ZoomInfo's Three-Tier Rule for Letting AI Agents Write to Your CRM

ZoomInfo’s Florin Tatulea, a GTM Engineer in Residence, offers a three-tier rule for how much freedom an AI agent should have when it writes to a CRM: “Automate where bad data costs time. Keep people in charge where it costs money.”

Tier one is fully automated work such as data enrichment and validation, because errors there are recoverable. Tier two is scoring and routing changes, which go through a review queue because edge cases need human judgment. Tier three is revenue-impacting changes, such as creating opportunities or moving deal stages, where a person makes the call. Tatulea frames the central difficulty this way: the hard question was never whether an agent could make a change, but whether it should.

The tiers are easier to trust when the evidence is visible. Outreach’s Deal Agent writes deal signals from recorded conversations into CRM fields, links every update to the source conversation with timestamps and transcript excerpts, and lets teams choose automatic updates or approval before anything is written. In 90 days of Outreach’s internal usage, users accepted 90.8 percent of AI-recommended field updates, and users clicked through to verify the evidence more than 1,000 times.

Those two sources point to the same design. Approval is a setting, not a verdict on the agent, and it should be set by the cost of a wrong write. A mislabeled industry field costs a few minutes. A deal moved to the wrong stage corrupts the forecast that leadership acts on.

For sales leaders putting agents into the CRM, the first step is a field-by-field sort into those three tiers, with the revenue-affecting fields held back from automation until the acceptance data justifies the move.