Cisco AI transparency technical notes

Verified 2026-09-05 · 83 claims · sources tier 2–4 · 3 disputed

Also known as AI Impact Assessment, AI Transparency Technical Note, AI Transparency Technical Notes, AII, LLM Proxy, Presidio, toxic hallucination filter.

Cisco publishes 5 AI Transparency Technical Notes covering Webex contact centre features: Call Summaries and Handoff to Agent, Agent Burnout Detection, Auto CSAT, Topic Analysis, and Webex AI Agent 67. The AI features landing page catalogue references 17 distinct Trust Portal transparency notes, including 5 contact centre features, 10 under Cisco AI Assistant, and 6 under AI for Real-Time Media 50.

Catalogue Structure and Scope

The mapping from a named Webex AI feature to its Trust Portal transparency note exists only in the AI features landing page embedded JSON payload, while the rendered page shows section titles without the mapping 51. Sources disagree regarding the scoping model: Cisco's transparency notes FAQ states that Webex publishes 1 AI Transparency Technical Note for every model used across its generally available Collaboration solutions and describes the notes as model-oriented rather than product-oriented 65. However, this is disputed by the documents themselves, as all 5 contact centre transparency notes are titled and scoped by feature rather than by model, and 1 model—GPT-4o mini—appears in 2 separately titled notes instead of a consolidated note 70.

Cisco states in its FAQ that Webex offers no service level agreement for AI features, framing accuracy and reliability as engineering practice rather than a commitment 64. Cisco also states in the FAQ that Webex does not by default use customer content to train models, relying instead on off-the-shelf datasets, Cisco-internal data, or synthetic data 66. Cisco further states that the contact centre AI features covered by these notes are optional and can be disabled by organisation administrators or left to individual user choice 69.

The public records for the Call Summaries and Handoff to Agent, Agent Burnout Detection, and Auto CSAT notes carried a document expiry date of 22 January 2026, which had passed at the access date of 5 September 2026 while remaining marked Active and latest version 47​23​36. Cover dates for 3 of the 5 contact centre notes are January 2025 (19 months old at the access date), while the Topic Analysis note is dated March 2026 and the Webex AI Agent note carries no cover date in its body 72. The PDF properties of the Call Summaries, Agent Burnout Detection, Auto CSAT, and Topic Analysis notes list a subject of Webex Messaging Security and keywords for a Webex Messaging security paper, neither of which matches their actual contents 71. WarmTransfer's reading of the sources is that none of the 8 transparency notes read in this tranche publishes a list or count of supported languages 68.

Summarization and Topic Analytics

The Call Summaries note covers 4 distinct sub-features: call drop summary, virtual agent to human agent handoff summary, mid-call summary, and wrap-up summary 44. Summarization uses GPT-3.5 Turbo and GPT-4o Mini through the Microsoft Azure OpenAI Service API 45. Cisco states it does not use customer content to fine-tune or train these models, and reports Microsoft's representation that Azure OpenAI models are not improved using customer data 46. Cisco states call transcripts are not retained after inference, though generated summaries are stored until a user initiates a new summary request 48. In the Call Summaries note, the statement defining who can view a summary ends mid-sentence, leaving the sharing rule unstated 49.

Topic Analytics chains 2 third-party models: GPT-4o mini on Azure OpenAI Service for call driver extraction, and Mistral-7B-Instruct-V0.2 on Amazon Bedrock for topic labelling 79. Input transcripts are stripped of personally identifiable information using Microsoft's Presidio prior to extraction 77. A toxic hallucination filter prevents the model from generating new unsafe language, though toxic language present in the source transcript can still surface because the filter leaves source input unaltered 78. Customer data is not stored or retained by the model during inference, producing only call drivers and topics 75. For European Union and India customers, transcription and analysis occur locally within those regions, while data for other customers is processed across relevant United States, European Union, and Asia Pacific Japan and China regions 74. The note references 2 sibling notes for feeding transcription steps: an In-House Transcription note and a Third Party Transcription and Real-Time Translation note 73. WarmTransfer's reading of the sources is that the Topic Analysis note publishes no evaluation cadence, threshold, or metric, stating only that Cisco frequently assesses the feature 76.

Burnout Detection and Auto CSAT

Agent Burnout Detection and Auto CSAT models are proprietary to Cisco, not publicly available, and usable by Cisco only within Webex applications or software development kits 37. WarmTransfer's reading of the sources is that the Model Architecture sections of both documents contain a figure with no descriptive text 35.

Agent Burnout Detection is a native Cisco algorithm based on agent performance metrics and behaviour, citing no third-party vendors 40. Customer content is used exclusively to customise the model for that customer's organisation and is not used to train models for other customers or general models 39. The model retrains programmatically without human intervention every 30 days using a rolling 90 days of data 43. Outputs consist of recommended corrective measures and a burnout index scored from 0 (no burnout) to 1 (high burnout) 42. Cisco states this output must not be used for performance appraisals, compensation or penalty decisions, or employment and right-to-work determinations 41. Cisco also states it is not aware of known biases in the model, while committing to monitor and rebalance training datasets for user-identified behaviours 38.

