Drinking our own champagne: using Digibee to mine Gong
I built agentic pipelines on Digibee to mine our Gong transcripts. What I found, and what I'd do differently.
PUBLISHED
June 5, 2026
I built agentic pipelines on Digibee to mine our Gong transcripts. What I found, and what I'd do differently.
Before joining Digibee, I worked as a data scientist. I built dozens of data pipelines, and I know firsthand the infrastructure decisions, credential management, and deployment headaches that come with the territory.
Click and Learn More
When I took on the role of Technical Product Marketing Manager at Digibee, I wanted to build something using the platform. So, I built an agentic workflow to mine our Gong call transcripts for market intelligence.
The pattern proved useful. I built another. Then two more.
This is the story of how I created a suite of agents focussed on extracting useful information from Gong transcripts, what they produced, and what the experience taught me about building agentic workflows with Digibee.
Our sales team logs dozens of calls on Gong every week. Sometimes they share standout moments, but no one has time to review all of them.
Those transcripts hold rich, unstructured intelligence. Customer concerns. AI hopes and fears. Unprompted opinions about our competitors and partners. This is exactly the kind of field-level signal that most strategy decks are built without.
Knowing the strengths of LLMs, and knowing that Gong exposes transcripts via API, I saw the opportunity: pull the transcripts, run them through an LLM, and store the results somewhere useful.
As enthusiastic Gong customers, we know that Gong offers its own AI tools. Our GTM team loves languagesGong’s automatically-translated call summaries, particularly as we work across multiple . But I had a hard time fitting the analysis I wanted to do into the tools Gong made available.
I ended up building four pipelines. Together, they produced:
None of this had been surfaced before. It was sitting in transcripts that no one had time to read.
I settled on a simple high-level architecture before dragging my first connector onto the Digibee canvas.
Each pipeline follows the same core pattern:
Trigger → Gather → Prepare → Extract → Judge → Store
A trigger kicks things off—either a daily scheduler running at midnight, or a REST API call. From there, the pipeline proceeds through five phases.
Gather
The workflow starts by pulling transcripts in bulk from a Gong endpoint. These documents are minimal, identifying each speaker only by a unique ID.
Each transcript then proceeds through its own subprocess. The first subprocess step gathers speaker details from another Gong endpoint. My goal was to learn about the opinions and experiences of our customers and prospects, not those of my colleagues. Therefore, I needed to resolve those IDs into names and company affiliations.
Prepare
I had IDs. I had speaker data. I needed to merge them into a coherent transcript. Digibee's JavaScript connector enabled me to handle this cleanly.
Extract
Using Digibee's Agent Component, I assembled a prompt instructing GPT-4o to extract structured summaries.
I did not get this right on the first attempt. I iterated through different versions of instructions directly in the Digibee canvas. Among other changes, I added a section that listed multiple ways competitors’ names might be transcribed incorrectly, along with instructions to “snap” those mis-transcriptions to the correct name.
I also used the component’s JSON Schema option to force the output to adhere to a specific structure. This resulted in a consistent dictionary for each mention that included the company name, a summary of the context, a key passage, and a 1–5 sentiment score.
Judge
I sent the extracts back to GPT-4 via a second Agent Component. It evaluates the quality of each extract, flagging weak or ambiguous mentions. Since the outputs of the LLMs are not deterministic, a judge layer It adds a relevant quality control before the results move on to the next stage.
Store
The pipeline sent results to a Google Sheet. A production version would route to a database.
Then, because I want my user experience to be strong, I built a Google Data Studio dashboard atop the sheet. My colleagues can easily filter to the data they’re interested in and click directly into the originating Gong call.
This agent-based workflow may seem expensive. Calls transcribed by Gong can last for hours. People talk, on average, about 10.000 words per hour. The token count grows even further when formatting, punctuation, and words that span multiple tokens are taken into account. Additionally, the workflow makes two calls to an OpenAI endpoint.
It must cost a fortune, right?
Not at all. My daily token costs fluctuate according to the number and length of calls transcribed (and they’ve definitely grown over time), but the workflow generally costs about 0,50 cents per day.
