A chatbot wired to your real Shopify: orders, catalog and qualified leads

Why your current chatbot annoys customers

Test it yourself: message your own bot with "where's my order 10432?".

If the reply is a generic line inviting you to check the confirmation email or contact support, your bot is resolving nothing. It's adding a step between the customer and the answer. And the customer knows it: that's why they type "human" in the second message.

The problem isn't that the bot is badly written. It's that it has no access to the information. It can sound incredibly natural and still know absolutely nothing about your business.

The technical difference: tools, not scripts

A traditional chatbot works on rules: if it detects certain words, it returns certain text. Everything it can say was written in advance by a person.

A wired agent works differently: it has tools it can decide to use. When someone asks about an order, the agent understands it needs to query the store, makes the query, gets the real data and replies with it. It wasn't written in advance — it was resolved in the moment.

That's the whole difference, and it's enormous. An agent with three well-built tools resolves more than a conversation tree with two hundred branches.

The three tools worth having

1. Live order status

It's the number-one message any store gets, by a lot. The agent queries the real order and replies with the status, the carrier and the estimated date — plus the tracking link if there is one.

With identity verification first. Nobody should be able to look up someone else's order just by typing a number. The agent asks for a piece of data confirming the person writing owns the order before revealing anything. This is not optional.

2. Real catalog search

This is where almost everyone fails in the worst way: the bot makes things up. It recommends a product that no longer exists, quotes an old price, or sends a broken link. A customer who clicks a dead link doesn't come back.

A wired agent queries the live catalog and sends the real link, with the current price and current availability. If the product is out of stock, it says so and offers the alternative that's actually in inventory. It never makes things up, because it isn't remembering: it's querying.

3. Lead qualification and capture

For high-ticket, the conversation doesn't end in a sale: it ends in a qualified lead. The agent asks the questions your rep would ask, saves the answer and notifies whoever has to call — with context, not with a "you have a new message."

Where this lives

In Latin America, the buying conversation happens on WhatsApp, Instagram and Messenger, not in a site widget. An agent that only lives on the web misses most of the real volume.

At Brüm the agent runs on those channels and queries the store live: it answers real order statuses with identity verification, and searches the catalog to send real links with price and availability.

The security rules that aren't up for negotiation

Wiring an agent to your store means giving it access to customer data. Our rules, no exceptions:

• Read-only, dedicated token. The chatbot never uses a token with write permissions, and never shares another system's token. If it's compromised, it's revoked without taking anything else down.
• Verify before revealing. No order data goes out without confirming identity.
• Minimal permissions. The agent can read orders and catalog. It cannot cancel, refund or modify anything.
• Human escalation always available. The moment the customer asks for it, or when the agent isn't sure.

What to measure to know if it works

The vanity metric is "messages handled." It says nothing. The ones that matter:

• Resolution rate without a human — of every 100 conversations, how many ended without escalating.
• Time to first useful reply — not the "hello," but the data the customer asked for.
• Order-status accuracy — audited against the store. It should be nearly perfect, because a wrong answer costs more than no answer.
• Qualified leads generated — how many reached the rep with enough context to call.

How to start without breaking anything

The order that works for us:

• First, measure the real volume. Export the last month's conversations and classify them. Three or four questions almost always concentrate the bulk of the traffic.
• Automate only number one. Order status, almost always. One tool, done well.
• Evaluate against real cases before releasing it. Assemble a set of real conversations and measure how often it answers correctly.
• Release with easy escalation and review the first week's conversations one by one.

Don't try to cover everything from day one. An agent that resolves one thing well is worth more than one that does five things halfway.

Frequently asked questions

Does this replace my support team?

No, it absorbs the repetitive. Most messages are the same status, price or availability question. Once they resolve themselves, your team is freed for the cases that do require judgment — and those get handled better.

What if the agent gets it wrong?

That's why it matters that it queries instead of remembers. An agent that reads the real data from your store has almost no room to make things up. The risk is with the ones that answer from memory.

Do I need Shopify?

Not necessarily, but you need a platform with an API. If your store doesn't expose its data, there's nothing to wire and you'll end up with a canned-response bot again.

How long does it take?

An agent with one well-built tool, evaluated against real cases, goes to production in weeks. What takes time is the evaluation set, not the connection.

NEXT ON THE BLOG → AMAZON MX FOR PREMIUM BRANDS: RETAIL READINESS BEFORE ADS