How I AIHow this CEO turned 25,000 hours of sales calls into a self-learning go-to-market engine
CHAPTERS
- 0:00 – 3:01
The real GTM problem: sales can’t find what customers care about
Matt explains the recurring complaint from his sales team: they lacked a reliable way to find customer insights, use cases, and messaging by segment. Claire frames the breakthrough: Suzy already had an enormous “source of truth” in recorded customer calls.
- •Sales and CS struggled to locate relevant customer information and patterns
- •The gap wasn’t effort—it was retrieval, organization, and repeatability
- •25,000 hours of calls represent high-signal, real customer language and needs
- •Goal: turn call data into an always-on GTM intelligence system
- 3:01 – 4:32
Why Zapier became the backbone (and why leaders must build)
Matt shares his background as a technical, non-coding operator and how Zapier let him stitch tools together long before LLMs. Both he and Claire argue that executives can’t delegate “AI” to engineering as a vague directive—leaders need hands-on capability to steer outcomes.
- •Zapier enabled multi-tool workflows without being an engineer
- •AI inside Zapier expanded what automations could do dramatically
- •Building on the “edges” of the org is an effective entry point to AI
- •Hands-on leadership improves judgment, speed, and cost awareness
- 4:32 – 6:03
Start with the core constraint: problem-first AI, not tool-first play
Matt lays out his approach: step back from the flood of tools and identify the single business constraint that blocks growth. For Suzy, the constraint was customer knowledge not being accessible and actionable for sales and customer success.
- •AI experimentation fails when it’s not anchored to a business bottleneck
- •Define what’s holding growth back or what opportunity is being missed
- •Translate the problem into a data question: what data would solve this?
- •Commit to a focused build once the core problem is clear
- 6:03 – 9:56
The data unlock: Gong transcripts as the ultimate source of truth
Remote work and Gong recording created a massive, underused asset: high-quality transcripts of real customer conversations. Matt describes building an operating system on top of both historical and new calls to stay responsive in real time.
- •Discovery: Gong transcripts were strong and plentiful—25,000 hours
- •Transcripts capture “customers in the wild” describing needs in their language
- •The system must serve past learning and real-time responsiveness
- •Automations run on every new completed call, not just batch analysis
- 9:56 – 14:02
Bootstrapping the trigger: hacking a Gong-to-Zapier feed with Browse.ai
Matt walks through the hardest step: creating a reliable trigger when a new call completes and obtaining a unique call ID. Because Gong didn’t expose what he needed directly, he used Browse.ai to scrape the transcript page using the call ID appended to a URL pattern.
- •Identify call ID as the key variable across Gong transcript URLs
- •Create a trigger when a new call occurs, then scrape the transcript page
- •Use Browse.ai to run a scripted scrape and return the raw transcript
- •Persistence matters: once the trigger hurdle is cleared, the rest accelerates
- 14:02 – 17:43
Processing the transcript: delays, HTML cleanup, and enrichment lookups
After the scrape completes, the Zap adds reliability and structure: short delays to prevent race conditions, formatting to strip HTML, and lookups to enrich context beyond what Gong provides. Matt also explains why he prefers sequential automation logic and how this differs from agents.
- •Add buffer delays to reduce errors when many tasks run quickly
- •Strip HTML and normalize transcript text for downstream analysis
- •Enrich missing context via Google Sheets and other lookup sources
- •Use Slack as a directory/lookup layer (e.g., emails) without posting messages
- •Mental model: deterministic automations first, then move toward agents
- 17:43 – 23:26
LLM step 1: a standardized call summary + sentiment score that predicts churn
Matt introduces the “Core Summary Generator” prompt: a structured summary that extracts participants, purpose, outcomes, sentiment, next steps, and rep performance notes. The key metric is a 1–10 sentiment score, which Suzy benchmarks against churn and upsell outcomes.
