Fields, not free text
The model fills a defined structure. The router never has to search a paragraph for a department name, and a caller receives the same shape every time.
Loading portfolio
AI and automation / 2026
AI Email Triage Router
A shared inbox gets sorted by whoever reads it first. Subject lines rarely say which team should act, so sorting needs the meaning of the message in a form software can use.
The assignment was to classify incoming emails into business departments. The scenario takes the email by webhook and makes one OpenAI call that must return three structured fields. The category is limited to Sales, Customer Service, HR, Finance, Operations or Other, spelled exactly, so a router can compare it directly. Each department has a branch and Other is the fallback. Each branch returns the department, summary and priority as JSON.
- Role
- Workflow builder
- For
- TS Academy coursework
- Stage
- Course project
Walkthrough
Watch it run
Built from the assignment recording: the scenario, the prompt and its structured fields, the six branches, a test email routed to Sales, and the run log.
My part
What I built
- Built the scenario: webhook, OpenAI module, six-way router and a response per branch.
- Wrote the classification prompt and the three field definitions, with example values for each.
- Set the router filters and marked the Other branch as the fallback.
- Designed the JSON response returned from each branch.
- Tested with sample emails sent by cURL and checked the run history.
Decisions
The choices that shaped it
A closed list of categories
The prompt names the six allowed values and asks for exact spelling. The field definition repeats them as examples. That keeps the model's answer inside what the router can match.
The model classifies, the router decides
The model's output is only an input to the filters. Which branch runs is decided by an exact comparison in the scenario, where it can be read and changed without touching the prompt.
A fallback instead of a forced fit
An email that belongs to none of the five departments is classified as Other and takes its own branch, so it is not pushed into the nearest wrong team.
What has been shown to work
- A sample email asking for pricing for 200 enterprise seats came back as Sales, priority Medium, with a one-sentence summary and the message 'Routed to Sales team'.
- A refund request for a broken order was routed to Customer Service in the same recording.
- The run history shows the model call completing, one branch responding and the other five filtered out.
- An earlier run in the history took the Other fallback.
Current stage and limits
- This is a course assignment. It is not connected to a mailbox.
- Each branch returns a routing result. Nothing is forwarded to a team's inbox.
- Two sample emails were recorded. There is no accuracy measurement over a labelled set.
- There is no confidence threshold, so an ambiguous email still gets one department.
Screens
From the product
Technical detailOpen the system view
SYSTEM ARCHITECTURE
Email Triage Router system view
The assignment was to classify incoming emails into business departments. The scenario takes the email by webhook and makes one OpenAI call that must return three structured fields. The category is limited to Sales, Customer Service, HR, Finance, Operations or Other, spelled exactly, so a router can compare it directly. Each department has a branch and Other is the fallback. Each branch returns the department, summary and priority as JSON.A webhook accepts a JSON body with the email's subject, sender and body.
The OpenAI module receives those three values and a prompt that allows exactly one of six categories.
The model's reply is constrained to three fields: category, summary and priority.
A router compares the category with each department name. Other is the fallback route.
The matching branch returns the department, summary, priority and a routing message as JSON.
- 01Input / actorconnects toInterface
- 02Interfaceconnects toCore service
- 03Core serviceconnects toOperational boundary
- 04Operational boundaryconnects toState / outcome
Security and privacy
- The test emails are invented and use made-up company addresses.
- The webhook address is cropped out of the published captures.
- A real deployment would send message content to OpenAI, which needs to be acceptable for the mail being handled.
Built with
Work with me
I build workflows like this one as a fixed-price service: one defined, tested n8n or Make.com workflow, with packages from $149.
More work
Gazelle's KitchenA WhatsApp business gatekeeper that qualifies genuine orders, enforces business and safety rules, contains off-topic or abusive conversations, and only surfaces the small set of cases that genuinely need the owner.PilotSellavie AIAn independent project: multi-tenant conversational-commerce software that connects social messaging with business knowledge, guarded AI, human handoff, orders, billing, analytics and operations.MaintainedAI Automation Control PlaneA runnable multi-tenant control plane for AI-assisted business workflows, combining FastAPI, Next.js, PostgreSQL, Redis, n8n, provider routing, human approval, idempotency, replay, cancellation, audit traces, and deterministic local demonstrations.Maintained