Summary of chat #CS-2847 (18 min) ready for review. Customer actions: bug on feature X, deadline Y, upgrade Z. Decisions: bug resolved with workaround, deadline Y for 30 May, upgrade Z to be formalised with addendum within 7 days. Follow-up: addendum → Carlo (7 days); adoption call → Marta (next week). Profile: new operations manager. Confirm or edit before saving to CRM?
AGENT · CONVERSATION SUMMARIZER
Every conversation closes with the summary ready in the CRM.
Conversation Summarizer activates at the end of a customer conversation (web chat, email thread, recorded call). It extracts actions, decisions, and follow-ups from the conversation, populates the CRM with the structured trace, and proposes next steps to the operator. The operator confirms; the summary enters the system without manual writing.
Conversation Summarizer at work.
All correct. Go ahead.
CRM updated: ticket closed, two tasks assigned, customer profile updated.
Why it exists.
Writing the conversation summary in the CRM is what customer service and commercial teams do after every interaction: what the customer asked, what the operator answered, what actions were agreed, what follow-up is needed. It is repetitive work, takes minutes per conversation, and does not scale with volume.
How it extracts the four elements.
The agent activates at the end of the conversation. It reads the full transcript (chat from the live chat system, email thread, call audio after transcription with a dedicated service) and extracts four structured elements: customer actions, decisions taken, required follow-ups with owner and deadline, and customer profile updates that emerged.
The operator confirms.
The operator remains the author of the customer conversation and is responsible for the relationship. They see the summary on the work channel, review it, confirm or edit. The confirmed version enters the CRM automatically. Review takes one to two minutes against five to ten for manual writing.
The functions that handle high volumes of customer conversations.
Customer service and commercial operator
Recovers time spent writing summaries. A twenty-minute conversation — which typically generates five to ten minutes of post-call writing — closes with one to two minutes of draft review.
Cross-channel · 7d 1,420 mentions
Site chat NPS 7.2
Reviews top theme: shipping
Social · IG sentiment +0.4
Brief to the brand owner
Customer service manager
Sees CRM summary quality improve structurally: consistent formats, complete fields, tracked follow-ups. Customer service metrics become measurable with confidence.
Claim type Auto liability
Severity Medium
Adjuster Dr. Ferrari · Northeast
Initial reserve € 4,200
Routed ✓
Account manager picking up the contact
Finds the history of previous conversations already structured. Context is recovered in seconds, not in ten minutes of CRM reading. Customer relationship continuity no longer depends on the individual operator's memory.
87 Tone too formal — try a direct opener
71 Long message: summary in the first 2 lines
94 Excellent — consistent with the customer history
Eighteen minutes of chat, two minutes of review, CRM populated.
Eighteen minutes on three different topics.
For the customer service team at a B2B SaaS company, a chat conversation with a customer lasts eighteen minutes. The customer reports a product feature issue, asks for clarification on a renewal deadline, and confirms an upgrade decision discussed the previous week. The operator resolves the issue, confirms the deadline, and notes the upgrade.
Four structured elements from the transcript.
At the end of the chat, the agent activates. Customer actions: technical support, administrative information, commercial confirmation. Decisions taken: issue resolved with workaround, deadline confirmed, upgrade to be formalised with addendum. Follow-up: account manager prepares the addendum (7 days); customer success specialist schedules the adoption call (next week). Profile: new operations manager.
Two-minute review, ticket closed.
The operator reviews the summary on the Slack channel, confirms with minor edits. The CRM is populated: ticket updated, tasks assigned, customer profile updated with the new manager. The trace of every summary stays in the runtime audit registry.
Declarative summary schema, written by the customer's customer service team.
Conversation Summarizer's rules are declarative. The customer service team defines in readable format the summary schema (which fields, in which order), the mapping rules for action and decision types, and the follow-up assignment rules. Rules live in the customer's repository, versioned, and validated at agent startup. HubSpot tools are native for CRM population (contacts, deals, tickets, meetings, tasks, notes). For other CRMs (Salesforce, Microsoft Dynamics, proprietary systems), integration is built during delivery. Audio transcription for calls is not a built-in capability: it is configured with an external service (Whisper API or a dedicated transcription service) during delivery, with audio data handling following the customer's GDPR constraints. Integration with company live chat systems (Salesforce Service Cloud, Zendesk, proprietary systems) is also built during delivery.
- Language
- TypeScript (Node.js)
- LLM model
- customer's choice: Anthropic, OpenAI, Mistral, open source models hosted internally, AWS Bedrock for a private model
- Built-in controls used
- pii-detector, credential-detector, message-length-limit
- Native channels
- Telegram, Slack, WhatsApp, OpenAI-compatible HTTP
- Live chat system integration
- dedicated adapter built during delivery (Salesforce Service Cloud, Zendesk, proprietary systems)
- Call audio transcription
- not a built-in capability; external service (Whisper API or equivalent) configured during delivery
- CRM integration
- HubSpot native (contacts, deals, tickets, meetings, tasks, notes); other CRMs in delivery
- Memory
- persistent per instance, pgvector + PostgreSQL FTS on customer context
- Registry
- append-only, queryable with a standard SQL client
Frequently asked questions about the agent.
No. The operator remains the author of the customer conversation and is responsible for the relationship. The agent reads the transcript at the end of the conversation, extracts the structured data, and prepares the summary. The operator reviews and confirms. Review typically takes one to two minutes against five to ten for manual writing.
Audio transcription is not a built-in capability. It is configured with an external service (Whisper API or a dedicated transcription service) during delivery by the Exelab team. For calls with customer consent (B2B commercial calls, authorised technical support calls), the transcript is processed and managed following the customer's GDPR constraints. Audio data retention follows the customer's policies.
HubSpot tools are native (contacts, deals, tickets, meetings, tasks, notes). For other CRMs (Salesforce, Microsoft Dynamics, proprietary systems), integration is built during delivery by the Exelab team. The typical pattern is writing the operator-confirmed summary via the CRM's API.
The agent reads the full thread message by message, extracting the evolution of the conversation (intermediate decisions, position changes, progressive agreements). The final summary traces the evolution, not just the last exchange. For very long threads (over twenty messages), the agent can produce a multi-level summary: a phase-by-phase summary and an overall summary.
From a 30-minute conversation to the squad in production.
A 30-45 minute conversation to understand how Conversation Summarizer would configure to the customer's channels. Which conversation channels, which summary schema, which target CRM.