AI Summaries
What the AI extracts from your conversations and how summaries appear in HubSpot.
What Gets Extracted
Conduits uses Claude AI to analyze each day's conversation and extract:
- Summary — A 2-4 sentence business-relevant overview of the conversation.
- Topics — Key subjects discussed (e.g., "Series A fundraise", "Q2 marketing plan").
- Action Items — Concrete to-dos, tagged by owner (you or the contact) with deadlines if mentioned.
- Meetings — Proposed meetings with times, locations, and whether they were confirmed.
- Introductions — When either party offers to connect someone with a third person.
- Deal Signals — Any mention of money, pricing, contracts, investment, or partnerships.
- Sentiment — Overall tone: positive, neutral, negative, or mixed.
- Urgency — High, medium, low, or none — based on language and context.
- Links Shared — URLs shared in conversation with context about why they were shared.
Summary Styles
Choose your preferred style in Settings:
| Style | Description |
|---|---|
| Business | Action items, meetings, deals — executive briefing style. Default. |
| Detailed | More context, conversation flow, quoted key phrases. |
| Minimal | Just the facts: who, what, when. Shortest format. |
How Notes Appear in HubSpot
Each digest becomes a rich HTML note on the contact's timeline. The note includes:
- A header with the contact name and date
- The summary paragraph
- Sections for action items, meetings, intros, and deal signals (only if present)
- A footer with message count and priority level
Multi-Language Support
The AI handles conversations in any language Telegram supports. Summaries are always generated in English for consistency in HubSpot, regardless of the conversation language.
Cost
Conduits uses Claude Sonnet for summarization. Typical cost per digest is ~$0.002 (0.2 cents). At 50 contacts synced daily, that's roughly $3/month in AI costs — included in your subscription.
When Summarization Fails
If the AI returns an unparseable response, Conduits retries once with a stricter prompt. If it still fails, a fallback digest is created with a generic message like "12 messages exchanged. Topics could not be automatically extracted." The digest is marked as needing review.