Uploaded July 2026 | Updated September 2026, 3 days ago
The default assumption in consumer AI has been that more autonomy is always better: let the agent act, let it send the email, let it book the meeting. Jean-Denis Greze, founder of Town and former CTO of Plaid, says that's backwards. The products that win will be the ones that earn autonomy slowly.
In this conversation, Jean-Denis breaks down how Town decided AI should never be allowed to send an email, only draft one: why the company pivoted from AI tax prep to a personal assistant after watching preparers drown in email and scheduling, how Town graduates an automation from "draft only" to "just handle it" only after a user proves they trust it, and why he thinks the inbox itself disappears as agents get better at filtering it. He also lays out a vision for agent-to-agent collaboration across companies, where LLMs share relationship and institutional signals without ever exposing anyone's raw inbox.
Topics covered:
• Why Town restricts its AI to drafts, never sends
• The pivot from AI tax prep to a personal assistant for everyone
• How Town's automations earn more trust and autonomy over time
• Why "not making the user look stupid" is a non-negotiable design rule
• What happens to the inbox once an agent is filtering it well
• Agent-to-agent collaboration and data silos across a company
To learn more about Town, visit www.town.com. To learn more about WorkOS, visit https://workos.com.
The default assumption in consumer AI has been that more autonomy is always better: let the agent act, let it send the email, let it book the meeting. Jean-Denis Greze, founder of Town and former CTO of Plaid, says that's backwards. The products that win will be the ones that earn autonomy slowly.
In this conversation, Jean-Denis breaks down how Town decided AI should never be allowed to send an email, only draft one: why the company pivoted from AI tax prep to a personal assistant after watching preparers drown in email and scheduling, how Town graduates an automation from "draft only" to "just handle it" only after a user proves they trust it, and why he thinks the inbox itself disappears as agents get better at filtering it. He also lays out a vision for agent-to-agent collaboration across companies, where LLMs share relationship and institutional signals without ever exposing anyone's raw inbox.
Topics covered:
• Why Town restricts its AI to drafts, never sends
• The pivot from AI tax prep to a personal assistant for everyone
• How Town's automations earn more trust and autonomy over time
• Why "not making the user look stupid" is a non-negotiable design rule
• What happens to the inbox once an agent is filtering it well
• Agent-to-agent collaboration and data silos across a company
To learn more about Town, visit www.town.com. To learn more about WorkOS, visit https://workos.com.
