Shopify · Comparison

One follows a script.
One reads the catalog.

Both sit in the corner of your storefront and both open with a greeting, which is why they get shopped against each other. The difference only shows up on the question nobody wrote down in advance.

Left: a coordinated colony working shared comb. Right: a single bee circling a closed box
The test

Ask it something nobody anticipated

Every evaluation should start here, because this is the only question that separates the two categories. Everything else — tone, avatar, install time, colour picker — is the same on both.

The scripted bot

Matches the sentence against flows a person authored. No match means a catch-all: a rephrased apology, a link to the contact page, or a menu the shopper already rejected. It cannot be wrong, and it also cannot help.

The store agent

Searches the catalog, resolves a variant, checks stock, returns a checkout link — or says it cannot confirm and offers a human. It handles what nobody scripted, at the cost of needing to be bounded properly.

Side by side

What each is actually good at

CapabilityScripted chatbotStore agent
Fixed FAQ (hours, shipping threshold)
Questions nobody wrote in advance
Search your live catalog
Resolve a specific in-stock variant
Return a real Shopify checkout link
Build a draft order for a quote
Callable by other software (MCP)
Structurally cannot invent an answer
Cheapest option for five known questions

Two of those rows favour the chatbot, and they are real. A scripted flow cannot hallucinate because it can only emit what a person wrote — if your support load is five repeated questions with exact answers, script them and spend nothing else.

The honest split

Most stores need both, doing different jobs

The failure is not picking the wrong one. It is using a single tool for three jobs that have different correctness requirements.

Script the exact answers

Shipping thresholds, returns windows, opening hours. Fixed, high-volume, and they must be exactly right. Cheapest and safest as a script.

Agent the long tail

Fit, compatibility, occasion, "which of these two" — the pre-purchase decisions that are unbounded in phrasing but bounded by your catalog.

Escalate the specific

Orders, refunds, account edge cases. A human, with the conversation attached, so the customer never repeats themselves.

2 of 9
comparison rows favour the scripted bot
Catalog
is what bounds an agent to being right
Both
is usually the correct answer
Questions

Agent vs chatbot, answered

01Is a store agent just a chatbot with an LLM in it?+
The difference is what happens off-script. A scripted chatbot matches an utterance to a flow someone wrote in advance; anything unanticipated falls through to a catch-all. An agent reasons over your live catalog, so it can answer questions nobody wrote down — and, just as importantly, decline the ones it cannot support.
02Are decision-tree chatbots ever the better choice?+
Yes. For a small, fixed set of high-volume questions with exact answers — opening hours, shipping thresholds, a returns window — a scripted flow is cheaper, faster and cannot hallucinate. If your inbox is dominated by five questions, script them. The agent earns its place on the long tail.
03What does "reasoning over the catalog" mean in practice?+
The agent can search products, resolve a specific variant, check what is actually in stock and produce a real checkout link. So it answers "will this fit a 12-month-old and can I get it by Friday" by consulting the catalog, rather than matching that sentence to a pre-written reply about sizing.
04Which one is more likely to say something wrong?+
A scripted bot cannot invent, because it can only emit what was written. That is a genuine advantage and worth stating plainly. An agent trades that safety for coverage, which is only an acceptable trade if answers are bounded to ingested catalog data and everything else escalates. Unbounded, an agent is the riskier product.
05Do I have to choose one?+
No, and most stores should not. Script the handful of exact-answer questions, let the agent take the long tail and the pre-purchase decisions, and escalate account-specific questions to a human. The failure is using one tool for all three jobs.
06What about an agent with no chat window at all?+
That is increasingly the case that matters. On a headless storefront, or when the shopper is talking to an assistant somewhere else entirely, there is no widget to install — but your catalog still needs to be reachable. That is what the MCP tools are for.

Test it on the question you never scripted.

That is the only comparison that separates them. Install the agent, ask it something awkward about your own catalog, and see whether it answers or escalates.