Voice Interaction Assistant

AI-Powered Client Conversation Capture, Summarization, and Task Generation

Why the Voice Interaction Assistant is Required

Relationship Managers (RMs) often spend significant time capturing, reviewing, and documenting client conversations. Key challenges include:

• Fragmented Recording Process: Conversations may exist in multiple formats — live calls, uploaded audio, or transcripts.
• Uploaded Audio Never Transcribed: Uploaded audio has no transcription path today — it stays at status ‘Pending Transcription’ indefinitely.
• Manual Summarization: Reviewing and extracting action points manually is time-consuming.
• Missed Follow-Ups: Without automated alerts and structured outputs, tasks and client commitments can be overlooked.
• Compliance & Audit Risk: Audio and transcripts must meet SEBI, DPDP, and RBI guidelines for retention and audit.

The Voice Interaction Assistant automates transcription, AI-based summarization, task generation, and sentiment analysis — ensuring RMs can focus on client engagement while maintaining compliance and full auditability.

• Capturing multi-channel client conversations efficiently.
• Producing accurate, actionable summaries from audio.
• Generating follow-up tasks in real time.
• Maintaining compliance with SEBI, DPDP, and RBI regulations.
• Scaling conversation captures across hundreds of clients.
• Transcribing both recorded and uploaded audio accurately, including Hinglish code-switching, without translating away the verbatim record SEBI expects.
• Bridging Salesforce and the external Sarvam STT service securely, without exposing the Sarvam API key.
• Handling short and long audio differently so long recordings never block the RM's UI.
• Summarizing transcripts that exceed both Salesforce field limits and AI prompt template token limits.
• Keeping Interaction access consistent with the existing FSC sharing model.

The assistant acts as an intelligent conversation manager:

• Captures audio live in-browser, or via uploaded audio or transcript files.
• Uses Sarvam AI (Saaras v3) for transcription, supporting Hinglish (Indian English-Hindi).
• Stores audio securely in AWS S3 (ap-south-1 Mumbai), with KMS encryption, versioning, and Object Lock for 7-year retention.
• Processes short clips (≤30 sec) synchronously and long recordings asynchronously via S3 event triggers.
• Generates AI summaries, action points, and sentiment Analysis via Agentforce / Einstein GPT prompt templates.
• Creates Tasks for each action point in Salesforce standard Activity timeline.
• Provides secure audio playback via presigned URLs.
• Fully compliant with SEBI, DPDP, RBI, and FSC sharing model guidelines.

The Voice Interaction Assistant lets a Relationship Manager capture a client conversation — by recording live in the browser or uploading an existing audio file — and turns it into a stored transcript, an AI summary, action points, and sentiment on the standard FSC Interaction object. The same Lightning Web Component (voiceInteractionManager) runs on Lead, Opportunity, and Account record pages; it passes recordId and objectApiName, and the Apex controller routes the parent lookup. Audio never traverses Apex to Salesforce storage — it is uploaded directly to AWS S3, and only the S3 object key is retained on the Interaction record.

Each capability is delivered through the following technologies:

