong Ng, and Tat-Seng Chua. 2024. On the multi-turn instruction following for conversational web agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8795–8812, Bangkok… find
  35. Ananya et al. Deshpande. 2023. Measuring and mitigating toxicity bias in large language models. In EMNLP. find
  36. Emily et al. Dinan. 2020. Multi-dimensional gender bias classification in dialogue agents. In EMNLP. find
  37. Reem Gody, Mahmoud Goudy, and Ahmed Y. Tawfik. 2025. Convogen: Enhancing conversational ai with synthetic data: A multi-agent approach. Preprint, arXiv:2503.17460. — [arXiv:2503.17460](https://arxiv.org/abs/2503.17460) find
  38. Bryce Goodman and Seth Flaxman. 2017. European union regulations on algorithmic decision making and a “right to explanation”. AI Magazine, 38(3):50–57. find
  39. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  40. Digvijay Ingle, Aashraya Sachdeva, Surya Prakash Sahu, Mayank Sati, Cijo George, and Jithendra Vepa. 2024. Probing the depths of language models’ contact-center knowledge for quality assurance. In Conference on Empirical Methods in Natural Language Processing. find
  41. Intersoft Consulting. 2024. Art. 22 gdpr automated individual decision-making, including profiling. find
  42. Hayate Iso, Pouya Pezeshkpour, Nikita Bhutani, and Estevam Hruschka. 2025. Evaluating bias in LLMs for job-resume matching: Gender, race, and education. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Techn… find
  43. Matt J. Kusner, Joshua R. Loftus, Chris Russell, and Ricardo Silva. 2018. Counterfactual fairness. Preprint, arXiv:1703.06856. — [arXiv:1703.06856](https://arxiv.org/abs/1703.06856) find
  44. Md Tahmid Rahman Laskar, Cheng Chen, Xue-Yong Fu, Mahsa Azizi, Shashi Bhushan, and Simon Corston-Oliver. 2023. Ai coach assist: An automated approach for call recommendation in contact centers for agent coaching. Preprint, arXiv:2305.17619. find
  45. Ninareh et al. Mehrabi. 2021. A survey on bias and fairness in machine learning. ACM Computing Surveys.
  46. Eric et al. Mitchell. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In NeurIPS.
  47. Moin et al. Nadeem. 2021. Stereoset: Measuring stereotypical bias in pretrained language models. In ACL.
  48. Nikita et al. Nangia. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In EMNLP.
  49. Ihor Pysmennyi, Roman Kyslyi, and Kyrylo Kleshch. 2025. Ai-driven tools in modern software quality assurance: an assessment of benefits, challenges, and future directions. Technology audit and production reserves, 3(2(83)):44–54. find
  50. Pooja S. B. Rao, Laxminarayen Nagarajan Venkatesan, Mauro Cherubini, and Dinesh Babu Jayagopi. 2025. Invisible filters: Cultural bias in hiring evaluations using large language models. Preprint, arXiv:2508.16673. — [arXiv:2508.16673](https://arxiv.org/abs/2508.16673) find
  51. Shengyue Guan, Haoyi Xiong, Jindong Wang, Jiang Bian, Bin Zhu, and Jian guang Lou. 2025. Evaluating llm-based agents for multi-turn conversations: A survey. Preprint, arXiv:2503.22458. — [arXiv:2503.22458](https://arxiv.org/abs/2503.22458) find
  52. Shourya Roy, Ragunathan Mariappan, Sandipan Dandapat, Saurabh Srivastava, Sainyam Galhotra, and Balaji Peddamuthu. 2016. Qart: A system for real-time holistic quality assurance for contact center dialogues. Proceedings of the AAAI Conference on Artificial Intelligence, 30. find
  53. Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2025. Comprehensive framework for evaluating conversational ai chatbots. Preprint, arXiv:2502.06105. — [arXiv:2502.06105](https://arxiv.org/abs/2502.06105) find
  54. John et al. Smith. 2022. Impartial: Evaluating fairness of pretrained language models via masked counterfactuals. In ACL. find
  55. Kaisong Song, Lidong Bing, Wei Gao, Jun Lin, Lujun Zhao, Jiancheng Wang, Changlong Sun, Xiaozhong Liu, and Qiong Zhang. 2019. Using customer service dialogues for satisfaction analysis with context-assisted multiple instance learning. In Proceedings of the 2019 Conference on Empirical Methods in Nat… find
  56. U.S. Equal Employment Opportunity Commission. 2024. Regulations and guidelines. find
  57. Ze Wang, Zekun Wu, Xin Guan, Michael Thaler, Adriano Koshiyama, Skylar Lu, Sachin Beepath, Ediz Ertekin, and Maria Perez-Ortiz. 2024. JobFair: A framework for benchmarking gender hiring bias in large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 322… find
  58. Yihao Wu, Tianrui Wang, Yizhou Peng, Yi-Wen Chao, Xuyi Zhuang, Xinsheng Wang, Shunshun Yin, and Ziyang Ma. 2025. Evaluating bias in spoken dialogue llms for real-world decisions and recommendations. Preprint, arXiv:2510.02352. — [arXiv:2510.02352](https://arxiv.org/abs/2510.02352) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mayilvaghanan-2026-counterfactual-fairness-evaluation-llm-based/