AI / ML Nov 22, 2026 • 6 min read

Building a support chat agent for Maybliss: what broke when it started calling Shopify, Gorgias and HubSpot

At Rocket Agency I build AI support agents for small businesses. One of the first was a chat agent for Maybliss, a brand that sells the BlissPod Pro, a magnetic modular pill case for people who take a lot of supplements or travel with them.

2 weeksconsult to first version live
~200chats a week
~50%resolved without a person
20+ hrssaved for the team each week

Most of the work wasn't the model. It was connecting it to the systems the business already runs on, and two of those connections broke in ways I didn't expect.

The setup

  • Shopify for orders, fulfillment and tracking
  • Gorgias as the helpdesk, where chats live and get handed to a person
  • HubSpot for customer records
  • Chat only, on the website

What customers actually ask

Before building anything I went through past tickets. Almost everything fell into a few buckets:

  • Where's my order, and when will it arrive
  • Can I cancel or change my order
  • Discounts and promo codes
  • Product questions: is it BPA-free, what's it made of, how strong are the magnets, are they safe
  • Insurance coverage (whether a pill case can be covered or reimbursed)

That split decides how you build. Product and policy questions get answered from a knowledge base: the inside is BPA-free polymer, the outside is anodized aluminum, lab reports are available on request, and returns are 30 days. Order questions need live data from Shopify. Cancellations and discounts involve money, so they need rules.

Failure 1: order number lookup

Customers type order numbers every way you can think of: "1042", "#1042", "order 1042", or the long ID from the confirmation email. In Shopify, the "#1042" a customer sees is the order's name, which is a different field from the order ID the API usually looks things up by. Early on, the agent would pass whatever the customer typed straight into the lookup, get nothing back, and tell someone with a real order that it couldn't find it.

What I changed:

  • The lookup tool cleans the input itself (strips #, spaces and words like "order") and tries it as an order name first.
  • If that fails it falls back to the customer's email, and checks that the email matches the order before sharing any details.
  • When nothing matches, the tool returns a plain sentence instead of an empty result, like No order found for 1042. Ask for the email used at checkout. The agent asks a follow-up instead of guessing.

The lesson: put the messy parsing in the tool, not in the prompt. The less the model has to decide, the more reliable it gets.

Failure 2: handing out discount codes without approval

This was the bigger one. Customers asked for discounts all the time ("any promo codes?", "I'm a returning customer", "the Black Friday deal didn't apply"). The agent was trying to be helpful, and it gave out discount codes that customer service had never approved for that situation.

It's an easy mistake to make. If a code shows up anywhere in the agent's context, like a knowledge base article, an old ticket or a promo page, the model treats it as something it's allowed to share. Being helpful and protecting margin pull in opposite directions, and by default the model picks helpful.

What I changed:

  • The agent can only mention public promotions that are already on the site, like the current sale or the subscribe-and-save discount.
  • Anything else (one-off codes, "make it right" discounts, fixing a code that didn't apply) creates a Gorgias ticket for a person to approve, and the agent tells the customer someone will follow up.
  • I took private codes out of everything the agent can read. A rule in the prompt isn't enough if the code is sitting right there in its context.
  • I added test conversations that try to talk the agent into a discount, and I run them before every change.

The lesson: anything that costs the business money should be enforced by what the agent can access, not just by what the prompt tells it to do.

What the other 50% looks like

The chats that still go to a person are mostly the ones I'd want a person on anyway: cancellations after an order has shipped, discount requests, and anything unusual. Cancellations and insurance questions follow a clear policy, so the agent explains it and passes the chat to the team in Gorgias with the order already attached. Nobody has to start from scratch.

What I'd tell someone building their first support agent

  1. Read the past tickets before you write a prompt. The question mix tells you what to build.
  2. Expect the lookups to break before the conversation does.
  3. Keep anything that costs money (discounts, refunds, cancellations) behind a person until you have a reason to trust the agent with it.
  4. Don't aim for 100% resolution. Half of 200 chats a week is 20+ hours back for a small team, and a clean handoff beats a confident wrong answer.

I'm looking for AI engineering roles where I can keep building agents like this. If you're hiring, here's my resume, or reach out on LinkedIn.

Tags:
AI AI agents Customer support Shopify
Published Nov 22, 2026