Avoma Notes vs Claude Extraction: Comparison Analysis

Comparing 5 calls across both systems to identify overlap, gaps, and integration opportunities.

1. Pain Points Overlap

Summary

Both systems extract pain points, but they focus on different granularities and sometimes surface entirely different items.

Brooks/Dropbox (2bbbfa5d)

Verdict: Strong overlap on 3/4 Claude items. Avoma catches one pain point Claude missed (certificate rotation burden). Claude captures the eBPF performance history as a pain point; Avoma files it under Risk -- arguably Risk is a better classification for a historical concern.

Starbucks (5facc469)

Verdict: Near-complete overlap. The main difference is where each system classifies things. Claude puts the Falcon collector issue under competitiveMentions; Avoma puts it under Pain Point. Both are valid.

Dropbox POC (5dd8e88a)

Verdict: Very strong overlap. Where they differ is classification: Claude files "external event ingestion" as a feature request; Avoma calls it a pain point. Both are defensible.

Airbnb (d5718c90)

Verdict: Weak overlap. The two systems are extracting from different moments in the conversation. Claude captures the strategic pain points (third-party risk, procurement); Avoma captures the tactical/immediate ones (Elastic integration, performance concerns). Claude's pain points are more aligned with what a sales team needs.

Wunderkind (b02a2f17)

Verdict: Moderate overlap. Each system finds ~3 items; they share 1-2 in common. Claude is more specific about UI/UX friction; Avoma captures broader deployment state issues.

Cross-Call Pain Point Assessment

• Both systems find the core pain points in most calls (3/5 strong overlap)
• Classification differs: Avoma sometimes files things under Risk or Background that Claude puts under painPoints, and vice versa
• Claude is more specific: Includes verbatim quotes and speaker attribution, making extracted pain points more actionable for sales follow-up
• Avoma catches operational/historical context Claude sometimes misses (certificate rotation, team departures as pain vs. background)

2. Feature Requests Overlap

Summary

Avoma's "Feature Mention" section conflates features of Qpoint being demoed with features requested by the prospect. Claude's extraction focuses specifically on what the prospect asked for.

Brooks/Dropbox

Starbucks

Dropbox POC

Airbnb

Wunderkind

Cross-Call Feature Request Assessment

• Avoma's Feature Mention is fundamentally different from Claude's featureRequests. Avoma captures features of the product that were discussed; Claude captures what the prospect asked for. These are very different signals.
• Claude's extraction is far more useful for product roadmap prioritization because it specifically isolates prospect demand.
• Avoma's Feature Mention is useful for tracking what was demoed -- which features the sales team presented and discussed. This could be valuable for demo optimization.
• Recommendation: Do not treat these as equivalent fields. They serve different purposes.

3. Competitive Mentions Overlap

Brooks/Dropbox

• Claude: No competitive mentions extracted
• Avoma: No competitors listed
• Match: Both agree -- no competitors discussed

Starbucks

• Claude: Falcon Log Scale Collector (negative), Humio/Log Scale (neutral)
• Avoma: Falcon Log Scale Collector (lacked data transforms), Datadog (mentioned as alternative target)
• Match: Strong overlap. Avoma additionally captures Datadog as a competitive mention -- Claude misses this because Lawrence frames Datadog as an integration target, not a competitor. Avoma's classification is arguably better here since Datadog is indeed a competitive platform.

Dropbox POC

• Claude: CrowdStrike Falcon (conflicts with BPF tools)
• Avoma: No competitors listed (CrowdStrike mentioned in Current Workflow and Risk sections instead)
• Match: Claude extracts it; Avoma does not classify it as competitive. Both capture the information somewhere.

Airbnb

• Claude: No competitive mentions
• Avoma: No competitors listed
• Match: Both agree

Wunderkind

• Claude: No competitive mentions
• Avoma: Datadog (complementary -- discovers uninstrumented traffic)
• Match: Avoma catches Datadog reference; Claude misses it. The Datadog mention is framed as complementary rather than competitive, so Claude's exclusion is defensible but Avoma's inclusion is more useful for market positioning.

Cross-Call Competitive Assessment

• Both systems have sparse competitive data (most calls don't have direct competitor discussions)
• Avoma is slightly better at capturing tools mentioned in passing (Datadog in Starbucks and Wunderkind) that Claude considers "integration targets" rather than competitors
• Neither system is comprehensive -- competitive intel extraction could be improved in both

4. Value Signals Overlap

Brooks/Dropbox

Key insight: Avoma's Impact section captures quantified business value ("hours to minutes") that Claude does not extract. The Apollo.io reference is social proof that Claude misses entirely. Claude's valueEchoes capture the prospect's own language, which Avoma lacks.

Starbucks

Key insight: Avoma's Positive Moments for this call are actually from Qpoint team members, not the prospect. This is a significant quality issue -- those are not buying signals at all. Claude correctly attributes all value signals to Lawrence Lucas (the prospect).

