Inside Gong's Custom Agents: Building No-Code Revenue AI That Starts With a Real Question
By Graham Miller, Head of GTM AI Strategy, Enablement, and Engineering at Gong ·
Roughly 88% of organizations are using AI today, but around 60% are not on track to realize outcomes from it. So the problem was never adoption. It is execution. In this GTM Clawllective sales session, Graham Miller from Gong ran a live, no-slides demo of Gong Custom Agents, the no-code agent builder that shipped as part of Gong's Big Dipper launch, and then walked through the AI fluency program his team built in a single week to close that execution gap. This recap covers how a Gong agent is built, when to build inside Gong versus outside it, and how Gong is teaching an entire go-to-market org to actually use AI.
What are Gong Custom Agents?
Gong Custom Agents are governed, no-code AI agents that any RevOps leader or sales manager can build and deploy inside Gong without engineering support. They arrived as part of Big Dipper, Gong's most recent product mission (following an earlier release called Andromeda, part of a space-themed campaign cadence).
The core idea Graham kept returning to: an agent should start with a real question from a real person, not with the technology. When you can build anything, most people tend to build everything and lose track of what they were solving for. Every agent he demoed traced back to a specific revenue question a leader actually asked.
How do you build a Gong Custom Agent?
Graham story-told the build through a live agent he had set up a few weeks earlier. An enterprise revenue leader had a suspicion: that the majority of closed-lost, SDR-generated pipeline came from people who were never seriously evaluating Gong, just curious. She wanted to know whether that was actually true. So he built an agent to find out.
The builder follows a clear top-to-bottom workflow that will feel familiar to anyone who has used an agent studio:
| Step | What it does | In this example |
|---|---|---|
| Trigger | Starts the agent running | A CRM change: an opportunity moves to Closed Lost |
| Filters | Scope the data to what matters | Narrowed to the enterprise leader's segment |
| Instructions | The prompt that tells the agent its job | Classify each SDR-generated opp as "formally evaluating" or "curious / wanting to be educated" |
| Response | Where the output is delivered | A dedicated Slack channel (email is also available) |
| Variables | What data to pull back | The classification plus an observation explaining why |
What can trigger a Gong agent?
At the time of the session, the available triggers were a CRM change, a date condition, and a smart tracker, with more on the way. The smart tracker trigger is powerful: you define a phrase or signal, Gong's tracker flags any call where it appears, and that flag fires the agent. Graham had used exactly this to build a second agent that surfaces deal-stalling language before forecast calls, catching risk that might never come up on the forecast call itself. One live example: AI budget overspend is starting to appear as a stall signal inside deal cycles, and a tracker-triggered agent can flag it early.
How much prompt engineering do you actually need?
Graham was candid that the instructions he showed were over-engineered on purpose. He likes to over-build when he first works with a new feature so he can trust what he sees. The four core tenets of a strong prompt still apply (clear task, clear role, context, and a desired format if you want one), but Gong is pre-instructed to understand much of this already, so you do not need to hand-hold it the way you would a raw frontier model.
Where do the results go?
Today, agent output can be sent to Slack or email to keep you updated over time. Write actions, such as directly updating fields in the CRM, were described as coming shortly. There is also an audit view showing every run: whether it completed correctly, a summary of what happened, and who built it.
Does a transcript-only approach actually work for sentiment?
A recurring question in AI-for-revenue builds is whether you need the audio file, not just the transcript, to judge sentiment. Graham's honest answer was that Gong has nothing audio-based for tone or inflection today, and nothing he is aware of on the roadmap.
But he made a sharper point about why the transcript is usually enough. Gong's AI is trained on a proprietary revenue-interaction data set, so it reads a transcript differently than a frontier model would. His example: friction or conflict during a negotiation looks negative to a raw LLM, which might flag "you're screwing me over" as a bad sign. Gong's data shows the opposite, that this kind of pushback usually means the buyer is engaged and fighting for the outcome they want. A general-purpose model pulls that back as a false positive. Gong does not, because it is not analyzing a moment in time. It indexes the full life of the customer and compares it against aggregated patterns from other deal cycles where similar language at a similar stage correlated with a win.
A practical lesson that surfaced in the discussion reinforces this: teams that build their own sentiment scoring find it takes heavy refinement to be useful. That means real feedback loops, narrowing the analysis window (for instance from 180 days down to 60), excluding irrelevant chatter and internal banter, and feeding false positives back in. It can take a couple of months of tuning before the output is trustworthy enough to fold into something like a customer health score.
When should you build inside Gong versus outside it?
This is the line Graham owns at Gong, and it is constantly evolving, but he framed it simply.
Revenue-generating activities stay in Gong. Deal reviews, forecasting, pipeline generation, and closing are not new workflows. They are tried-and-true processes that drove revenue long before AI, and Gong's proprietary AI and data set make it the expert for accelerating them with better insights.
Knowledge work often lives outside Gong. How a go-to-market operations or enablement team does its job includes things that belong in the platform and things that do not. A traditional learning-management journey, for example, is not something you run inside Gong. That gap is exactly what led to his second build.
