Why we focused on knowledge-grounded email drafting instead of raw AI replies
One pattern we kept seeing while building MailsRAG is that support and sales teams do not usually fail because the model is weak. They fail because the workflow lets the model invent answers from the incoming email alone.
That creates replies that sound fine but drift away from pricing, policy, product facts, or security boundaries.
So we focused on a retrieval-first workflow:
- search the most relevant company knowledge first
- draft from that evidence
- keep review-first approval in place
- expand automation only where the results stay stable
We wrote up the reasoning here:
https://www.mailsrag.com/en/knowledge-based-email-replies/
Product page:
https://www.indiehackers.com/product/mailsrag
Also on dev.to:
And on Medium:
If you are working on AI email automation, I would love to hear whether your team trusts generic drafts or grounded drafts more in real inboxes.
