AI for ecommerce: practical starting workflows

Most small shops do not need an AI agent to start. They need help with the work around the catalogue: product copy, reviews, repetitive customer emails, messy product data and working out why things come back. Each job has a clear input, a draft a person can check and a point where a mistake would reach a customer or a regulator.

This guide maps five starting workflows in that shape. It is the broad starting map; our tutorial on building an ecommerce product-question agent is the deep build for one of them. If you have not run an AI pilot before, our guide on how to start using AI at work covers choosing a task and measuring the work that remains. AI Chronicle has not run these workflows in a shop. The example later on is labelled hypothetical, and where we mention a platform feature we report what its documentation says, with the date we read it.

What makes a workflow a safe first step

A good starting workflow has four properties. The input already exists in your shop: product fields, review text, past emails, order and return records. The output is a draft or a suggestion, not a live change. A named person checks it against the source before it is used. And the workflow stops short of anything with money, legal or safety consequences.

That last property matters in retail because so much of what a shop says is regulated: descriptions are advertising, shipping times are promises, and review displays have their own rules in the US and the UK. None of those rules exempts text a model wrote. Shopify's own help page on generated product descriptions is blunt about where responsibility sits:[1]

“You're responsible for the accuracy of all of the content that you publish to your store, even when you use automatic text generation to create it.”

— Shopify Help Center, Automatically generating product descriptions, read 16 October 2026.[1]

The same page warns that generated text "can include things such as product benefits even when you didn't list any explicitly".[1] That is the failure to design around: fluent additions the source data never supported.

Each is a fixed sequence (gather input, draft, check, use), not a system that decides its own steps. Our explainer on AI agents versus AI workflows explains why that predictability helps when you start.

The five workflows at a glance

The order is a suggested sequence. The first two never reach a shopper, so a mistake costs time rather than trust; the last three publish or send words in your shop's name.

1. Returns-reason analysis

Input. Your return records for a period: product, variant, structured reason and any free-text note. Shopify's help centre says return reasons vary by product category, with apparel showing options such as "Too big" or "Too small", and that you can view return reasons in analytics "to identify trends".[2] It also notes that "a refund without a return doesn't record a return reason", so data quality depends on staff creating returns properly.[2]

What AI does. The structured reason says that something was too small; the free-text notes often say why: "sleeves short", "colour darker than photo". A model can group those notes into themes you can act on, such as a size chart to remeasure or a photo to retake.

Human check. Let the model label notes; let a spreadsheet count them. Language models are unreliable at arithmetic over long lists, and a theme's frequency is a number someone will act on. Read a random sample of labelled notes to check the labels match what customers wrote.

Do not automate. Refund and exchange decisions stay in your returns process. And do not let a theme become a conclusion: "five notes say the stitching failed" is a reason to inspect stock, not proof of a supplier defect.

2. Catalogue cleanup

Input. A product export: titles, descriptions, variant options, categories and attributes.

What AI does. It drafts a list of suggested fixes: inconsistent size labels ("M", "Med", "Medium"), materials buried in description text that belong in a field, missing care instructions, titles that break your own pattern. Schema.org's Product vocabulary is a useful checklist of what a clean record can hold in structured form, including size (for example "XL"), material, color, gtin, sku and hasMeasurement for measurements such as "the inseam of pants".[3] Clean fields are what the other four workflows depend on, which is why this one comes early.

Shopify's documentation, read on 16 October 2026, says it uses a product's name, description and images "to generate an AI-powered product category suggestion", which you can accept or replace.[4]

Human check. Approve each proposed change before import, line by line.

Do not automate. Two fields deserve particular care. The same Shopify page says product category determines the rate at which a product is taxed when using Shopify Tax, and that collecting too much or too little "can create financial or legal liability".[4] It also says return reasons are based on product category, so a wrong category corrupts workflow 1.[4] Treat category changes as decisions, not tidying. Never let a model fill in a GTIN or measurement it cannot see in a source document.

