How Suzy's CEO Turns 25,000 Hours of Sales Calls into Automated Marketing and Coaching with One Zapier Workflow

Matt Britton, founder and CEO of the consumer insights platform Suzy, built one Zapier workflow that turns recorded customer calls into summaries, churn-risk alerts, employee feedback, follow-up drafts, ad keywords, and marketing content.

Matt has worked at the edge of digital marketing for decades. He says he sold early Facebook ads directly to Mark Zuckerberg and Eduardo Saverin in 2005 and bought some of Google's early keywords for his first agency. He is not a software engineer, but he likes finding business problems he can solve by assembling existing tools. His book, Generation AI, covers the generation growing up alongside this technology.

The raw material was already inside Suzy: 25,000 hours of recorded Gong calls collected over five years. Matt saw that those transcripts captured what customers cared about, how they described their problems, and where accounts were getting stuck. He used Zapier to route each new transcript into several downstream jobs.

In this episode of How I AI, Matt shows how a customer call becomes a structured summary, a low-sentiment alert, feedback for the employee, a drafted follow-up email, suggested Google Ads keywords, and a draft blog post. The workflow treats conversations as reusable operational data rather than meeting records that disappear after the call.

The go-to-market workflow: From Gong call to shared business data

How to Build an Automated AI Sales Coaching System from Call Transcripts: Give sales and customer-success reps timely private coaching by comparing each call with a transparent behavior rubric, citing specific moments, and keeping development feedback separate from automated performance decisions.

How to Automate Customer Call Analysis and Predict Churn with Zapier: Turn each authorized customer call into a traceable account-health signal by cleaning the transcript, resolving the account and owner, extracting evidence-backed risks and next steps, and routing the result to the right team.

Matt's strategy starts with customer conversations. The workflow captures each recorded call, adds account and employee context, runs several analyses, and sends the resulting outputs to sales, customer success, and marketing systems.

The trigger, hacking Gong data with Browse AI

The first challenge was getting data out of Gong and into Zapier. The trigger Matt needed did not exist at the time, so he assembled a workaround with Browse AI. Suzy's recording announces that the call is being recorded for quality assurance. Any team adapting the workflow should also set its own recording notice, consent, retention, and access rules before processing customer conversations.

"I could have given up, Claire, like at that point. That one step probably took me the longest... But you just have to know that there's a way. Just 'cause the tool doesn't do it, doesn't mean it can't be done."

Here's the workaround he built:

1. Identify the URL Pattern: Matt noticed that every Gong call transcript had a unique URL, and the only part that changed was a specific call_id at the end.
2. Trigger on New Calls: He set up a Zap to trigger whenever a new call was logged in Gong. This initial trigger provided the essential call_id.
3. Scrape with Browse AI: He then used a tool called Browse AI to scrape the page. The Zap passes the call_id to Browse AI, which builds the full URL, goes to the Gong page, and pulls the entire raw transcript.

That two-part trigger became the foundation for the rest of the Zap. It also shows how Matt approached a missing integration: inspect the available identifiers, then bridge the gap with another tool.

Data prep and enrichment

Once Zapier receives the raw transcript, Matt cleans the data and layers in additional business context before any AI analysis starts.

Core analysis, generating summaries and sentiment scores

This is the first major LLM step. Using a ChatGPT (GPT-4 Turbo) action inside Zapier, Matt feeds in the cleaned transcript along with a detailed prompt designed to pull out operationally useful insights.

The model handles several jobs in a single pass:

Analyze the customer success call transcript between Suzy and the client. Summarize the company, participants, purpose, main topics, outcome, and next steps, excluding small talk. Describe customer sentiment and any concerns. Give sentiment a score from 1 to 10, where 10 indicates high satisfaction and 1 indicates possible discontinuation. Identify one thing the employee did well and areas they could improve.

The prompt produces a structured summary, sentiment score, next steps, and observations about the employee's performance. Matt says Suzy benchmarks average sentiment scores against actual churn and has found them highly predictive of churn and expansion. That is Suzy's reported result, not a guarantee that an LLM score will predict outcomes in another business.

Internal distribution and alerts

Once the analysis is complete, the results are pushed to the team in real time through Slack.

AI-powered coaching and performance tracking

A separate LLM step drafts feedback on the employee's performance and sends it after the call.

"It actually creates a feedback note to the person on the call... 'Here's what you did right, here's what you did wrong,' and actually sends it to them right afterwards so they understand how to get better."

The prompt looks for behaviors such as strong questions, interruptions, or short calls. The feedback goes to the individual employee and is logged so managers can examine patterns over time. AI-generated coaching can miss context, so it should inform a manager's judgment rather than become an automatic performance decision.

Fueling the marketing engine

The customer insights don't stay inside customer success. The same call data feeds directly into marketing workflows too.

Closing the loop with automated follow-ups

The final step drafts a personalized follow-up email from the call context and sends it to the employee. Matt deliberately kept a person in this loop: the employee can edit the message, change recipients, or decide not to send it.

What changes when one call feeds several systems

Matt's automation makes one customer call useful across coaching, account-risk monitoring, marketing, ad targeting, and internal knowledge sharing. The leverage comes from structuring the conversation once, then routing distinct outputs to the people and systems that need them.

How to Automatically Generate Marketing Content from Customer Calls: Turn recurring customer problems into marketing drafts by extracting source language, minimizing confidential data, generating a generalized use-case article, and requiring privacy, editorial, SEO, and campaign review before release.

Add Keywords to Google Ads Campaigns From these authorized call excerpts, propose customer-language search terms. For each, include exact source phrase, number of distinct accounts, search intent, matching approved product capability, landing page, ambiguity or negative-keyword risk, and confidence. Return a campaign proposal only; do not activate or change budget.

Matt also treats the automation as something that needs an owner. Each downstream output can fail differently: coaching can become generic, risk signals can overreact, and marketing language can lose the customer's meaning. The system creates leverage only when someone reviews quality, corrects the prompts, and tracks whether the routed output helps the team receiving it.

Matt describes the people who build systems like this as "super individual contributors": proactive operators who understand a business problem and can assemble a solution without waiting for a full engineering roadmap. He also expects teams to need owners who monitor the quality of each automation's output.

The place to begin is the highest-signal dataset your business already produces. For Suzy, it was recorded customer calls. Before automating downstream actions, define who may access that data, which outputs require human review, and how the team will measure whether each output is actually useful.