You can build a dashboard that pulls together your forecast, call transcripts, and CRM. You can build an agent that handles one task surprisingly well. But if you don't trust the information underneath it, a confident answer doesn't get you very far.
“So it's just a good reminder, too, that just 'cause it sounds confident doesn't mean it's right.”
About This Episode
Heather Robinette joined me to talk about what changed after she completed Overclock's 10-week AI for RevOps course. Before it, she mostly used AI through chat. By the end, she was connecting APIs, building applications, and working with agents. The useful part of that story is what she started doing with those skills. She describes pulling call information, Salesforce data, and weekly forecast discussions together to give leadership more context around a forecast. Her example of a risk-adjusted number is illustrative, not a reported business result.
She also walks through a marketing dashboard that started with sourced versus influenced pipeline. Then came campaign performance, conversion rates, goals, and the questions her team actually wanted to answer. She estimates putting about 20 hours into it through repeated iteration. Now the team has somewhere to start when it wants to understand a campaign, instead of asking her to pull every number. That is a practical change in how the work gets done, even if building it took more than a clever prompt.
I share some of the smaller agents I've been building for my own work: triaging email, drafting replies, and finding duplicate tasks in Asana. We also get into the context problem. Someone helping with a project needs more than a task name. They need to understand the conversations and decisions behind it. I explain how I'm connecting meeting summaries, Google Docs, and NotebookLM so the people working with me can find that context themselves.
The conversation comes back to boundaries. Which systems should an agent be able to read? Which should it be able to change? Heather talks about learning to work inside a virtual machine and being deliberate about connectors. We discuss giving agents specific jobs, including checking other agents' work. She describes an instructor's team of agents as one example; she is clear that she has not built that whole setup herself. We also compare models by the actual task and cost, instead of assuming the newest option is automatically the best fit.
Then we get to the data. In one client example, I describe finding that 10,000 of 14,000 company records were missing revenue, even though revenue mattered to the targeting criteria. An AI-generated account list still has to work with that gap. Heather's point is straightforward: get the data to a place where you can trust a basic report before expecting AI to produce dependable insight from it. The work can get faster. The responsibility for checking it stays with us.
About Heather Robinette
Heather Robinette works in revenue operations and is now Senior Revenue Operations Manager at Zafran Security. When we recorded this conversation, she was at AppOmni. Her examples come from the operational work of helping teams understand forecasts, pipeline, campaigns, and the data behind their decisions. After completing Overclock's AI for RevOps course, she began applying what she learned to dashboards, applications, and agents. In this conversation, she shares what she built, what she had to learn, and why security and data quality became a bigger part of how she thinks about AI.
Key Takeaways
- Start with a question your team needs answered. Heather's marketing dashboard grew from a specific pipeline question into a place the team could explore campaigns and progress against goals.
- Give agents a clear job. Narrow responsibilities made it easier for me to see what was breaking and fix one workflow without disturbing another.
- Treat access as a design choice. Heather distinguishes what an agent can connect to from what it should be allowed to read or change.
- Test models against the work. We discuss comparing the same task across models and weighing the result against cost.
- Check the data and the answer. Missing fields can undermine a useful-looking output. Confidence in the wording is not evidence that the underlying answer is right.
Chapter Markers
00:00 - Cold open
00:18 - AI for RevOps course
03:30 - CRO and marketing dashboards
05:59 - Travis's agents and shared context
13:01 - Security guardrails and sandboxing
15:15 - Single-task agents that check each other
18:41 - Choosing models and managing cost
22:58 - What companies want from AI
25:43 - Bad data means bad AI
31:01 - Wrap-up
Resources & Links Mentioned
Tools & platforms
Claude, Claude Cowork, Claude Code, ChatGPT, Gong, Salesforce, Clari, Asana, AskElephant, Google Docs, NotebookLM, DigitalOcean, Perplexity, Manus, HubSpot, and HubSpot Breeze.
Books & podcasts
None.
Other resources
Overclock's AI for RevOps course.
Connect with Heather Robinette
Connect with Travis
If This Episode Resonated
Recorded August 14, 2026. Platform references and examples reflect that conversation.
AI-Ready or Not is a production of Rainier RevOps.
Recorded in Zoom. Edited in Descript. Hosted on Transistor.