Auto CSAT is a native Cisco deep learning model that predicts customer satisfaction scores for each interaction without third-party vendors 60. The output scale is defined by the customer, using a 1-to-5 scale as an example in the documentation 62. Auto CSAT is evaluated every 30 days, retrained every 30 days to counter data drift and bias, and retrained on the last 180 days of data if accuracy drops below an acceptable threshold 59. WarmTransfer's reading of the sources is that the acceptable accuracy threshold value triggering retraining is not published in the note 63. Cisco specifies that Auto CSAT scores must not be used in isolation to evaluate performance, penalise or compensate agents, or inform employment decisions 61.

Webex AI Agent Architecture and Locality

Webex AI Agent supports autonomous agents powered by large language models for state and dialogue management, alongside scripted agents using conventional machine learning for natural language understanding 19. Autonomous agents use Microsoft Azure OpenAI Service models, specifically naming GPT-4o, GPT-4.1-mini, GPT-4.1, and a GPT-4.1 global variant 22​24. Scripted agents run on Mindmeld or Swiftmatch natural language understanding engines hosted inside Webex Contact Center infrastructure 31. In addition, 3 sub-models run on Cisco's own infrastructure: an interim response model on GPT-4.1-mini, a query generator fine-tuned from facebook bart-large-xsum, and a turn detection ensemble fine-tuned from SmolLM2-135M-Instruct and bert-base-uncased 34.

Speech vendor integrations depend on the engine tier:

• Webex AI Pro 1.0 uses Azure for speech to text and text to speech 33.
• Webex AI Pro-US 1.0 and Webex AI Pro-EU 1.0 use Deepgram for speech to text and Eleven Labs for text to speech 33.

Cisco states that payment card and protected health information are redacted by default, with optional customer masking for names, locations, general health data, email addresses, phone numbers, and IP addresses 29. Cisco notes outputs may contain inaccuracies and advises validation of critical outputs 28. Cisco states that autonomous and scripted agents support multilingual capabilities and publishes no list or count of supported languages anywhere in the note 27. Call recording for voice interactions can be enabled by the customer; Cisco publishes no retention period for recordings or agent transaction data, deferring retention to Offer Disclosures on the Trust Portal 30.

A data locality table defines 8 regional deployments: produs1, prodeu1, prodeu2, prodca1, prodjp1, prodanz1, prodsg1, and prodin1 25. Only London and Frankfurt pin Azure OpenAI inference locally, whereas the United States region lists US Regional plus Global 21. In the remaining 5 regions (Canada, Japan, Australia and New Zealand, Singapore, and India), the Azure OpenAI column reads Global, which footnote text defines as processing across any data centre worldwide where the models are hosted 20. For Singapore (prodsg1), the internal model proxy is in Singapore while Azure OpenAI inference is Global 32. For India (prodin1), the internal model proxy resides in Sydney, Australia, while speech and cloud components reside in Mumbai and Central India 26.

Webex Connect Nodes

Both Webex Connect transparency notes describe their feature as an optional beta feature that clients can have disabled 53. Both features run on GPT-3.5 turbo via Azure OpenAI Service based on Cisco internal latency, cost, and quality benchmarks 58. Both notes state that abuse monitoring is disabled, avoiding data logging for human review 52. Both notes commit to updating their text when underlying models or data processing change 55.

The text summarization node applies no filtering for personal or health data, requiring users to mask sensitive information beforehand 56. Training data augmentation requires at least 1 existing utterance and the intent name, generates up to 20 utterances per run, and is disabled once an intent contains more than 50 training utterances 57. Both Webex Connect notes contain boilerplate text cautioning that unsafe content may output if requested in the code description, despite neither feature involving developer code generation 54.

See also

• See also The Cisco Trust Portal as a monitorable document source.
• See also Cisco offer disclosures as an evidence class.
• See also Webex Contact Center AI agent assist and virtual agents.
• See also Third-party model and speech vendors behind Webex contact centre AI.
• See also Contact centre AI data residency retention and training boundaries.
• See also Employment decision safeguards on automated agent scoring.

Applicability

Across the claims in this article, the evidence covers Cisco Webex Contact Center, Cisco Webex, Cisco Webex AI Agent, and Cisco Webex Connect across cloud deployments, verified as of 2026-09-05. The 8 named regional deployments for Webex AI Agent are produs1, prodeu1, prodeu2, prodca1, prodjp1, prodanz1, prodsg1, and prodin1 25. Topic Analytics transcription and analysis are processed locally within the European Union and India for customers in those regions, while data for other customers is processed across the United States, European Union, and Asia Pacific Japan and China regions 74.

What remains uncertain

The exact sharing rules for generated call summaries are not covered by any claim in this article, as the source text terminates mid-sentence 49. Specific retention schedules for Webex AI Agent voice recordings and transaction data are not covered by any claim in this article. Supported language counts and lists across the Webex AI features are not covered by any claim in this article.

WarmTransfer. (2026, September 5). Cisco AI transparency technical notes. WarmTransfer. https://warmtransfer.net/knowledge/webex-ai-transparency-notes

@misc{warmtransfer-webex-ai-transparency-notes,
  title  = {Cisco AI transparency technical notes},
  author = {{WarmTransfer}},
  year   = {2026},
  url    = {https://warmtransfer.net/knowledge/webex-ai-transparency-notes},
  note   = {Verified 2026-09-05}
}