AI email triage
An inbox workflow that reads each incoming message, assigns a category and priority, extracts the key details, drafts a reply and waits for a person to approve it.
- Business
- Fictional home-services company
- Status
- Working prototype
01 · Business problem
The problem
A shared inbox receives booking requests, billing questions, complaints and supplier messages all mixed together. One coordinator reads everything, forwards most of it and types similar replies all day. Urgent issues wait behind routine ones.
Proposed approach
Every incoming message is classified and prioritised first, with the reasoning shown. Key details like names, addresses, invoice numbers and dates are pulled out. A draft reply is prepared in the company's tone. Nothing goes out until a person approves it, and uncertain messages are flagged instead of guessed.
02 · System architecture
How it fits together
- 01
Inbox connector
Reads new messages from Gmail or Microsoft 365 through a dedicated service account.
- 02
Classifier
Assigns category, priority and confidence. Low-confidence messages go to a person.
- 03
Extractor & drafter
Pulls structured details and drafts a reply from approved templates and policies.
- 04
Approval queue
A person approves, edits or rejects. Approved items are routed and logged.
03 · Proposed workflow
Step by step
- 1
Message received
New email or channel message arrives.
- 2
Classification
Category, priority and confidence assigned.
- 3
Draft
Details extracted and a reply drafted.
- 4
Human approval
A person approves, edits or rejects.
- 5
Routing
Sent to the right team, ticket created and logged.
04 · Interactive demo
Try the triage workflow
Pick a fictional message and run it through classification, drafting and approval. You decide whether the reply goes out.
1 · Choose a message
Choose a message from the inbox, or write your own, to run the triage workflow.
05 · Scope
Deliverables
- Category and priority rules
- Detail extraction for common request types
- Response drafting in the company's tone
- Approval queue with approve and reject actions
- Routing rules and activity log
This prototype is built with
- Next.js
- TypeScript
- Deterministic local simulation (no AI model)
- Fictional sample messages
Likely production stack (proposed)
- Gmail or Microsoft Graph API
- A large language model called from a secure server
- Helpdesk or CRM API
- Audit log storage
Results
No results reported. This is a working prototype, not a live deployment, so there are no real measurements to share.
Want something like this for your business?
We’ll talk through how it would work with your tools, your data and your team, and be honest about whether it’s worth building.