• Dual-Path Transcription: The LWC measures clip duration as recording stops or a file is selected. Clips ≤30 seconds upload to voice-interactions/audio-short/ and explicitly call the Lambda's /voice endpoint for synchronous transcription; clips >30 seconds upload to voice-interactions/audio/ and rely on an S3 ObjectCreated event to trigger asynchronous batch transcription. A transcript-upload skips both Sarvam paths and goes straight to summarization.
• AWS Integration Layer: An API Gateway HTTP API exposes two routes — /voice for Salesforce and /voice-callback for Sarvam — both invoking the voiceTranscribe Lambda (Node.js 22.x, 5-minute timeout), which handles four invocation patterns: presign, sync transcribe, S3 event, and Sarvam callback.
• Secure Service-to-Service Bridging: Sarvam never receives a direct call from Salesforce; the Lambda is the integration boundary and holds the Sarvam API key in AWS Secrets Manager. Two team-generated secrets — API_SHARED_SECRET and SARVAM_CALLBACK_TOKEN — authenticate Salesforce-to-Lambda and Sarvam-to-Lambda calls respectively, and the Lambda writes results back to Salesforce via OAuth 2.0 Client Credentials Flow against a dedicated Connected App.
• Audio Storage & Retention: AWS S3 in ap-south-1 Mumbai is the production system of record for audio, with SSE-KMS encryption, Object Lock enforcing a WORM retention period, versioning for recovery, and a lifecycle policy that moves audio to Glacier after 90 days and deletes it after 7 years to meet SEBI retention.
• Presigned URL Playback: Because the bucket is private, playback is served through short-lived S3 presigned URLs (1–2 hour validity) generated either by the Lambda via the AWS SDK or by Apex via SigV4 signing — the audio bucket is never made public and audio is never streamed through Apex.
• Large Transcript Handling (Map-Reduce): Transcripts beyond the 131,072-character field limit are stored in full in S3 with only a preview kept in Salesforce. For summarization, oversized transcripts are split into chunks on speaker/sentence boundaries, each chunk is summarized in a MAP phase, and the mini-summaries are combined in a REDUCE phase into one overall summary, action points, and sentiment — all inside GenerateSummaryQueueable.
• Data Model & Status Workflow: New fields on the standard Interaction object (Audio_S3_Key__c, Transcript_S3_Key__c, Language_Detected__c, Sarvam_Request_Id__c, STT_Provider__c, Raw_AI_Output__c, Stage__c, Error_Message__c, Audio_Storage_Location__c) carry the pipeline state through a six-value status lifecycle — Uploading, Transcribing, Summarizing, Complete, Needs Review, and Failed — with no new custom object required.
• Compliance Mapping: A consent gate captures Consent_Timestamp__c before recording for DPDP; all AWS resources sit in ap-south-1 Mumbai and Sarvam runs on Indian infrastructure for data localisation; S3 Object Lock, versioning, and CloudTrail provide the SEBI audit trail; and Interaction sharing continues to inherit from the parent record under the existing FSC sharing model.
• Architecture Principle — Storage Decoupled from Synthesis: Storage is decoupled from synthesis, and Sarvam is never called directly by Salesforce. Audio is uploaded straight from Apex to S3 and becomes the durable system of record first; the AWS Lambda is the sole integration boundary that talks to Sarvam, holding its API key in Secrets Manager and bridging Salesforce and Sarvam over authenticated, server-to-server calls. Only once a transcript exists does Agentforce run — entirely server-side from Apex — to generate the summary, action points, and sentiment, with each action point written back as a governed Salesforce Task. This keeps every conversation retained, auditable, and India-resident regardless of when or how the AI processing completes.

• Multi-Mode Capture: Live recording, file upload, transcript upload.
• AI Transcription: Sarvam AI handles short and long audio, including Hinglish support.
• Automatic Summarization: Generates structured summary of conversation, including actionable tasks.
• Sentiment Analysis: Classifies interaction tone as Positive, Neutral, or Negative.
• Task Generation: Each action point is converted into a Salesforce Task linked to the conversation.
• Secure Storage & Playback: Audio stored in S3; playback via short-lived presigned URLs.
• Compliance & Audit: Full audit trail, 7-year retention, and DPDP/SEBI/RBI adherence.
• Scalable for Large Volumes: Handles long transcripts with map-reduce summarization for AI prompt limits.

• Faster Client Documentation: Summaries and tasks generated automatically, reducing RM effort.
• Proactive Follow-Ups: Action points surfaced in real-time for timely engagement.
• Enhanced Compliance: Audio retention, audit logs, and transcript traceability meet regulatory standards.
• Improved Advisor Productivity: RMs focus on client conversations instead of administrative work.
• Scalable Operations: Can handle short clips and long recordings efficiently, including high-volume calls.

• Time Savings: Reduces manual review and summarization effort.
• Actionable Insights: Immediate generation of tasks and next steps for client engagement.
• Consistency & Accuracy: AI-driven summarization and sentiment analysis ensure reliable outputs.
• Compliance Assurance: Full adherence to SEBI, DPDP, RBI, and FSC guidelines.
• Secure Storage: Audio safely stored in India-resident AWS S3 environment with encryption and versioning.
• Improved RM Effectiveness: Enables focus on client advisory rather than documentation.
• Scalability: Supports multiple RMs and high-volume interactions simultaneously.