How I AIHow this CEO turned 25,000 hours of sales calls into a self-learning go-to-market engine
CHAPTERS
Sales team can’t find answers—25,000 hours of calls become the source of truth
Matt Britton explains the core pain: sales and CS teams lacked easy access to customer insights and patterns. The breakthrough was realizing the company had amassed ~25,000 hours of recorded Gong calls, creating a high-signal dataset for a self-learning go-to-market engine.
- •Sales/CS struggle to find customer interests, use cases, and messaging by segment
- •Recorded customer calls are the most authentic “in the wild” voice-of-customer data
- •Shift from ad-hoc notes to a systematic operating system around call data
- •Focus on being responsive to what’s happening now, not just historical analysis
Why Zapier became the automation backbone (and the AI tipping point)
Matt shares why Zapier works for him as a non-coder who likes stitching tools together. Once Zapier integrated AI/LLMs, it expanded from simple glue automation into an orchestration layer for sophisticated GTM workflows.
- •Matt’s technical curiosity without being an engineer (ads, keywords, early platforms)
- •Zapier enables multi-tool workflows without custom engineering
- •AI in Zapier dramatically increases leverage and scope of automations
- •Preference for sequential, step-based thinking vs. branching-style builders
Start with the business problem, then choose tools and data
They emphasize avoiding “tool wandering” and instead identifying the single constraint holding growth back. Once the problem is clear, the right data sources and automations become obvious—calls were the missing link.
- •AI adoption fails when teams play with tools instead of solving a core constraint
- •Define what’s blocking growth or where the opportunity is
- •Once the problem is named, become “tunnel-visioned” on solving it
- •Data selection matters more than the app layer
CEO hands-on building: the leadership skill shift in the AI era
Matt and Claire argue leaders must build and understand workflows themselves, not just delegate to engineering. Hands-on automation teaches practical AI fluency, improves judgment, and reduces dependency on “black box” estimates.
- •Delegating “build AI” to engineering alone often goes nowhere
- •No-code/automation tools are an accessible entry point for executives
- •Car-repair analogy: understanding basics prevents bad tradeoffs and waste
- •Hands-on work builds vocabulary (JSON, APIs) and confidence to push further
Triggering on a new Gong call: hacking the call ID feed with Browse AI
Matt walks through the hardest early step: getting a reliable trigger that captures each new call and its unique ID. Because Gong didn’t expose an easy path, he used URL patterns plus Browse.ai scraping to pull transcripts programmatically.
- •Create a trigger when a new call completes and retrieve its call ID
- •Recognize Gong transcript URLs differ only by call ID
- •Use Browse.ai to visit the transcript page and scrape the raw transcript
- •Persistence through the first “hard step” unlocks everything downstream
Cleaning and enriching transcripts: delays, formatting, and lookups
After scrape completion, the automation buffers with a short delay to prevent errors, then cleans HTML and normalizes text. Matt enriches the record by pulling additional context from Google Sheets and other systems to “round out” missing data.
- •Add a 1–2 minute delay to avoid race conditions and broken tasks
- •Remove HTML/noise so LLM analysis isn’t confused
- •Use Google Sheets lookups to connect IDs → brand → owner/supervisor context
- •Think of automations as a journey where you collect needed “supplies” via enrichment
LLM selection and operations: model choice, stability, and cost tradeoffs
Matt explains how he chooses models pragmatically: don’t change what works, but test outputs in ChatGPT across models for speed and quality. He also flags the operational challenge of maintaining many automations without perfect handoffs and documentation.
- •Prefer “latest stable” versions when possible, but keep working systems intact
- •Older models can be faster and less error-prone; test for best output per cost
- •Zapier exposes multiple model/version options with different compute costs
- •Operational hygiene becomes important as automation count grows
Core Summary Generator: call overview, sentiment scoring, and next steps
A central LLM prompt transforms each transcript into an actionable brief: participants, objectives, outcomes, sentiment, what went well, improvement areas, and next steps. The sentiment score becomes a measurable signal that can be benchmarked against churn and expansion outcomes.