Due to the way users orchestrate workflows in Digibee, this application is not at risk for a runaway spiral. No LLM decides when to call another agent, or when to ask itself more questions. The workflow prepares the context deterministically, asks the LLM to use its creativity only where it adds value, and moves on.
The only way the cost of this workflow could explode is if we suddenly had more conversations with customers and prospects—which I would consider a good problem to have.
The first pipeline took me about three days, spread around other work. I was learning the platform, figuring out Gong's API quirks, working through how to structure prompts for consistent extraction, and solving for things I hadn't anticipated.
By the second pipeline, I had a template. Copy the core pattern, adjust the prompt, evaluate the outputs, and reconfigure the destination. By the fourth pipeline, I could go from plan to deployment in 2-3 hours.
This reflects Digibee's design philosophy. The platform is built around reuse: components, patterns, and configurations carry forward from one project to the next. Later builds benefit from earlier ones. The activation energy is front-loaded.
I want to be clear. I could have built these pipelines in Python. From a purely code-level, I’m confident I would have built a functional POC of the first workflow faster that way.
I'm less confident that I could have built the second, third, and fourth pipeline faster in Python. And the quality of the pilots on Digibee was much higher than what I would have strung together in my IDE.
Digibee manages credentials so thoroughly that the keys I used to access Gong, OpenAI, and Google Sheets never surfaced in a place I could see them. Had I built these workflows from code, those credentials would have been floating as environment variables or (worse) hard-coded directly into the script.
The Agent Component simplified prompt iteration and version control. The ability to drag each component onto a linear canvas facilitated the correct composition of the deterministic elements before handing them over to the LLM. It also prevented me from calling out functions out of order, something I've been guilty of before on large orchestration roles.
The platform also handles infrastructure invisibly. In previous roles, I spent days wiring together Lambdas, EC2 instances, and S3 triggers to build pipelines I'd trust in production. Deploying the workflow on Digibee’s cloud infrastructure took less than a minute and just a few clicks.
What I'd do differently
I stumbled a few times on this project. Embarrassingly, I have to chalk most of them up to “user error.” If you would like to build similar projects in Digibee, learn from my mistakes.
Take Digibee's courses.
In my review of early prospect calls, one of the questions that came up was whether Digibee provided training. The answer is emphatically “yes.” And my advice is to take the online courses!
In my eagerness to build something, I charged into the platform without any education or preparation—even clicking out of the in-canvas training wizard. Don’t do that.
Digibee simplifies a great deal, but all powerful tools have learning curves. Investing 3-4 hours on tutorials would likely have saved me a day and a half on the first build.
Learn bracket notation
When I presented my projects internally, our engineers noted that my workflow included several unnecessary connectors. They actively stored data and actively retrieved it later. Digibee’s bracket notation allows users to access data from any previous connector. This would have simplified my canvas.
Be careful with LLMs. They can be silly.
In one experiment, several transcripts caused the extraction connector to exceed its maximum token output. This caused the LLM endpoint to return a brief error message instead of its expected JSON.
The judge node, not receiving its expected input (but understanding the intended shape of its output) hallucinated, inventing snippets of conversation. Digibee offers two solutions to this problem, extending the LLM context window and setting the workflow to fail if the agentic step fails—both of which I enabled after deleting fantastical comments from Mark Zuckerberg.
I built all of this as a PMM trying to understand a product I'd just signed up to market. I wasn't trying to build something impressive. I was trying to learn.
What I didn't expect was to build four things that people actually use:
One caution: Once I delivered these early projects, requests started rolling in. Now I have my own integration backlog to manage!
If you're thinking about your first integration or agentic workflow on Digibee, I'm happy to talk through what I learned.
Matt Casey, Senior Product Marketing Manager at Digibee, creates blog posts, videos, e-books, and other content about enterprise technology. Before moving into product marketing, Matt built data products as a data scientist and produced stories for magazines, newspapers, and radio stations. In his spare time, he plays board games, parents two sons, and builds the occasional AI side project.
RECOMMENDED POSTS
An AI engineer explains some of what he had to do to build the Digibee Digital Worker.
AI will not replace deterministic workflows in financial services. It complements them with reasoning and context capabilities where rules alone are not enough.
What differentiates an AI-native platform from an AI-powered one? A platform rebuilt for agents, with specialized workflows and collaboration between people and AI.