Every agency claims to be AI-first. Almost none will tell you which tasks the AI performs, which model, and what happens when it is wrong. The phrase has become a way of sounding modern without committing to anything.
So here is the specific version: what AI actually does inside the Shopify stores we run, what it is genuinely bad at, and where using it costs more than it saves.
Where AI does real work
Product research and competitive scanning
Reading dozens of competitor product pages and pulling out positioning, price bands and description patterns is work that used to take a specialist a full day. It now takes an hour, most of it spent checking output rather than gathering it.
This is the clearest win because the task is genuinely mechanical: read a lot, summarise, find the pattern. Nothing about it requires judgement until the summary exists.
First-draft listing copy
A product description that follows an existing template, in an established brand voice, from a spec sheet, is a good first draft task. The draft is never what ships. A specialist edits it, and the edit is faster than writing from nothing.
The failure mode is obvious and common: publishing the draft. AI-written product copy that nobody edited reads exactly like AI-written product copy that nobody edited, and customers notice before you do.
Support triage
Sorting an inbox — where is my order, refund request, product question, angry — is a classification problem, and classification is what these models are best at. Sorting is safe. Replying is not.
Reporting
Turning a week of orders, ad spend and support volume into a summary a founder reads in three minutes is mechanical work that used to eat a specialist's Friday afternoon.
Where AI is bad, and we do not use it
This list matters more than the one above, because it is the part agencies leave out.
Talking to your customers unsupervised. Not because it cannot produce a plausible reply, but because it produces a plausible reply when it should escalate. A confidently wrong answer about a delayed order is worse than a slow correct one, and it is your brand that carries it.
Anything involving money. Refunds, discounts, price changes. The cost of a mistake is immediate and the benefit of automation is small.
Supplier negotiation and relationships. Obvious, but worth stating, because "AI handles supplier comms" is a claim people make.
Deciding what to sell. It can gather the evidence. The decision is a judgement about your brand, your margins and your appetite for risk.
The rule we actually run on
AI drafts, gathers and sorts. A person decides and sends.
That single line settles most questions about where to use it. If the task ends in something a customer sees or a number that moves money, a specialist owns it. If the task is reading, summarising or classifying, the model does the first pass.
What this is worth in hours
The honest answer is that it compresses the preparation, not the work. Research that took a day takes an hour. A report that took an afternoon takes fifteen minutes. Listing drafts arrive written rather than blank.
What it does not do is remove the specialist. The store still needs someone accountable for the outcome, and the time saved goes into the parts of the job that were being skipped for lack of hours.
Any agency telling you AI lets them run your store with fewer people is telling you they have reduced their own costs. Ask what you get from it.
The questions worth asking any agency claiming to be AI-first
- Which specific tasks does the AI perform end to end, with no human step?
- What happens when it is wrong, and who notices?
- Does anything it produces reach my customers without a person reading it first?
- Are you using AI to improve my store, or to reduce your headcount?
The fourth question is the one that separates the two kinds of AI-first agency, and the answer is usually visible in how quickly the first three get answered.