Airbnb

Key insight: Avoma's "Positive Moment" section for Airbnb contains zero actual positive moments from the prospect. Every item is a product feature description. Claude correctly identifies Luke's "I feel like something like this should exist" as a value signal. This is a major quality gap in Avoma's extraction.

Wunderkind

Cross-Call Value Signal Assessment

• Claude is significantly better at identifying genuine buying signals. Claude correctly attributes signals to prospects and captures the emotional weight ("should exist", "bane of my existence").
• Avoma's Positive Moment section has quality issues. In 2/5 calls, it contains statements from the Qpoint team (not the prospect) or product descriptions (not positive moments). This makes it unreliable as a buying signal source.
• Avoma's Impact section adds unique value. It captures quantified business outcomes ("hours to minutes") and broader value framing that Claude does not extract. This is genuinely useful data.
• Claude's valueEchoes are unique -- the concept of tracking when a prospect articulates value in their own words has no Avoma equivalent and is high-value for messaging validation.

5. NEW Data Only in Avoma Notes (No Claude Equivalent)

For each call, here are Avoma sections that have no corresponding field in our CallData schema:

Action Items

All 5 calls have detailed, timestamped action items with owner attribution:

Assessment: High value. Our urgency.nextSteps captures some of this but is less granular -- it records commitments, not specific deliverables with deadlines. Action items are the #1 candidate for schema addition.

Budget

Assessment: Low incremental value. Only 1/5 calls has budget data, and it is the price we quoted, not a prospect budget signal. Our budgetSignals field is actually more useful because it captures the prospect's reaction to pricing.

Decision Making

Assessment: Medium-high value. Understanding who has authority and how decisions are made is critical for deal progression. We do not capture this. 3/5 calls have useful decision-making data.

Current Workflow

Assessment: Medium value. Understanding the prospect's current stack helps with positioning and technical fit assessment. Our company.description captures some of this but not the detailed technical workflow. This would be most useful during deal preparation rather than post-call analysis.

Background

Assessment: Medium value. Deal history and organizational context. Some of this lands in our summary field but not systematically. Most useful for account management rather than PMF analysis.

Timeline

Assessment: Medium value. Our urgency.timelineSignals captures prospect timelines but not Qpoint's own commitments (e.g., "Datadog guide by afternoon"). The prospect timeline overlap means incremental value is limited to tracking our own deliverables.

Risk

Assessment: Medium-high value. Risk captures deal blockers and technical obstacles that could prevent close. Our objections field partially covers this, but Risk is broader -- it includes technical compatibility issues (CrowdStrike/eBPF conflict) and deployment prerequisites (object store setup) that are not objections per se but are real blockers.

Follow Up

Mostly empty across calls. Not valuable as a separate field.

Custom Sections (AI Security Insights, Deployment Strategy, etc.)

Avoma generates call-specific custom sections based on content:

• Brooks: "AI Security Insights", "Product Roadmap Discussions", "Customization Needs"
• Starbucks: "Project Ownership", "Technical Standards Enforcement", "Security Data Streaming"
• Dropbox: "Retention and Storage", "Data Classification", "Integration with Infrastructure"
• Airbnb: "Deployment Strategy", "Data Classification Focus"
• Wunderkind: "Technical Needs Assessment", "Sensitive Data Scanning Interest"

Assessment: These are often the most insightful parts of Avoma's output -- they represent emergent themes from the conversation that don't fit standard categories. However, they are unpredictable (different for every call) and therefore hard to integrate into a structured schema. They are best consumed as supplementary reading, not structured data.

6. Quality Comparison

Where Claude is better:

1. Speaker attribution on value signals. Claude correctly identifies who said what. Avoma's Positive Moment section attributes statements to Qpoint team members as positive moments (Starbucks: Jon's "sounds like a dream" and Devin's "cut through the red tape"). These are not buying signals.
2. Distinguishing feature requests from feature descriptions. Claude's featureRequests captures what the prospect specifically asked for. Avoma's Feature Mention mixes product capabilities being demoed with actual requests. Example: Airbnb's Feature Mention includes "Identify unauthorized flows, such as SSNs sent to uncontracted AI providers" -- this was Qpoint describing what the product does, not Airbnb requesting it.
3. Verbatim quotes with context. Every Claude insight includes the exact quote, speaker, and confidence level. Avoma provides timestamps but paraphrases everything. For sales follow-up and messaging validation, verbatim language is far more useful.
4. Interest level and funnel stage. Claude provides a quantified interest score (1-5) and funnel classification. Avoma has no equivalent signal.
5. Value echoes. Claude uniquely captures when a prospect articulates value in their own words (distinct from general value signals). This has no Avoma parallel and is critical for messaging validation. Example: Brooks articulating two distinct value props (DLP + AI governance) with different organizational buy-in timelines.

