AI agent use cases in ecommerce: where to start, job by job

AI agents help most in ecommerce where three things are true: the job comes round every day or week, the data it needs already sits in your platform and tools, and a person can check the output quickly before it reaches customers. On most stores that points first to stock mismatches between channels, SEO and product content, merchandising fixes and draft replies to customer questions. Pricing, personalised offers and forecasting can pay off too, but they need cleaner data and firmer rules before you hand them over.

To choose your first one, pick the job where money leaks quietly today, where a change is easy to undo, and where you can compare results against a baseline within a few weeks. If you cannot name the person who will approve the agent's work each week, the job is not ready for an agent yet.

Each section below covers one job: what the agent watches, what it prepares, what a person checks before anything goes live, and what to measure.

Pricing

A pricing agent watches what a trading team would check by hand if it had the time: competitor prices on matched products, your stock depth, sell-through, cost prices and margin, and any promotion already attached to a product. When one of those moves, it prepares a proposed price change with the reasoning written out.

Example: a competitor cuts its price

A competitor cuts the price of one of your best sellers on a Tuesday morning. A rules-only repricer matches it within minutes, whatever that does to margin. A useful agent checks two things first. Does the competitor actually have stock? A low price on something nobody can buy is not competition. And what is your own position? With several weeks of stock and steady sales, matching gives margin away. It then sets out the options (hold, match or move part of the way) with the margin per unit for each and the floor price you set for that product.

Someone on the trading team picks an option and approves it, after confirming the competitor listing is the same product, size and condition, and that no bundle or price promise is tied to it. The approved price goes to every channel that sells the product, so your site and marketplace listings agree. The same agent can model a "what if", such as a short markdown on a slow line, as an estimate to compare options.

Where to draw the line on pricing

Older advice on AI pricing pushed per-person prices, where each shopper sees a different price based on their history. Leave that alone. Shoppers who notice see it as unfair, and it uses their data in a way they did not expect. Loyalty discounts and openly set offers for a group of customers, approved like any other price, do a similar job.

Measure gross margin per order rather than revenue alone, how often an approved change is reversed within a week, and sell-through on repriced products against similar products that kept their price.

Stock and inventory

On a store that sells in several places, most stock problems are sync problems before they are forecasting problems. The platform, the ERP or warehouse system and each marketplace hold their own number, and those numbers drift.

Before. Someone exports stock from the platform and the warehouse system each morning, lines them up in a spreadsheet and fixes the worst gaps by hand. A top seller sells out while a marketplace still shows it as available, and the team finds out from a cancelled order. Reorders go out when a buyer notices a line looking thin.

After. The agent compares stock across every connected system through the day and reports each mismatch with both numbers and the system it believes is right. It flags products that will run out before the next delivery lands, from current sell-through and the supplier's lead time. Once a week it drafts a reorder list: product, quantity, supplier and the reason for each line.

The agent does not place the order. A buyer checks the draft against what the data cannot see: minimum order quantities, a supplier's price rise, cash this month, a product about to be discontinued. For sync fixes, a person confirms which system is the source of truth before any number is overwritten, because an agent that corrects the wrong side copies the error to every channel.

Measure days out of stock on your top sellers, orders cancelled because of stock errors, open mismatches between systems, and the age of stock that has stopped selling. Count the reorder drafts approved without changes, too: that tells you when the agent has earned more trust.

Personalisation

Take a customer who bought a coffee machine from you in the spring, has looked at grinders twice this week and opens your emails on a phone in the evening. A segment-based system files them under "coffee" and sends the same newsletter as everyone else in that group. An agent with the whole history can prepare something better: recommendations led by grinders and descaling tablets rather than another machine, and an email built around them, timed for when this customer usually reads.

To do that, the agent watches purchase history, browsing, recommendation clicks, returns, email engagement and marketing consent across your website, app and marketplaces. That shared view is what stops your app recommending the product a customer has just returned on your site.

Because it works across every customer, approval sits at the level of rules and samples. Your team approves what counts as a related product, how often anyone can be contacted and what the copy may say, then spot-checks a sample before each send. Keep these out of the recommendations:

products that are out of stock or about to be
items the customer has just bought or sent back
anything personal they would not want shown back to them, such as health products
customers who have not agreed to marketing
products whose return rate wipes out the margin

Personalisation results are easy to overstate, because people who click recommendations were often going to buy anyway. Hold back a slice of customers who get the standard experience, then compare add-to-basket rate, conversion and repeat purchase between the two groups. Watch unsubscribes and complaints alongside them.

Customer service

Customer service is where agents are easiest to picture and easiest to get wrong. The useful version starts with context. When a ticket arrives, the agent pulls in the customer's orders, delivery tracking, previous contacts and the relevant returns policy, so nobody asks for an order number twice.