- •Use a consistent, structured prompt to summarize calls (exclude small talk)
- •Generate a sentiment score and track it over time as a health indicator
- •Sentiment trends are predictive of churn vs. expansion
- •Model selection is pragmatic: stable, cost-effective, and fast beats novelty
- •Prompt includes what to produce plus evaluation criteria and next steps
- 23:26 – 26:54
Real-time visibility: Slack call intel feed + churn early warning channel
The workflow posts summaries into a dedicated Slack channel, giving leadership and teams continuous customer pulse without manual writeups. Low sentiment calls route into a churn early warning stream, enabling intervention even when a rep might otherwise downplay or delay escalation.
- •Auto-post a standardized call recap into a company Slack feed
- •Leadership can monitor customer reality at scale (300-employee org context)
- •Route low sentiment into a churn alert channel for proactive response
- •Acknowledge false positives and calibrate with rep context
- •Eliminate reliance on manual reporting of “how the call went”
- 26:54 – 28:34
Marketing loop: extract customer language into Google Ads keywords automatically
Matt shows how transcripts become paid growth fuel: an LLM analyzes the call for areas of interest and outputs keyword ideas to bid on. These keywords are automatically added to Google campaigns, creating a closed loop from customer conversations to acquisition targeting.
- •Use customer calls to capture high-intent language and problem framing
- •LLM outputs keyword lists aligned to what prospects actually say
- •Automation pushes keywords into Google Ads campaigns
- •Turns GTM conversations into continuous market research and demand gen input
- •Reduces dependency on a dedicated analyst monitoring every call
- 28:34 – 31:27
AI enablement for reps: call coaching + follow-up email drafts (human-in-the-loop)
Beyond insights, the workflow directly upgrades execution: reps receive post-call coaching feedback and an editable follow-up email draft. Matt emphasizes keeping a human in the loop for outbound messages to prevent context mistakes and preserve judgment.
- •Generate individualized coaching notes: what went well and what to improve
- •Aggregate feedback into a dataset to spot patterns and support reviews
- •Draft follow-up emails immediately after calls to save rep time
- •Human-in-the-loop: reps edit before sending to avoid misfires
- •AI “democratizes” coaching quality independent of manager strength
- 31:27 – 35:25
From single-call actions to an aggregate customer intelligence database (and CRM implications)
Matt describes structuring call-derived attributes into a database (role, product interests, trends) that can power retrieval and prep for future calls. He and Claire discuss how this starts to resemble (and potentially surpass) traditional CRM workflows by auto-generating structured context without manual data entry.
- •Create a structured customer profile database from each call
- •Enable queryable insights by segment (e.g., role + industry patterns)
- •Structured data improves downstream retrieval/RAG-style use cases
- •Salesforce still exists, but the automation begins to replace core CRM value
- •Core thesis: applications matter less than the data foundation
- 35:25 – 39:17
The controversial automation: redacted blog posts + SEO assets from call transcripts
Matt reveals a heavily-tested system that turns calls into publishable marketing content while removing identifying details to protect confidentiality. The workflow delays publishing (e.g., 21 days), generates headlines/graphics/CTAs, scales to thousands of posts, and feeds paid traffic via dynamic search ads.
- •Redact all identifying customer/company details to avoid confidentiality breaches
- •Generate SEO-optimized blog posts reflecting real use cases discussed on calls
- •Publish on a delay to reduce risk and add an extra safety buffer
- •Scale content production dramatically (thousands of posts) without human labor
- •Use the content for both organic discovery and Google Dynamic Search Ads
- 39:17 – 42:53
Org design shifts: the rise of the “super IC,” automation orchestrators, and guardrail prompting
In closing, Matt explains how automation changes hiring: fewer order-takers, more proactive builders with hands-on learning mindset. He proposes new roles—GTM automation “general contractors”—and shares his prompting method: define the goal, set guardrails (what not to do), then iteratively refine.
- •Hiring priority shifts to ambitious, proactive individual contributors
- •“Super ICs” become disproportionately valuable in AI-native orgs
- •New role concept: GTM orchestrators managing automation blueprints
- •Outputs should be owned by functional teams (marketing owns blog performance)
- •Prompting technique: intent + guardrails (“do not do”) + iterative refinement
