Intuit’s AI agents and the case for keeping humans in the loop

In April 2026, VentureBeat reported that Intuit shipped AI agents to about three million customers and saw roughly 85% repeat usage. Leadership framed the outcome as evidence that combining AI with human expertise drove adoption more than model quality alone. Treat those metrics as company-reported.

The design pattern

• Automation handles volume and drafting.
• Humans handle exceptions, judgment, and accountability.
• Finance and tax workflows punish silent errors.
• A system that routes uncertain cases to people can earn more trust (and more reuse) than one that guesses confidently.

Metrics beyond deflection

1. Track re-open rate.
2. Track human correction rate.
3. Track time-to-resolution for escalations.
4. Track customer trust scores where you can measure them.

Expose intermediate agent steps to reviewers. Allow override. Log which policy version produced an action. Train on approved knowledge with retrieval. Match autonomy to blast radius. This is the product counterpart to the agent-security checklist and sandbox-escape analysis.