3. Product copy from structured data

Input. The structured fields from workflow 2, plus your size chart, care text and brand tone notes. Writing from fields means the model rewrites facts you supply instead of inventing typical ones.

What AI does. It drafts a description in your voice. Shopify's documentation describes a generate-text tool in the product editor and advises giving "a product title and at least 2 product features or keyword items" as a minimum.[1]

Human check. Trace every factual claim in the draft back to a field. If a sentence says "waterproof" and no field says so, cut it. If you list products on Google, check the draft against Merchant Center's rules too. Its description guidance, read on 16 October 2026, asks that information be "consistent" with your title and landing page, and says not to include promotional text such as price, shipping or delivery dates.[5]

Label AI-written descriptions where a platform asks. Google Merchant Center now says descriptions created using generative AI must be submitted in the structured_description attribute, with the digital_source_type value "trained_algorithmic_media".[5] If a person substantially rewrites a draft, decide internally how you will classify it, and apply that rule consistently.

Do not automate. Claims the data does not support, and anything that becomes a health, safety, environmental or performance claim. Those need evidence you hold, not a plausible sentence.

4. Review summaries

Input. Every published review for a product, not a selection.

What AI does. It drafts a short summary of recurring themes, including complaints: "Most reviewers mention the handle stays cool; several say the lid rattles."

Why this one needs the most care. In the UK, the Digital Markets, Competition and Consumers Act 2024 added a banned practice on reviews, in force since 6 April 2025; the CMA describes it as behaviour that is "automatically unfair and illegal".[6][7] The Act defines "consumer review information" as information "derived from, or is influenced by, consumer reviews", and bans publishing it "in a misleading way", which includes giving positive reviews greater prominence than negative ones.[6] The CMA's short guide for businesses lists "review summaries" as an example of consumer review information.[8] Its full guidance goes further: where reviews are found to be fake, publishers should correct the information built on them, naming "AI-generated review summaries" alongside ratings and review counts.[7]

In the US, the FTC's rule on consumer reviews and testimonials took effect in October 2024. The FTC's announcement says it covers reviews by someone who does not exist, "such as AI-generated fake reviews", and bars businesses from writing, creating or selling such reviews.[9] The rule also prohibits misrepresenting that displayed reviews represent most or all of those submitted when reviews are being suppressed based on rating or negative sentiment.[10] The rule does not mention summaries, but the principle carries over: a summary that quietly drops the complaints misleads just as a filtered review page does.

Human check. Compare the summary against the full set of reviews, with negative reviews read deliberately. Regenerate when reviews change, and remove fake reviews from the input before summarising.

Do not automate. Writing, editing, rewording or selecting reviews. A model must never produce review text in a customer's voice, and the summary must not be presented as a review.

5. Support-reply drafts

Input. The customer's message, your published policies, and the facts of their order pulled by a person or a read-only lookup.

What AI does. It drafts a reply that a staff member edits and sends. Shopify Inbox, for example, offers AI-generated suggested replies that its documentation calls "a composing aid for staff", generated from store information such as product listings and policies, and not used for the automatic responses its agent sends.[11] The same page says you are responsible for the accuracy of what you send and need to read a generated reply closely before sending it.[11] For the fully automated version of this idea, an assistant that answers shoppers directly, see our product-question agent tutorial. Grounding answers in your own documents is the technique our RAG explainer covers.

Human check. Read every draft against the order record and policy. Watch for promised dates, invented exceptions and loosened policy wording.

Do not automate. Several replies carry specific legal duties in the US. Under the FTC's Mail, Internet, or Telephone Order Merchandise Rule, any shipping time you state needs a "reasonable basis", including updated shipment information given online.[12] If you learn you cannot ship on time, you must seek the customer's consent to the delay, and the first delay notice must include a revised date (or say you cannot give one) and offer cancellation with "a full and prompt refund".[12] You cannot swap in "materially different" merchandise without the customer's express agreement, and differences in "design, style, color, fabric, or promoted end use" count as material.[12] Rule-required refunds cannot be replaced with store credit.[12] Delay notices, substitutions, refunds and complaints are exactly the replies a person should write or approve from scratch, with the model at most tidying the wording. Other countries have their own rules; check those where you sell.

