← Back to Blog

Inside the AI System That Turns Every Gong Call Into Competitive Intel

By Tom Leader & Morgan Cliburn, Chili Piper ·

Most competitive intelligence still lives in an AE's head, or in a one-line Salesforce note nobody fully trusts. Tom Leader and Morgan Cliburn built something different at Chili Piper: an automated pipeline that turns every closed deal's Gong calls, emails, and Salesforce data into structured, queryable competitive intelligence, surfaced in Slack and Claude. Here's how it works, and what they learned building it.



What problem does this solve?

Before this system existed, competitive intelligence at Chili Piper came from two places: whatever an AE typed into a Salesforce closed-lost reason field after a deal ended, and manual keyword searches through Gong's built-in AI summaries. Both had the same limitation. They only worked account by account. Getting a quarter-wide view of why the company was winning or losing against a specific competitor meant a human sitting down and reading through calls one at a time.

How does the pipeline actually work?

Every morning, an n8n workflow checks for opportunities that closed in the last day, marked as new business, and not yet processed. For each one, it pulls every Gong call tied to that opportunity in Salesforce, using Gong's own call-to-opportunity tagging, pulls the full transcripts and speaker data, pulls the email thread from Salesforce, and pulls the opportunity's line items.

All of that gets merged into one long markdown document and sent to an LLM with a structured prompt. The prompt requests the primary win or loss reason, the competitors mentioned, competitive advantages, value drivers, and successful messaging snippets, each one tagged with the Gong timestamp it came from so anyone can jump straight to that moment in the call. The output is forced through a JSON schema rather than left as free text, and the model also returns a confidence score for its own answer.

Why full transcripts instead of Gong's own AI summaries?

Gong already generates AI summaries of every call, so the obvious question is why not just use those. The answer is control over the output shape. Raw transcripts let the system define its own structured fields, like talk tracks and buying committee makeup, rather than being limited to whatever Gong's summary format happens to include.

What happens when the model isn't confident in its answer?

If the confidence score comes back low, usually because a deal didn't have much call or email data, a second agent searches through the rep's other calls for more context and tries again. A retry cap keeps that loop from running forever.

Why store this in a database instead of just posting it to Slack?

The first version of this system posted straight to Slack and a Google Sheet. It was better than nothing, but it was still a human reading a document. The current version pushes structured fields, not paragraphs, into Supabase. Primary win reason, competitors, messaging, and more all become their own queryable fields. That means the team can isolate a single theme, like every deal that mentions a specific competitor, across the entire account base instead of rereading transcripts by hand.

A nightly job also generates vector embeddings for anything new in the database, so a downstream agent can search by meaning and intent rather than exact keywords. That embedding step is what makes the Slack bot possible. Tag the bot with a question, and it searches the full dataset with a recency-weighted filter, then answers with linked Gong snippets instead of a generic summary.

Why not surface these insights live, on the call?

An earlier version of this project was a Zoom app that would have shown suggestions to reps in real time. Adoption was poor, which the team now reads as confirming a wider point: reps do better with deep, focused listening during a call, not pop-ups competing for their attention. Real-time coaching didn't hold up in practice, and it turned out to be more useful to have insights ready before and after the call instead of during it. That live Zoom app has since been dropped in favor of the Slack bot, an MCP interface in Claude, and a dashboard.

Where does this data come from besides Gong?

Closed-deal calls and emails are the backbone, but the system also pulls from a few other sources: an internal Slack channel where anyone at the company can post competitive intel they've noticed, a monitor that watches competitor websites for changes, and marketing's existing library of customer stories. Combining these reduces how much any one call transcript has to carry on its own, which also helps with a known limitation. The system can't see non-measurable factors, like whether a buyer already had a relationship with the rep going in. That gap gets addressed partly by volume of data over time, and partly by separate work already underway to model sales-process factors more directly.

Does it actually work?

There's no single clean before-and-after conversion number. What the team can point to: quarterly board reporting on competitive deals used to take about a week to build by hand. Now it runs in minutes, and Morgan can also just ask the system on demand whenever someone wants to know how a specific competitor is trending. Win rates against tracked competitors have been moving up over time too, though that's multifactorial and not attributed to this tool alone.

Key takeaways for GTM teams

If you're building something similar Do this
Where to start Closed-won and closed-lost deals, not every call. You get a clean success metric to build around.
Output format Structured fields, not free text. Free text can't be reliably queried later.
Data organization Keep everything about one deal, calls, emails, line items, together in one place.
Where to surface it Wherever people already work, like Slack and Claude, not a new destination they have to remember to visit.

Want to Attend the Next Session?

GTM Clawllective is Chili Piper's weekly community session where a GTM builder shows something real they've shipped with AI. The live discussion stays off the record, and every recap includes the speaker's demo video. Join the community to get the next one on your calendar.


About the Speaker

Tom Leader is Director of Product Growth at Chili Piper, where he builds internal automation and tooling, including the n8n workflows behind this competitive intelligence system. He describes himself as someone who can't code, and instead built the entire flow node by node, with Claude generating each piece as he tested it.

Morgan Cliburn is Head of New Business Sales, Mid-Market at Chili Piper. She uses the system's output for win-loss analysis, competitive battlecards, and the company's quarterly board digest on competitive performance.

Watch the session replay: Competitive Intel with Gong