- •Summarize purpose, main topics, and outcome; exclude small talk
- •Assess customer sentiment and provide a 1–10 sentiment score
- •Identify what the rep did well and what could be improved
- •Extract concrete next steps and stakeholders for follow-through
Real-time Slack visibility: company-wide call intelligence + churn early warnings
The workflow posts summaries into Slack to create a live feed of customer reality—useful even for a 300-person company. Low sentiment scores route into a churn early-warning channel so leadership can intervene before problems become churn events.
- •Automatic Slack posts create shared context and a “pulse” on customers
- •CEO and teams can scan calls without chasing reps for updates
- •Separate churn-risk alerts when sentiment falls below a threshold (e.g., <7)
- •Acknowledges false positives; humans can annotate context in-channel
From customer language to demand gen: extracting keywords for Google Ads
Matt shows how each call also fuels marketing: an LLM extracts high-intent terms customers actually use. Those terms are automatically added to Google campaigns, tightening the loop between voice-of-customer and paid acquisition.
- •Customers reveal intent and problem language directly in calls
- •LLM outputs keyword candidates aligned to Suzy’s positioning and offers
- •Automation adds keywords into Google Ads campaigns programmatically
- •Closed-loop system to find more prospects similar to successful customers
AI coaching for Sales/CS: individualized feedback and trend tracking
The automation generates immediate coaching notes for the rep: strengths, weaknesses, and improvement suggestions. This feedback is also stored so managers can spot behavior patterns and make performance reviews more objective and continuous.
- •Post-call coaching delivered quickly while context is fresh
- •Helps motivated reps improve faster; raises baseline performance
- •Store feedback to detect recurring issues (interrupting, ending early, missing topics)
- •Creates continuity even if managers change; reduces reliance on coaching quality
Human-in-the-loop follow-up email writer: faster, better post-call execution
To reduce the burden of post-call admin, the system drafts a high-quality follow-up email the rep can copy, edit, and send. Matt keeps a human approval step to prevent misfires and allow contextual judgment (timing, recipients, tone).
- •Automates a high-friction task: writing polished follow-up emails
- •Delivers draft immediately after the call along with other outputs
- •Human-in-the-loop avoids incorrect context, premature sends, or wrong CCs
- •Boosts speed of execution without sacrificing accountability
Building an aggregate customer profile database for RAG and prep
Beyond per-call actions, the workflow structures each call into a database: roles, product interests, trends, and use cases. Sales can then query aggregated patterns (e.g., what automotive brand managers care about) to prep smarter and standardize playbooks.
- •Create structured fields per call (role, interests, trends, product areas)
- •Structured data improves retrieval/RAG compared to messy notes
- •Enables fast pre-call intelligence by segment (industry, title, persona)
- •Shifts value from single call outputs to compounding organizational memory
Redacted content engine: auto-generating SEO blog posts and ads from calls
The most ambitious workflow converts calls into anonymized, SEO-optimized blog posts that remove all identifying details. Posts publish after a delay (e.g., 21 days), scale into thousands of pages, and can feed dynamic search ads—turning conversations into reusable market-facing assets.
- •Extract reusable use cases without exposing client identity or strategy
- •Aggressive redaction + extensive testing to prevent confidentiality breaches
- •Generate headline, graphics, CTA, and SEO structure; publish with a safety delay
- •Scale content library and run Google Dynamic Search Ads against it
Org design implications and prompting approach: super ICs + guardrails
Matt closes on how automation changes hiring and ownership: fewer “order takers,” more proactive builders and operators, plus a small group of GTM automation orchestrators. For prompting, he uses a guardrail-first method—clearly stating goals and what the model must not do, then iterating toward the desired output.
- •Hiring shifts toward ambitious, proactive “hands-on keyboard” contributors
- •Need GTM automation orchestrators to manage the blueprint and reliability
- •Functional teams own outputs (marketing owns blog performance; sales owns usage)
- •Prompting framework: define objective, add “do not do” constraints, then refine