What is the AI fluency program Gong built?
The starting problem: most people mistake AI literacy for AI fluency. Literacy is knowing what the tool does. Fluency is navigating with it seamlessly, the way you would react in real time if a Paris patisserie offered you a pain au chocolat instead of the croissant you expected. To sell credibly in today's market, a field team has to show up as genuinely fluent experts.
Graham put fluency on a five-level maturity model, transparent that just "full sending" AI (pumping in customer data, reaching for the highest models with no security mindset) is itself a risk:
| Level | Name | What it looks like |
|---|---|---|
| 1 | The Risk | Using unvetted tools with no security mindset |
| 2 | The Novice | Can write a prompt, but is just a user |
| 3 | The Practitioner | Co-creates with AI, fact-checks outputs, iterates |
| 4 | The Strategist | Orchestrates multi-agent workflows |
| 5 | The Architect | Builds new systems and leads the transformation |
To operationalize it, he and one other person built an internal platform (nicknamed "brainstorm") using Claude Code and Codex. It ties together the parts of learning he believes in: gamification, a leaderboard of the first people to finish each new course so you can flex on your peers, a daily challenge modeled on an "AI Wordle" to keep people coming back, and a curated journey that moves learners from AI 101 to 201 to 301, each with hidden activities and a knowledge test. Underneath, it handles the reporting the team actually wanted.
The most striking part: the ask came on April 29th, and it was live by May 5th. About a week, two people.
What were the results?
The reporting is where it paid off. Gong's go-to-market team progressed through the program at nearly 4X the speed of the rest of the company, because they are the ones frontline to customers who want consultative help on AI strategy. The bet is that improving the company's own AI fluency directly improves its ability to generate and close revenue.
Access to AI tooling is gated behind that fluency:
| Milestone | What it unlocks |
|---|---|
| Complete AI 101 | Can request a frontier LLM: Claude, Gemini, or ChatGPT |
| Pass the final AI 501 | Access to coding tools: Claude Code, Codex, and Antigravity |
| No training completed | Gong AI only; no other AI tooling out of the box |
The logic is straightforward: handing someone a powerful tool before they know what to do with it just produces expensive experimentation.
How does Gong think about build versus buy?
Build versus buy came up as a major theme, and it is another reason AI fluency matters, because you cannot make the call well without understanding the tradeoffs. Gong's position is not to force customers one way or the other, but to be consultative about the questions that actually decide it: maintenance, critical-workflow support, time to value, and inference cost.
Graham was direct that cost tends to become the single greatest deterrent to a heavy build. If you pull everything into a Claude-centric sales hub, you often end up with three or four layers of inference running at once, and context fractures like a game of telephone, drifting in meaning by the time it reaches the end. Unless you are genuinely well-versed in AI architecture (a hybrid of open-weight and closed models, deliberately delegating tasks, confident in your hosting), the economics get hard fast. A related caution applies underneath all of it: none of this works if the underlying data is not normalized, because a model that cannot find what it needs will confidently reason from the wrong data and hand you an answer that looks right and is completely wrong.
On the natural follow-up, whether Gong might build its own CRM, Graham was careful to say he is not in the highest-level product conversations. But he pointed to Gong's data extractor as a hint of the direction: a lightweight, effective system of action and record that is orchestrated on your behalf. Wherever that goes, he said, the anchor stays the same: AI with clear revenue impact.
Key Takeaways for GTM Teams
- Start every agent with a real question from a real person. When you can build anything, the discipline is deciding what you are actually solving for. Every agent worth building traces back to a leader's specific revenue question.
- Match the tool to the job. Keep revenue-generating workflows (forecasting, deal reviews, pipeline, closing) where the proprietary data lives. Push knowledge work and general enablement to the tools built for it.
- Context beats raw transcript reading. A model that only sees a moment in time will flag false positives. Judging signals against the full customer lifecycle and aggregated outcomes is what makes them trustworthy.
- Gate access behind fluency. Requiring a baseline course before provisioning any frontier LLM, and reserving coding tools for advanced users, turns AI spend into capability instead of experimentation.
- Build fast, but respect what breaks. A two-person team shipped a full learning platform in a week. The same speed makes maintenance, security, and inference cost the questions to answer before you commit to building over buying.
Want to Attend the Next Session?
The GTM Clawllective is a community of go-to-market operators sharing how they actually build and use AI, live and unscripted. Sessions are not recorded, so this recap is the written record. If you want to be in the room for the next one, join the community.
About the Speaker
Graham Miller leads go-to-market AI strategy, enablement, and engineering at Gong, an AI-native revenue platform. He has spent roughly 12 to 13 years in the field's evolution from sales operations to revenue operations to today's AI-driven go-to-market engineering layer. He is based in Provo, Utah, and studied at Brigham Young University's Marriott School of Business.
He approaches the work as a constant learner and self-described expert practitioner, splitting his time across setting AI architecture, teaching reps how to use AI, and building the workflows, agents, and applications that put it to work against clear revenue outcomes.
- LinkedIn: Graham Miller
- Gong Custom Agents launched as part of Mission Big Dipper: Gong announcement