---
title: 'Minerva CQ: Proactive AI for Voice Calls'
url: https://www.emergentmind.com/topics/minerva-cq
type: topic
---

# Minerva CQ: Proactive AI for Voice Calls

Minerva CQ is a real-time Agent Assist system purpose-built for voice-based customer support centers and presented as a case study in "Redefining CX with Agentic AI: Minerva CQ Case Study" [2509.12589]. It is defined as an instance of the Agentic AI paradigm: rather than acting as a passive, reactive agent-assist, it operates as a goal-driven, autonomous, tool-using AI co-pilot that proactively supports human agents throughout the entire customer conversation lifecycle. In this formulation, Minerva CQ is positioned against static rules, simple prompting, or retrieval-augmented generation (RAG) without deeper contextual reasoning; it identifies customer intent, triggers modular workflows, maintains evolving context, and adapts dynamically to conversation state.

## 1. Operational problem and design objective

The system is motivated by persistent CX failures in contact centers, notably high average handling time (AHT), low first-call resolution (FCR), and poor customer satisfaction (CSAT) [2509.12589]. The paper attributes a key share of this degradation to agent cognitive load: agents must navigate fragmented tools, troubleshoot manually, and frequently place customers on hold. Existing AI-powered agent-assist tools are described as reactive and limited to isolated suggestions, because they are driven by static rules, simple prompting, or RAG without temporal or goal awareness.

Minerva CQ is designed as a response to that constraint. Its stated objective is not merely to answer explicit prompts but to provide proactive, real-time support during live voice interactions. The proposed shift is from passive agent-assist to continuous intent tracking, workflow guidance, and context maintenance. This framing places Minerva CQ within a broader transition from prompt-response assistance to systems that observe, infer, decide, and intervene while the conversation is still unfolding.

## 2. Pipeline architecture and control loop

The system is implemented as a modular pipeline spanning the full call lifecycle [2509.12589]. The four main phases are Call Initiation, Problem Discovery, Resolution Guidance, and Call Closure. During Call Initiation, the caller is identified, CRM context is retrieved, and entity recognition extracts key customer IDs from the transcript. During Problem Discovery, the system performs continuous real-time intent detection and contextual analysis to infer customer goals. During Resolution Guidance, it triggers proactive workflows, performs context-aware query generation and knowledge retrieval, and uses an FAQ cache for validated, low-latency responses. During Call Closure, it generates an automatic final summarization and performs sentiment and compliance checks, including PII redaction.

The architecture also includes support modules: customer profiling for sales calls, an FAQ cache for rapid answers, and an agent guidance dashboard for live sentiment, CSAT/NPS, and related signals. The human agent remains central to the system. The paper describes Minerva CQ as operating through a continuous loop:

$$
\text{Observe} \rightarrow \text{Understand} \rightarrow \text{Decide} \rightarrow \text{Act} \rightarrow \text{Assist} \rightarrow \text{Learn}
$$

This loop formalizes the system’s dynamism. It does not simply retrieve information; it builds and updates operational state over time, with bidirectional interaction between the agent and context-aware AI modules.

## 3. Core modules and technical functions

Minerva CQ integrates real-time transcription, intent and sentiment detection, entity recognition, contextual retrieval, dynamic customer profiling, and partial conversational summaries [2509.12589]. The real-time transcription module supports multilingual speech-to-text for English, Italian, Hindi, “Hinglish”, and more. It also handles code-switching by dynamically transcribing mixed-language utterances without added latency. Text normalization converts transcript data to agent-friendly English for non-Hindi-reading agents, allowing dashboard use independent of the original language of the customer utterance.

Entity recognition and CRM integration are used to reduce repetitive verification. Live entity parsing extracts identifiers such as name, email, and account number from spoken input, and CRM integration fetches and updates customer data in real time. Intent and sentiment detection run continuously: streaming intent classification monitors conversation for intent cues, while sentiment tracking infers customer affect and exposes live indicators on the agent dashboard. The same layer supports predictive scoring for CSAT/NPS during and after the call.

Knowledge access is structured around AI-suggested query generation and contextual retrieval. The system rewrites under-specified customer requests into actionable, KB-compatible queries such as “How to activate a travel plan?”. RAG is used selectively only for new or complex queries. Frequently asked and answered questions are instead served by an FAQ cache, where validated entries are returned in less than 0.5 seconds, compared to 5–9 seconds for RAG retrieval via OpenSearch. FAQ entries are continuously checked through heuristics and ontology before being added, with the stated purpose of ensuring accuracy and avoiding KB drift.