What the agent can prepare:

a tag and priority for each incoming ticket
a draft reply that quotes the actual order and delivery facts
a refund or replacement proposal, with the line of policy it relied on
a handover note for complex cases, summarising what has been tried
a weekly list of the most common reasons customers got in touch

What stays with a person:

sending replies, at least until drafts regularly go out without edits
every refund, credit or change to an account
complaints, anything with a legal angle, and customers who may be vulnerable
changes to the policy itself

The weekly list of contact reasons is often worth more than the replies, because it points at fixes elsewhere. If many tickets ask where an order is, the dispatch email may need work. If people ask about delivery costs, check that product pages match what checkout charges: we have seen product pages promise free delivery above one amount while orders showed a different threshold in use.

Measure first response time, time to resolution, repeat contacts about the same order, the share of drafts sent without edits, and satisfaction on tickets the agent drafted against those it did not.

Merchandising

A merchandising agent watches what sells, what is in stock, what earns margin and what shoppers type into your site search. From that it prepares changes to what shoppers see first: the sort order of a category page, the products in a homepage block, cross-sell pairings and the products a promotion should cover. A good one also catches the embarrassing problems: a promoted product that is out of stock, a collection full of hidden products, a seasonal page still live a month after the season ended.

Merchandising changes are usually quick to reverse, which makes them a sensible early job. Before one goes live, someone should still confirm:

every featured product is in stock in the sizes and colours people actually buy
the promotion leaves margin above your floor once delivery costs are counted
it does not clash with another live offer or a supplier's terms on price or placement
the end date is scheduled rather than left to memory
the page has been checked on a phone as well as a desktop preview
the previous arrangement can be put back in one step

Measure revenue and margin per category page visit, average order value on pages that show cross-sells, sell-through of promoted stock, and margin on promoted orders after the discount.

SEO and content

SEO suits agents because the work is large, repetitive and easy to review. A catalogue of a few thousand products needs titles, descriptions, category copy and product data that search engines and AI assistants can read, and very few teams keep all of it current by hand. A weekly cycle looks like this:

The agent reads your Search Console data, catalogue and existing content, and lists pages losing clicks, products with missing or duplicated titles, and broken links, including old URLs that started failing after a replatform.
It drafts the fixes: titles and descriptions in your brand voice, category copy, a redirect list for broken URLs, and briefs or drafts for articles answering common customer questions you have no page for.
A person reviews the batch: every claim true (essential in regulated categories), the voice recognisably yours, no URL changed without a redirect, nothing stuffed with keywords.
Approved changes are published, and the agent records what changed and when.
A few weeks later it compares clicks, impressions and organic sales on the changed pages with similar pages left alone.

Search does not change by the minute, so a weekly cycle suits most stores.

The same work matters for AI answers. Shoppers now ask ChatGPT and other assistants what to buy, and those answers favour stores whose product data is accurate and current. Getting that right improves the odds of being cited; nobody can promise a citation. At Vortex IQ this is the job of SEO and AI Visibility (Beacon), which prepares titles, descriptions, product data and briefs across the catalogue for your specialist to review. To see what assistants can read on your store today, the agent-ready store audit covers the checks.

Cart abandonment

Abandoned carts have two causes that need different work: shoppers who were never going to buy today, and shoppers your checkout lost. Start with the second, because it is within your control.

For checkout problems, an agent can run real shopper journeys in a browser (add to cart, apply a code, choose delivery, pay) and report where the journey breaks or slows, with screenshots. At Vortex IQ, Store Experience (Prism) does this. An agent watching orders can also flag failed payments nobody followed up and delivery costs that first appear at checkout.

For shoppers who leave anyway, four approaches still hold, each with its own check:

Spotting a shopper about to leave. A long pause on the checkout page can trigger a help message. A person approves the wording and decides whether it may ever carry an offer.
Follow-up by basket and behaviour. The agent groups abandoned carts (high value, first visit, returning customer) and drafts a reminder for each group. Someone checks consent, tone and how many reminders one person can receive.
Offers based on risk. A discount on every abandoned cart teaches customers to abandon on purpose. The agent proposes offers only for carts unlikely to come back otherwise, with the margin cost shown, and a person sets the limit.
Timing. Some customers respond within the hour, others the next day. The agent proposes send times per group, and the team approves the schedule.

Measure recovered orders against a holdout group that gets no reminder, discount cost per recovered order, checkout completion rate, and unsubscribes from reminder emails.

Forecasting

A forecasting agent reads order history, seasonality, your promotions calendar, planned marketing, supplier lead times and outside factors where they matter, such as weather for garden furniture. It prepares a demand forecast by product or category, ideally as a range, and flags where this week's forecast has moved sharply from last week's, with the likely reason.

