You will be able to build inbox rules and automations that label, forward and file email by type.
Arif handles the shared inbox at a renovation firm in Ubi. By his own count, it gets around sixty emails on a busy day: supplier invoices, quotation requests from homeowners, site photos from supervisors, condo management notices, delivery confirmations, newsletters, and the occasional neighbour complaint about noise. All of it lands in one list. He spends the first half hour every morning reading subject lines and forwarding mail to the right person, and when he is on leave, things get missed.
Email sorting is one of the most common automations people want. It is also one of the easiest to get wrong by using AI too early. The approach that works is to sort in layers. Cheap, predictable rules go first, and AI handles only what the rules cannot sort.
A large share of incoming email can be sorted by the sender or by the subject line. Suppliers send from the same addresses every time. Condo management offices send from their own domains. Delivery confirmations come from a handful of logistics companies with predictable subject lines. Newsletters come from addresses you can list.
For this kind of mail, use the rules (also called filters) built into your email app, as covered in lesson 3.2, "Use the automation you already pay for". A rule matches on the sender's address, the sender's domain, or words in the subject. It then applies a label or moves the message to a folder. Rules cost nothing extra, run instantly and never vary. They are easy to check, and when one misfires you can see exactly why.
To write them, go through a sample period of your email and group it. Every group with a clear sender or subject pattern gets its own rule. Arif went through two weeks of email this way, and the rules he wrote covered most of the volume. Start with the rules that cover the most mail.
The mail left over usually comes from unknown senders, such as new customers, complainants, job applicants or new suppliers pitching. For Arif, it was homeowners asking about renovation, a neighbour complaining, a job applicant and a new supplier pitching. Many of these are general enquiries with vague subjects like "Hi" or "Question", or no subject at all. No rule can sort them, but a model reading the body can. AI classification earns its cost on this remainder.
Set up the AI step as taught in lesson 5.2, "Make the model pick from a fixed list". Give it a fixed list of labels, each with a one-line definition and an example. Check its answer against the list before anything is routed on it. Send unsure, the fallback label for mail that fits no category, to a person. Arif used five labels: quote request, complaint, job application, supplier pitch and unsure. Each has a one-line definition and an example, and unsure goes to him.
Send only unsorted mail to the AI step. There are two ways to do this. You can have your email rules apply an "unsorted" label to anything they have not already sorted, and trigger the automation on that label. Or you can filter at the start of the flow so that mail which already has a label stops there. Running every email through AI would cost more. For mail that a rule already handles well, it would also be slower and less predictable.
Once a message has a label, whether from a rule or from AI, route it to whoever deals with it. The destination can be a person (by forwarding, or by assigning it in a shared inbox), a folder, or a task list (by creating a task with a link to the email). Arif's routes are:
Quote requests create a task for the project manager with the homeowner's name, the email subject and a link to the message. Complaints go to the boss and to Arif, flagged. Job applications are labelled and moved to a hiring folder. Supplier pitches are labelled and archived. Unsure stays in the inbox with a label so Arif sees it.
Two safety habits apply to every route. First, never delete or edit the original email, or move it in a way that makes it hard to find. Label it and file it, but keep it in the account so anyone can go back to what the sender actually wrote. Second, the routing step only ever forwards a link or creates a task. It never replies on the business's behalf. If you want drafted replies, add them as described in lesson 5.3, "Draft replies a person approves before sending", with a person in charge of sending.
New routing rules make mistakes that only show up with real mail. A supplier changes their sending address. A homeowner writes "invoice" in the subject of a question about a quote. A newsletter rule catches a real message from the same company.
In his first week, Arif found two problems. A supplier who also did renovation work for the boss's relatives sent personal emails, and his rule filed them as invoices. The AI also labelled a homeowner's angry follow-up about a delayed quote as a quote request when it was a complaint. He fixed the first by narrowing the supplier rule to emails with attachments. He fixed the second by adding an example of a frustrated follow-up to the complaint definition.
For the first two weeks, check your sorted folders and labels every day, which takes a few minutes. When a message has been sorted or routed to the wrong place (a misroute), fix the rule or label definition that caused it so the same kind of message lands correctly next time. Keep a short note of each misroute and its cause. After two weeks with few or no misroutes, drop to a weekly check.
In the activity below you will write the routing rules for your own inbox, split them into simple rules and the cases that need AI classification, and set up the first three. Start with the rules that cover the most mail.
List the routing rules for your inbox. Split them into simple rules and cases that need AI classification, then set up the first three.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).