What it does in a normal day
A new email arrives in Gmail or Outlook. The assistant reads it and does three things:
- Labels it. Quote request, booking change, billing question, complaint, vendor, newsletter, spam. You choose the categories.
- Drafts a reply for the routine ones, using your FAQ, price sheet and policies, saved as a draft in the thread, not sent.
- Escalates anything sensitive or unclear, with a one-line summary posted to Slack or a daily digest email.
You open the inbox, review the drafts, tweak a word or two and hit send. The time saved comes from not starting every reply from a blank screen.
Drafts first, autopilot later (maybe)
I build every email assistant in draft mode first. After a couple of weeks of reviewing its output, you will know which categories it handles well. Some businesses then let it auto-send a narrow set, such as "thanks, we received your documents" confirmations. Others keep it in draft mode forever, which is a perfectly good outcome.
What it should not do, ever:
- Invent prices, availability or delivery dates that are not in its source material
- Make commitments on refunds, contracts or legal matters
- Reply to angry customers without a human reading the message first
- Forward or summarize sensitive data to places it was not meant to go
Feeding it the right knowledge
The quality of drafts depends almost entirely on what you give it. Before the build, gather:
- Your 20 to 40 most common customer questions and how you answer them
- Current prices or price ranges, service areas and hours
- Policies: cancellations, refunds, warranties, payment terms
- Five or six real replies you were happy with, so it learns your tone
If you have never written your FAQs down, the chatbot script generator is a quick way to turn business details into a structured FAQ knowledge base. It was built for chat widgets, but the same document works well as the assistant's reference.
How it is wired together
The assistant runs as a workflow in n8n, Make or Zapier, connected to your mailbox and to the OpenAI or Anthropic (Claude) API. A trigger fires on each new email, a model call classifies and drafts, and the result goes back into the thread as a draft plus a label. It is the same pattern as the AI qualification step in the lead follow-up demo build, pointed at your inbox instead of a web form.
A note on privacy: emails often contain personal or financial details. We limit what is sent to the model to what it needs, use API accounts you own and avoid storing message content outside your existing systems. If you work in a regulated field, check your obligations before routing client email through any AI service.
Cost and timeline
ChatGPT or Claude integrations start at $150 for a focused build, such as triage plus drafts for one inbox. Most take 3 to 7 days including a testing period on your real email. You own the workflow and API accounts, and AI usage is billed by the provider directly. Start a project and tell me which inbox is eating your day.