Sharing a pattern I’ve been using in production that’s saved a significant amount of manual work.
The problem: Customer emails arrive at all hours. Most are variations of the same 5-6 questions — pricing, how to install, compatibility, refund requests. Writing individual replies was eating 45+ minutes a day.
The approach:
The workflow does three things in sequence:
-
Classify the intent — a Claude node reads the email and categorizes it: technical question, billing, feature request, or generic. Each category gets a different prompt and tone.
-
Pull context before drafting — for technical questions, it fetches the relevant docs/FAQ from a Supabase table before generating the reply. This keeps the AI grounded in actual product behavior, not hallucinations.
-
Draft + human gate — the reply goes to a Telegram message with Approve / Edit / Skip buttons. I never send anything blind. But in practice I approve ~80% without editing.
The key trick is step 2. Generic “reply to this email” prompts hallucinate product details. Feeding the relevant doc chunk first cuts that to near zero.
What I’d do differently: I’d add a confidence score node up front. Low-confidence emails (ambiguous intent) should go straight to human review, not waste tokens on a bad draft.
Happy to share the workflow structure if useful. I packaged it as a ready-to-import template here: soloflow.tools/products/ai-support-inbox-workflow/ — but the pattern itself works for any inbox automation use case.
Anyone else using a similar approach? Curious how others are handling the “don’t hallucinate product details” problem.