Case study

A shop page that answers in 20 seconds, even after hours.

A six-person engraving shop sold through Messenger and couldn't reply fast enough. The fix was a bot built from the owner's own answers, with the team still in charge of every sale.

79%

faster page response time, 3h 7m to 38m 54s

96.5%

response rate, up from 84.4%

~20 sec

first reply, against a 10.4-hour wait after hours

Engagement: AI Messenger assistant build

Does an AI Messenger chatbot work for a small business?

Yes, as a front desk that answers when the team can't. On Engrave Me Bukidnon's Facebook page, response time went from 3h 7m to 38m 54s and the response rate rose from 84.4% to 96.5%. The bot replies in about 20 seconds, takes the order details, and passes the chat to the team, who still close the sale.

Who they were

Engrave Me Bukidnon, a made-to-order engraving shop in the Philippines. A team of six, working six days a week, making awards, signs, and souvenirs for schools, offices, and events. Most of its sales start in Facebook Messenger.

The problem

Messenger was the shop's main sales channel and its biggest bottleneck. Customers asked about prices, lead times, and photos, then ordered right in the chat. Every message waited until someone on the team was free, and the whole team was busy making orders.

Many buyers write in after their own work day ends. In the first week, 37% of chats had messages sent after hours, and the team's next reply to those messages came a median of 10.4 hours later. For a new customer, that wait was the first answer they got.

Scoping it honestly

A viral video promised a Messenger chatbot in two minutes. Connecting the bot is the quick part. Running one on a real shop's page took a build log from October 9 to October 28, about 12 to 15 hours of build sessions, two rounds of questions for the owner, and a separate practice page for test rounds before any customer saw it.

What we built

  1. 01

    The owner's knowledge file

    Everything the bot knows sits in one plain file, written from the owner's own answers: hours, lead times, payment, shipping, and a price list built from the team's own quotes. The shop's facts are short and simple, so the bot reads the whole file with every reply instead of searching documents.

  2. 02

    Reply engine

    n8n moves each message through the steps, and a small AI model writes the replies in the Bisaya-English mix the team already uses. A check in code blocks any price that isn't on the list, and every price comes with a note that the team gives the final quote.

  3. 03

    Order form

    Code tracks the team's job order form field by field: name, contact number, item, material, and quantity. The bot reads the details back to the customer before it passes them on.

  4. 04

    Team handoff

    A customer can type “team” at any time. The bot keeps answering basic questions while they wait, never pretends to be a person, and goes quiet once the team replies.

  5. 05

    Safety nets

    136 automated tests, an email alert when something errors, a daily health check, a privacy notice before any contact details are asked, and chat data deleted after 90 days.

What changed

The bot handled 68 chats in its first 6.5 days on the real page. It reached the customer before any person in 62 of 65 chats, with a first reply in about 20 seconds. The page's response time fell from 3h 7m to 38m 54s, and its response rate rose from 84.4% to 96.5%. The AI cost for that first week was about 50 US cents.

It didn't take over the sale. The team still replied in 61 of 68 chats. The first week also showed what to fix: returning customers were treated as new ones, and three answers were wrong where the owner's file had gaps. Both were fixed within a week. For handoffs the busy team missed, an SMS alert to the owner is built and waiting on the SMS provider's approval.

A bot is only as good as what the owner writes down. Give it the answers people actually ask for, and keep a person on every sale.

Read the full test on Substack →

More case studies

Your version of this

Every one of these started as a conversation.

Tell me what you're sitting on — expertise, archives, a content bottleneck — and I'll tell you what I'd build.

Not ready to talk yet? Start with the AI Readiness Diagnostic.