A forecast earns its keep through the decisions it changes: reorder quantities, stock for peak, staffing, ad budgets. So the agent should also list where the forecast and your current plan disagree, for the buyer or planner to review. Before acting on any forecast, a person should ask:

Does it know what history cannot show, such as a new launch, a discontinued line, a price change or a large promotion next month?
Which products moved most since the last forecast, and does the reason make sense to someone who knows the range?
How far out was last month's forecast on the top products, and in which direction?
Is the range wide because demand is uncertain, or because the data has gaps?
Which decision changes if this number is right?

Measure forecast error on your top products month by month, bias (whether it runs consistently high or low), and the reorder decisions it changed, alongside stockouts and overstock on those products.

Real-time data

Every job above needs accurate data, and some also need current data. Those are different problems, so know which one you have before you start.

How fresh does each job need its data?

Minutes: stock across channels, checkout and payment failures, site errors. A problem here costs sales for as long as it lasts.
Hours: competitor prices, customer service context, abandoned carts, live promotions.
Days to weeks: SEO and content, forecasting, merchandising reviews.

Fresh data is wasted if it is wrong. Mismatches we come across include ad platforms claiming more sales than the store recorded in total, heavy traffic with almost no orders (bots or broken tracking), and far fewer products available to buy than the catalogue suggests, because many are hidden.

Here the agent watches the feeds themselves: platform, ERP, marketplaces, ads, analytics and help desk. The Model Context Protocol (MCP) gives agents one standard way to connect to many tools, where each used to need its own integration. Vortex IQ reaches these systems through over 200 connectors in the Vortex IQ connector catalogue.

When the agent finds a gap, it prepares an alert with the evidence (both numbers, where each came from, and since when), and a person decides which system is right before anything is corrected. At Vortex IQ, Store Health (Pulse) is the crew that finds store issues, verifies them and writes them up with evidence. Measure the time between a problem starting and someone knowing about it, and the mismatches still open at the end of each week.

How to choose your first use case

List where money leaks quietly today. Stock errors that cancel orders, pages losing search traffic, a checkout step that fails on mobile, tickets asking the same question every day. Choose from that list, not from a list of AI features.
Score each job on how much money is at stake, how clean the data is and how easily a change can be undone. A wrong product title can be changed back in minutes. An email cannot be unsent.
Name the approver: one person with the authority to say yes and the time to review every week. Without that, the agent's drafts become the new backlog.
Agree how a change is undone before the first one goes live: a platform undo, a pull request your developers or agency review, or written steps.
Set a baseline and a holdout, so you can tell the agent's effect from the season's.
Run it for a fixed period, then decide whether to widen it, change it or stop.

Leave per-person pricing, and anything that spends money or messages customers without review, until your first job has a track record.

This is how we work at Vortex IQ. Specialist crews find the problem in your own data, prepare the exact change, and apply it once approved, with an undo point where the platform supports one, as a pull request, or as clear steps. Then they check the result. It proposes. Your team approves. Nothing goes live without your say-so. For the full picture, read AI agents for ecommerce, or start with a free store audit to find where your own store leaks money first.

Frequently asked questions

What is an AI agent in ecommerce?

Software that reads data from your store and connected tools, works out what a specific job needs, and prepares the change: a price, a reorder, a reply, a product description. A dashboard leaves that work with you, and a simple automation only follows fixed rules. A well-run agent also waits for approval before anything reaches customers.

Can AI agents change prices or send emails without asking?

Many tools allow it. Do not start that way. Begin with drafts and proposals that a person approves. Some teams later let one narrow, low-risk action run inside limits they set, such as order-tracking replies in approved wording, and that should be a decision made on evidence.

Are AI agents worth it for a small team?

Often, because a small team is the one with no time to look. The limit is review time rather than store size: someone still has to approve the work each week. Start with one job that saves more time than it takes to check, and add the next only when that one runs smoothly.

Do we need clean data before using AI agents?

You need accurate data for the job you pick, not for the whole business. An SEO agent needs your catalogue and Search Console. A stock agent needs your platform, warehouse system and marketplaces to agree. Check that slice first, because an agent working from wrong numbers produces confident, wrong proposals.

How do we know if an AI agent is working?

Record a baseline before you start and keep a holdout where you can: pages, products or customers the agent does not touch. Compare the two over a fixed period, and track how often its proposals are approved without edits.

Will AI agents replace our agency or team?

No. Agents take on the gathering, checking and drafting that fills a working week. Your team decides what goes live, and your agency or developers still review code changes as pull requests.

AI agent use cases in ecommerce, job by job