A request arrives by email with the intent, data and urgency buried in the text. Someone has to read it, understand it, turn it into a task and remember to reply, and requests get lost along the way.
Who it is for
Small businesses that receive orders, quote requests, complaints or support requests by email in a shared inbox (sales@, support@, info@) and handle them by hand.
The blueprint has five stages: classify, extract, create the task, check for risk and close the loop with the customer. You can test it today by copying emails into an AI chat; once it works, it connects to your email and task manager with n8n or Activepieces.
The protocol, step by step
-
Define your request types
Go through the last 30 emails in the inbox and write 4 to 6 types. Always add "other" for what does not fit.
Copy and pasteExample: quote · order · complaint · tech support · invoice · other
-
Define the fields for each type
For each type, write down which data you need to be able to act. Whatever is missing is what AI will ask the customer for.
Copy and pasteQuote: name, company, product or service, quantity, required date, city. Complaint: name, order number, what happened, what the customer expects.
-
Classify and extract with one prompt
Paste each email with this prompt. Ask for the answer in a fixed format so it can go into a task without rewriting.
Copy and pasteClassify this email into one of these types: [LIST]. Then extract the fields for that type: [FIELDS]. Return: type · urgency (high/medium/low and why) · fields found · missing fields · one-line summary · requires human review? (yes if there is a complaint, a high amount, a cancellation threat or doubt about the type). Email: [PASTE]
-
Create the task with owner, date and evidence
Each type has a default owner and a deadline. The task carries the summary, the fields and the link to the original email as evidence.
Copy and pasteTitle: [TYPE] · [CUSTOMER] · [SUMMARY] Owner: [by type] Due date: [by urgency] Description: extracted fields + link to the email
-
If there is risk or doubt, ask for human review
Everything the prompt flagged for human review goes to a person before replying. So does anything that landed in "other". AI does not answer complaints on its own.
-
Send the result back to the customer and log it
When the task is created, send an acknowledgement with what you understood and what is missing. When it closes, reply with the result and keep the reply linked in the task.
Copy and pasteHi [NAME], we received your [TYPE] request: [SUMMARY]. [OWNER] has it and we will reply by [DATE]. To move forward we still need: [MISSING FIELDS].
-
Automate once it works by hand
Once a week of emails goes through steps 3 to 6 without surprises, connect the inbox to n8n or Activepieces: trigger on new email, AI step with your prompt, task creation and draft reply. Replies to complaints stay under human approval.
Common mistakes
- Letting AI invent data that is not in the email. Always ask for the list of missing fields.
- Creating tasks without an owner. A task with no owner is an unread email under another name.
- Answering complaints automatically. Always review before sending.
- Connecting the whole company's email at once. Start with a single shared inbox.
Who this is NOT for
If your problem is a personal inbox full of newsletters and notifications, this is not for that: see the Inbox Zero card in the Vault. This is for customer requests that must end in work done.
Want it applied to your business?
If you would rather not build it alone, book 30 minutes with me and we will review it with your real case. If you still do not know where to start, the map tells you what to automate first in 5 questions.