AI Engine
This is the brain of Spear — how it thinks about your prospects, writes your emails, and learns from every interaction. If you’re curious how an AI agent goes from “here’s my product URL” to “meeting booked,” this is the page that explains it.
AI Processing Pipeline
Section titled “AI Processing Pipeline”AI Call Costs Per Prospect
Section titled “AI Call Costs Per Prospect”| AI Task | Input | Output | Estimated Cost |
|---|---|---|---|
| Product analysis | Scraped pages (~5K tokens) | Value prop + personas (~1K tokens) | ~$0.02 (one-time) |
| Prospect scoring | Prospect profile + ICP (~2K tokens) | Score + reasoning (~500 tokens) | ~$0.008 |
| Email generation | Research dossier + style profile (~3K tokens) | 3 emails + subjects (~1.5K tokens) | ~$0.015 |
| Reply classification | Email thread + context (~1K tokens) | Category + response (~500 tokens) | ~$0.005 |
| Total per prospect | ~$0.03-0.05 |
Key AI Design Decisions
Section titled “Key AI Design Decisions”Prospect Scoring is Not Keyword Matching
Section titled “Prospect Scoring is Not Keyword Matching”The LLM evaluates each prospect against ICP criteria with contextual reasoning. It understands nuanced signals:
- “This company just raised Series A and is hiring SDRs” → scaling sales, likely needs tooling
- “CTO just posted about migrating to microservices” → technical decision-maker, likely evaluating dev tools
- “Company has 12 employees and just launched on Product Hunt” → early-stage, founder-led, matches ICP perfectly
Voice Matching
Section titled “Voice Matching”Before generating emails, the AI analyzes the founder’s existing writing from:
- Gmail sent folder (with permission)
- Website copy
- Any provided sample emails
This creates a “style profile” that ensures generated emails sound like the founder, not like an AI.
Cross-Customer Learning
Section titled “Cross-Customer Learning”Subject line optimization and message template selection use anonymized aggregate data from all customers. The system learns:
- Which subject line patterns get opens in which industries
- What personalization approaches convert vs. get ignored
- Which objection-handling tactics lead to meetings
- What send time/day patterns maximize response rates
Factual Verification Layer
Section titled “Factual Verification Layer”Before including any company-specific claim in an email, the AI verifies it against source data. If confidence is low, it uses a generic personalization approach rather than risk a factual error.
AI Voice Detection
Section titled “AI Voice Detection”Generated emails are run through a separate check to detect AI-sounding patterns. Emails that score as “obviously AI-generated” are regenerated with different approaches. The founder’s reputation is on the line — one robotic email can damage trust permanently.