Skip to main content
AI2026-09-22 · 5 min read

An AI copilot that drafts, never sends

The send button belongs to a human. See where copilots help in a support inbox and where full-auto replies fail.

Field notes · Built by Arche Axon

The pitch for AI in support inboxes is usually full automation: let the model reply, save the labor cost. The failure mode is equally predictable: a confident wrong answer sent to a customer in your name, with no human in the loop until the complaint.

Drafts, not sends

Vendale's copilot drafts; a human sends. That design keeps the decision with the team member. The copilot can use the thread and available customer history to propose a reply. The human decides whether the draft is accurate and what the business wants to promise.

Where copilots genuinely help

  • A starting point for repetitive questions: review and edit a suggested draft instead of beginning with a blank response.
  • Context recall: what this customer bought, what they were promised, what the last ticket said.
  • Tone rescue at 6pm on a Friday, when the honest answer is 'not yet' and the tired answer is shorter.

The failure mode to refuse

Full-auto replies to payment disputes. Full-auto promises about stock. Full-auto anything where being wrong is expensive. A copilot that cannot say 'draft only' will eventually say something your business has to apologize for.

Every module in these notes ships.

Request access and put the notes under your own finger.

Request access