Answer support mail with your refund policy
For each new support mail, this workflow labels the mail, finds the order, applies your refund and return policy, and saves a draft reply in the customer’s thread. It never sends a mail. You read each draft and send it.
What it does
Section titled “What it does”- It reads the new mail and puts one of your Gmail labels on it: refund, return, damaged or wrong item, delivery, cancellation, billing, complaint or other.
- It finds the order in your order sheet.
- One AI step reads your policy and the order, decides what is due, and writes the reply with the right amount. The step makes one model call for each mail.
- The workflow saves the reply as a draft in the customer’s thread. A mail that needs no reply gets no draft.
The workflow has 6 steps, and one run handles one mail. When 30 mails come in, 30 runs start together.
What you fill in
Section titled “What you fill in”- the mail address that support mail comes to;
- your order sheet;
- your refund and return policy, as a Google Doc;
- the Gmail labels that you sort mail into.
What it needs
Section titled “What it needs”Gmail, Google Sheets and Google Docs, from one Google account that you connect.
Where the result goes
Section titled “Where the result goes”A label on each mail, and a draft reply in the customer’s thread in your Gmail. Nothing goes out until you send the draft.
Why it runs by itself
Section titled “Why it runs by itself”Support mail comes in all day. The workflow starts on each new mail to your support address, so the draft is ready when you open the mail. What starts a run shows how to start a workflow on new mail.
Each incoming mail goes through a check for text that tries to give orders to an AI model. In the test, one mail was a phishing mail. The check stopped its run, and the run went to Action items for a person to look at.
How it was built
Section titled “How it was built”One chat in claude.ai with the Toolscaled connector: 8 messages from the person, 27 minutes. The person typed these messages. We replaced the mail address and the file links with brackets. The [bench support] tags mark the test data.
-
Which Toolscaled workspace are you connected to?
-
I run a small online shop, Larkspur Home & Outdoor. Customer support mail comes to [support address]. For each support mail I want it sorted into one of my Gmail labels bench/support/refund, bench/support/return, bench/support/damaged_or_wrong, bench/support/delivery, bench/support/cancellation, bench/support/billing, bench/support/complaint or bench/support/other, the order looked up in my order sheet ‘[bench support] Orders’, my refund policy applied (the Google Doc ‘[bench support] Refund policy’; both are in my Drive folder ‘Toolscaled benchmark/support queue’), and a draft reply saved with the decision and the right amount. Never send anything, only drafts. Please do this now for the 30 mails in my inbox with [bench support] in the subject, then set it up so Toolscaled does the same by itself for new support mails.
-
I don’t have the links handy, sorry. Both files are in my Drive folder ‘Toolscaled benchmark/support queue’. Can you find them there?
-
Here are the links. Orders sheet: [link] and refund policy: [link]
-
Yes, go ahead. Drafts only, nothing sent.
-
Yes, please do a test run on today’s mail first. Please leave it off for now; I will turn it on myself later.
-
Yes, that looks right.
-
Is the workflow saved, and what is it called?
Claude did the 30 mails once in the chat and showed each decision with its amount. After the person’s yes, it saved the drafts, saved the workflow, and made the trigger that starts it on new support mail. A test run on the day’s mail held its label and draft and showed them in the chat. The trigger stayed off until the person turned it on.
The numbers
Section titled “The numbers”We ran the workflow on 30 test mails that the build chat never saw. The test mails held three kinds of case that the build did not see: an item change before shipping, a lost parcel with no wish stated, and an order number written with a space. We compare it with an agent that did the same task from the same request. How we measured gives the method.
| For 30 mails | Workflow | Agent |
|---|---|---|
| Time | about 6.5 s | 196 s |
| Cost | 73 credits, about $0.14 | about $0.26 |
| Tool calls | 6 steps a mail | 108 |
| Labels right, of 30 | 29: the phishing mail went to a person | 30 |
| Decisions right, of 17 | 16 to 17 | 16 |
| Replies right, of 13 | 13 | 11 |
| Same decisions in 5 batches | 14 to 15 of 17 in each batch | one run |
The build chat cost about $5.65 at list price, and your chat plan pays for it. At these numbers the build pays back after about 46 batches of 30 mails. In one of the 5 batches, one run waited about 130 s on Gmail, so that batch took 135 s.