Everyone has had the moment. You ask an AI tool something, it answers instantly and cleanly, and you believe it for about a day before finding out it was wrong.

Mollie Bodensteiner, Drew Schoettler, and Jessica Buchwald run GTM operations, AI, and strategy at ZoomInfo, and they spent an hour on what sits underneath that moment.

Jessica's line is the one that stuck with me: “The system never tells you when it is lost. Absence at least announces itself.”

Jessica named the failure mode nobody designs around. A model that is missing context does not stop and ask.

"It doesn't raise its hand when it's confused. It will just continue to make mistakes over and over."

A human analyst working from a bad spreadsheet eventually says something looks off. The system does not. It produces a clean, plausible, wrong answer at the same speed it produces a right one, and nothing in the output distinguishes the two.

That puts the burden on monitoring what comes out and feeding corrections back in, rather than on trusting the setup once and moving on.

Drew separates missing data from wrong data, which most teams treat as one category of problem.

"A stale record is not neutral. It can actively cause a system to prioritize wrong accounts. It can contact somebody who has changed the role. It can reference an outdated technology."

Absence produces inaction, where staleness produces confident action in the wrong direction. His position is that freshness, accuracy, and consistent definitions matter more than volume, though he grants volume still helps with pattern analysis.

The governance principle Jessica landed on is narrower than most policies, and it starts before the data question.

"The AI spit this thing out. It's correct, but should it be used for this use case? In some cases, no."

Correct and appropriate are different tests, and the second one has no automated version. From there she gets specific: agents and LLMs should have access only to the data points you have decided they should have, particularly once a workflow runs without a human in it.

She is also candid that telling people to stop using random external AI tools mostly failed. Building supported internal alternatives worked better than prohibition did.

Mollie pushed back on the premise, from the receiving end:

"I'm okay with being AI email prospected. I actually expect it, and I'm kind of disappointed, honestly, if you don't. Because I'm like, well, you guys aren't really efficient over here."

The objection was never automation itself, it was what the automation produces. Send her four paragraphs and the answer is, in her words, cool, nope. Her read on the wider argument is the same:

"I don't think email's dead. But crappy emails are dead."

Jessica's prospecting advice follows from that, and it narrows rather than scales. Work the signals down to the top 10%, put real messaging in front of those, and nurture the rest instead of pretending you can personalize at full volume.

The line the panel kept returning to was Drew's, that done is never done, and it sounds like a platitude until you pair it with a system that never raises its hand.

Something that will not tell you when it is wrong cannot be validated once and trusted afterward. The data goes stale, the model changes, the workflow drifts, and none of it announces itself. What replaces the one-time sign-off is a habit of checking, which is slower and nobody enjoys it.

Want to learn from more Revenue Creators? Join the RevGenius Community, and for live discussions we have Coffee Talks, next Thursday at 12pm ET.

What Should Actually Be Agentic in Your Outbound Motion (And What Shouldn't)
Wednesday 19 August, 11:30 AM EST

Every outbound tool now calls itself agentic, but a lot of what they've automated was never the problem. A working session on where autonomy earns its keep and where a human stays in the loop.

Are you hiring and want your job featured? Submit a role here.

Want to get your brand in front of 60,000+ revenue leaders?

Partner with us and put your name at the center of the conversation.