AI Chronicle

A worked example: a two-person teaware shop

This is a hypothetical example. The shop, its products and its process are invented to show how the workflows fit together. No system was built or run, and no results are claimed.

Imagine a two-person shop selling teapots, cups and loose-leaf tea. Over one month the owners try two workflows, in order.

First, returns-reason analysis. They export three months of returns. The structured reasons are mostly "Damaged or defective" and "Not as described". They ask an AI tool to label each free-text note with one theme from a short list they wrote ("arrived chipped", "smaller than expected", "colour differs", "lid fit", "other") and to quote the phrase that justified each label. They count the labels with a spreadsheet formula and spot-check 20 notes against the labels. Their acceptance rule: if more than a couple of labels in the sample are wrong, they fix the theme list and rerun before trusting any count. Whatever the counts show, the action stays human: remeasure, repackage or rephotograph.

Second, product copy from structured data. For ten teapots, they first move capacity, material and dimensions out of old descriptions into fields (workflow 2, approved line by line). Then they generate drafts from those fields only. Their check is a two-column review: each sentence in the draft on the left, the field that supports it on the right. Sentences without a field are deleted. Because they also list on Google, AI-drafted descriptions go into the structured-description attribute with the AI label.

Review summaries and support drafts wait until the catalogue is clean.

How to roll these out

• One workflow at a time. Start with returns analysis or catalogue cleanup. Our marketing guide covers the adjacent jobs of campaign copy and ads.
• Write the check before the prompt. Decide what a person will compare the draft against, and what counts as a failure.
• Keep a short log. Note what was drafted, who checked it and what changed.
• Count the review time. A workflow that saves drafting time but doubles checking time is not a saving. Our planned analysis on measuring AI productivity looks at how to keep that time visible.
• Stop at the line. Each workflow's "do not automate" column is the boundary. Moving past it is a separate decision, with its own design and tests.

Questions readers commonly ask

Do I have to tell customers that a product description was written with AI?

We found no general rule that says so for a shop's own website. Google Merchant Center does require AI-generated descriptions submitted to it to use the structured-description attribute with an AI label.[5] Whatever you disclose, the description must still be accurate; responsibility for published content stays with you.[1]

Can AI write review responses or ask customers for reviews?

Drafting your shop's public reply to a review is a support-style task; a person should check and send it. Asking customers for reviews is allowed, but the FTC rule prohibits incentives conditioned on reviews "expressing a particular sentiment", and the UK rules cover concealed incentivised reviews.[10][6] Never let a model write review text itself.

Is a review summary risky if every review in it is genuine?

It can be. UK law treats a summary as consumer review information, and publishing it in a misleading way, such as giving positive reviews more prominence than negative ones, is banned.[6] A summary of genuine reviews that leaves out the common complaint can still mislead.

Which workflow should a very small shop start with?

Returns-reason analysis, if you have enough returns with notes to read; otherwise catalogue cleanup. Both stay internal, and both improve the data the customer-facing workflows rely on.

When does this become an agent?

When software chooses its own steps and acts without a person checking each output, such as answering shoppers live. That needs read-only tools, a handoff to a person and a test set.

The next step

Once your product data is clean, the next build is the one shoppers see. Follow the full design for an assistant that answers sizing, materials, stock and shipping questions from your catalogue only: build an ecommerce product-question agent.

About these sources

All pages were accessed on 16 October 2026. Legal sources are cited for the principles they state, not as legal advice; consult a lawyer for the places you sell. Platform documentation establishes what a feature is described as doing, not how well it works. AI Chronicle has not run these workflows; the workflow table, sequence and teaware-shop example are illustrative.

Related reading

• How to measure AI productivity without hiding review work
• Build an invoice-exception investigation agent