Two additional modules extend the system beyond retrieval. Dynamic customer profiling is applied in sales contexts by detecting expressions of interest, hesitation, or intent and building a lightweight profile throughout the conversation; suggested actions and product offers are then tailored to that evolving profile. Partial and final conversational summarization provide persistent context. After each conversational turn, the system generates a streaming partial summary capturing facts, intent, and decisions for use by both humans and downstream modules. At call end, a final summary records primary intent, resolution path, agent actions, and sentiment trajectory, with PII redacted for compliance and documentation.

## 4. Real-time interaction pattern during calls

A Minerva CQ-augmented call begins with entity and context recognition plus CRM retrieval, so the agent does not need to ask redundant questions [2509.12589]. As the customer describes the issue, intent recognition and profiling modify the dashboard live. The system then suggests next-best actions, including contextually auto-generated “click prompts”, and retrieves answers by first consulting the FAQ cache and escalating to RAG if needed. Sentiment monitoring exposes mood and CSAT shifts, enabling real-time adjustment for de-escalation or upselling. The call ends with an automated compliant summary that is produced and documented, reducing after-call work.

The system’s adaptivity is explicitly tied to conversation state. Agents see evolving suggested actions as new intents or sentiments are detected; for example, when tone turns negative, guidance may shift toward de-escalation workflows. Partial summaries preserve prior facts and decisions for all modules, functioning as a shared context layer and reducing the need to repeat or re-ask information. The paper also notes live FAQ validation: the system learns new FAQs from conversations, validates them, and adds them to the cache over time.

This interaction model is central to the paper’s distinction between Minerva CQ and reactive agent-assist systems. The latter typically provide isolated suggestions after an explicit query, whereas Minerva CQ is described as continuously observing the call, updating contextual state, and changing its assistance strategy as the conversation progresses.

## 5. Production deployment and quantitative evaluation

The case study reports live production deployment and an A/B test with 100 agents: 50 with Minerva CQ and 50 controls without AI assist [2509.12589]. Both groups handled approximately 40,000 production voice calls, normalized for routing and products, and KPIs were measured from CRM and call logs.

| KPI | Control / Baseline | Minerva CQ / Relative change |
|---|---|---|
| Average Handling Time (AHT) | 4m 43s | 2m 55s; ↓ 38% |
| Lead-to-Enquiry (L2E) Conversion | baseline | +33% |
| Booking Conversion | baseline | +4.8% |

The evaluation is summarized in terms of difference of means, expressed as relative change percentages:

$$
\text{Relative Gain} = \frac{(\text{AI metric} - \text{Control metric})}{\text{Control metric} \times 100\%}
$$

For AHT reduction, the extract gives the worked example:

$$
\text{AHT\_R} = \frac{2m\,55s - 4m\,43s}{4m\,43s} = -38\%
$$

Additional system-level measurements concern retrieval latency and operational savings. FAQ cache response time is reported as less than 0.5 seconds, versus 5–9 seconds for RAG retrieval via OpenSearch. Approximately 7,000 out of 10,000 pilot queries were answered from the FAQ cache, with cumulative latency savings of about 11.7 hours of system or agent wait time. The paper also reports operational cost reduction, attributing it to lower agent time per call and reduced LLM API costs.

The multilingual evaluation is operational rather than benchmark-oriented. Minerva CQ handled mixed-language scenarios such as Hindi/English conversations and normalized them to English for dashboard display, thereby supporting agents regardless of reading ability in the source language.

## 6. Feature-to-outcome mapping, limitations, and research significance

The paper provides an explicit feature-to-KPI mapping [2509.12589]. Entity recognition and CRM context reduce repeat verification and accelerate call initiation, corresponding to lower AHT. Intent recognition and workflow triggering provide proactive guidance and consistent resolution, corresponding to lower AHT and higher FCR. AI query generation with FAQ cache reduces search and manual typing, providing validated quick responses and supporting lower AHT and more consistent CX. Partial and final summarization reduce after-call work while supporting compliance through PII handling.

A common misunderstanding would be to treat Minerva CQ as a conventional RAG layer attached to ASR. The paper argues against that characterization. RAG is used selectively only for new or complex queries, while the system’s distinctive behavior lies in the integration of partial summarization, dynamic workflow triggering, live sentiment and intent detection, and FAQ caching. This suggests that the contribution is as much about online state maintenance and action selection as about retrieval quality.

The evaluation is not presented as free of caveats. The paper notes that some temporal confounders remain in the pilot, including seasonality and agent skill differences. It also identifies future work in expanding standard voice-centric benchmarks and further automating FAQ cache governance. Within those limits, the case study positions Minerva CQ as a deployable architecture for real-time agent assistance in voice-based customer support: a modular yet connected system that links transcription, inference, retrieval, workflow guidance, profiling, and summarization to measurable operational changes in production.

Source: https://www.emergentmind.com/topics/minerva-cq