Where Avoma is better:

1. Action items with owners and timestamps. Avoma's action items are detailed and assignable: "Send resource usage figures for large deployments to Brooks McMillin" (Devin, 21:27). Our nextSteps captures commitments but not at this granularity.
2. Decision-making authority. "Lawrence Lucas has unilateral authority to enforce network standards" is critical deal intelligence. We have no equivalent field.
3. Current Workflow detail. Avoma systematically captures the prospect's current tech stack and processes. This helps assess technical fit and identify integration opportunities.
4. Risk/blockers. Avoma's Risk section captures deployment blockers (CrowdStrike conflicts, missing object stores, custom Envoy filter requirements) that fall between our pain points and objections.
5. Pricing quoted. Avoma captures the price we offered ("$360/node/year"). Our extraction focuses on prospect reaction, not what was offered.
6. Quantified impact. Avoma's Impact section occasionally provides quantified business value ("troubleshooting time from hours to minutes") that Claude does not extract.

Where both are weak:

1. Competitive intelligence. Both systems extract minimal competitive data. Most calls have 0-2 mentions, and the context is thin.
2. Budget/procurement details. Budget sections are mostly empty in both systems. This likely reflects the calls rather than extraction quality.
3. Multi-stakeholder dynamics. Neither system tracks relationships between attendees or champions vs. skeptics well.

7. Recommendations

Fields to Add to Our Schema

Priority 1 -- High Value, Clear Structure:

1. actionItems -- Array of { text: string, owner: string, deadline?: string, timestamp?: string }. This is the single highest-value addition. Avoma captures these well, and they fill a gap in our data. Our nextSteps field captures intent; action items capture commitments with deliverables.
2. decisionMaking -- Object with { decisionMaker?: string, authority?: string, process?: string, approvers?: string[] }. Understanding who has authority (Lawrence at Starbucks, Cecilia at Dropbox) is critical for deal progression and not captured anywhere in our current schema.
3. risks -- Array of { text: string, category: 'technical-compatibility' | 'deployment-blocker' | 'organizational' | 'security-privacy', severity?: string }. Distinct from objections (which are prospect pushback) and pain points (which are pre-existing problems). Risks are obstacles to deployment that need mitigation. Examples: CrowdStrike eBPF conflicts, missing object store, custom Envoy filter requirement.

Priority 2 -- Medium Value, Consider Adding:

4. currentWorkflow -- Object or free-text describing the prospect's current tech stack, tools, and processes. Useful for technical fit assessment and demo preparation. Could be a simple string field on the company object alongside description.
5. pricingDiscussed -- { priceQuoted?: string, model?: string, reaction?: string }. Our budgetSignals capture prospect reaction but not what was offered. Both sides of the pricing conversation matter.

Should Avoma Notes Supplement or Replace Parts of Our Extraction?

Supplement, not replace. Here is why:

• Our Claude extraction is higher quality for the core PMF analysis signals: pain points, feature requests, value signals, interest level, and positioning. The verbatim quotes, speaker attribution, confidence levels, and structured categories make our data significantly more useful for cross-call thematic analysis.
• Avoma captures operational/deal-management data that our extraction was not designed for: action items, decision-making authority, timeline commitments from our side, and deployment risks/blockers.
• Avoma's Positive Moment and Feature Mention sections have quality issues (misattributed speakers, product descriptions labeled as positive moments, feature demos labeled as feature requests) that would degrade our data if used as a replacement.
• The right model: Use Avoma notes as a supplementary data source that feeds specific new fields (action items, decision-making, risks) while keeping our Claude extraction as the primary source for insight-level data.

Top 3 Highest-Value Integrations

1. Action Items extraction from Avoma notes into our schema. Impact: Enables tracking of deal-specific deliverables and commitments. Every call has 3-6 action items with clear owners. This data is immediately actionable and completely absent from our current extraction. Implementation: Add actionItems field to CallData, extract from Avoma notes via API or parse from markdown.
2. Decision-Making and Risk fields added to our extraction prompt. Impact: Rather than relying on Avoma for these, add them to our Claude extraction prompt. Claude would produce higher-quality output (with verbatim quotes and confidence scores) than Avoma does. The prompt already extracts urgency, positioning, and other complex structured data -- adding decisionMaking and risks sections is a natural extension. Implementation: Extend the prompt in pipeline/prompts/extract-call.md with new sections; add types to schemas/call.ts.
3. Avoma notes as supplementary context in the extraction prompt. Impact: When both transcript and Avoma notes are available for a call, include the Avoma notes as supplementary context in the Claude extraction prompt. The Avoma notes provide a "second opinion" summary that could help Claude catch things it might otherwise miss (e.g., the certificate rotation pain point in Brooks, the Datadog competitive framing in Starbucks). Implementation: If Avoma notes are available at data/raw/{uuid}.notes.md, inject them into the prompt as an appendix. Minimal prompt change; potentially significant quality improvement on edge cases.

Appendix: Call-by-Call Raw Comparison

Brooks/Dropbox -- Extraction Quality Score

Starbucks -- Extraction Quality Score

Dropbox POC -- Extraction Quality Score

Airbnb -- Extraction Quality Score

Wunderkind -- Extraction Quality Score