Related products that actually fit
Most "you may also like" widgets guess. Set related-product rules in plain words, like same brand or parts that fit, and measure what they really sell.
The promise of AI in customer service is usually pitched as "let the bot answer." The version that actually works in a mid-market store is quieter: the AI writes the first draft with the order in front of it, and a person reads it before it goes out. This post walks through what that looks like across a normal day in the Communications Center, from the morning queue to the comment that comes in at nine at night.
The inbox opens with every channel in one list: email to your support and sales addresses, Facebook and Instagram comments, YouTube comments, and replies to the emails your store sent overnight. Each row shows the customer, a subject, a status (open, pending, resolved, snoozed) and a small marker when the AI has a draft waiting.
Above the list are three chips. Unread shows new customer replies. AI pending shows threads where a draft is ready for review. SLA under one hour shows customers who have waited too long. Most mornings start with the third chip and then the second.
Say your store gets 120 messages a day. By 8:30 the AI has read the overnight ones and drafted replies for most of them. Your job is not to write. It is to review.
Open a thread. On the right is the AI rail: a summary of the conversation, the draft, and small chips showing which skills the AI used to write it. "Order Status" means it looked up the order. "Returns Handling" means it checked the policy and the order's eligibility. "Product Recommendation" means it searched the catalog.
Below the draft is the customer card, pulled from the CRM: what they bought, when, how many orders, any open cart, any prior conversations. The draft was written with that card in front of it, so "your order shipped Tuesday and should arrive Thursday" is a fact, not a guess.
Read the draft. Edit if needed. Click Approve and Send. That is the whole loop for a routine message. Thirty seconds instead of four minutes.
Not every thread gets a draft. Some inboxes are set to Off on purpose: legal, executive escalations, anything you always want written by hand. And in inboxes set to draft, the AI declines when it is not confident, so the thread arrives with no draft and a note.
That is the design working. The three per-inbox settings are:
| Setting | What happens |
|---|---|
| Off | No drafts. A person writes every reply. |
| Draft only | The AI drafts every new customer message. A person approves before it sends. |
| Auto-send | The AI drafts and sends without review. |
Draft only is the default for shared support inboxes. Auto-send is for narrow, repetitive patterns where the grounding data is strong, for example a dedicated order-status inbox, and even then with a person watching for outliers.
An email arrives asking whether a particular part fits a particular model and what else the customer would need. A smart routing rule sends it to the Sales inbox, because it contains one clear sales intent. If it had asked about a return and a new purchase in the same message, it would have stayed in General, because mixed intents are not guessed.
The Sales inbox has a different persona (more enthusiastic, still accurate) and a restricted skill list: product recommendation and fitment, never warranty claims. The draft pulls live stock and price for the suggested variant from the store, so it does not recommend something that sold out this morning.
The rep reviews it, adds one line about a promotion, and sends.
"I want to return this and find something better for my truck." Two skills fire in one draft: Returns Handling explains the policy and confirms the order is inside the window, and Product Recommendation suggests the alternative. The draft is longer because this inbox is set to Detailed rather than Concise, since return conversations tend to spawn follow-ups if the first reply is too short.
The person reviewing it can see both skill chips, so if the wrong one fired it would be obvious. It did not. Approve, send.
A comment lands on an Instagram post: "Do you ship to Canada?" It appears as an Instagram conversation in the Facebook Comments and Instagram Comments inboxes that were created automatically when the channel was connected.
Drafts for public comments follow stricter rules. Plain text only. No order lookups. Never include an order number, a tracking link, an email address or a phone number, because the reply is public. The draft says yes, with a link to the shipping page, and invites the customer to email for anything order-specific.
When the reply posts, the platform reports the store's own comment back as a new event. The inbox recognizes it as your outbound reply and resolves the thread rather than treating it as a new inbound message. The AI never drafts a reply to its own reply.
Overnight, the outbound system sent a cart-recovery email to a shopper who reached checkout twice and left. She replied: "What's the shipping cost to Ohio?"
The reply lands in the inbox with the cart, the email that was sent, and the reason it was sent attached. The draft answers the shipping question with the actual rate for that cart. Because the reason she left was, apparently, shipping, that fact also feeds a weekly digest: if enough people ask about shipping, the shipping line moves up in the next version of the email.
A message arrives. The AI drafts it. It sits in the AI pending chip until someone opens the inbox in the morning. Nothing is sent. If the customer's wait crosses an hour, the SLA chip will show it first thing.
If drafting had silently stopped for some reason, the admins would know. The system watches its own drafting: if customer messages keep arriving in inboxes with drafting turned on and no drafts appear, it raises an alert and assigns a task to each admin. A quiet evening never triggers it. Messages without drafts do.
The reply count per person goes up, but that is not the point. Three things change:
The Communications Center reads the same order record as the pixel and the CRM. That is why the drafts know what they know. And when the owner wants to set up a new inbox, its skills and its routing rule, they can describe it to Claude or ChatGPT through the workspace connector, see the exact configuration before it is created, and confirm once.
Want to see your inbox with drafts in it? Talk to us.
Want this for your business? Talk to us How the pixel works
Most "you may also like" widgets guess. Set related-product rules in plain words, like same brand or parts that fit, and measure what they really sell.
What AI assistants read, how to feed them one clean catalog, and how to see the sales ChatGPT, Gemini and Perplexity send you.
The pixel steers the feeds, measures related products and tells the Outbox what each shopper left behind. CRM lists become ad audiences. The company brain keeps learning.
One or two a month. No hype, no spam.