AI agent for ecommerce: what it does and the hours it saves

In short: an AI agent for ecommerce is a program such as Claude Code, ChatGPT Codex, Gemini CLI or Cursor that takes a task written as one plain sentence and carries it out in your store: prices, stock, reports, customer segments, campaign drafts. In CloudCart it connects with a personal access token (PAT) from the admin. In a sample store with 400 products it saves about 19 hours of work a month.

Most pages on this topic describe chatbots that answer shoppers. This guide is about the agent that works for you inside the admin: what it can do, a worked calculation of the time saved, and the sentences to give it each day, week and month.

What is an AI agent, and how is it different from a chatbot?

An AI agent is software that receives a goal, plans the steps itself and carries them out with the tools it has access to. IBM describes it the same way: a model that reasons, uses tools and acts on a person's behalf. An ordinary chat only replies with text, which you then copy and paste by hand.

For a store the difference is practical. A chat writes you a product description. An agent finds the product, saves the description, changes the SEO title and tells you which other 37 products have no description.

	Chat (ChatGPT or Claude in the browser)	Rule-based automation (workflows, scenarios)	AI agent with access to the store
How you give it a task	A question in the chat	You set up a trigger and steps in advance	One plain sentence
Does it write to the store?	No, you copy by hand	Yes, only along the configured scenario	Yes, within the permissions of its key
A new task you did not plan for	Gives you text	Needs a new scenario	Carries it out straight away
Sees the store's data	Only what you paste in	Only the fields in the scenario	Orders, products, customers, reports

In CloudCart the AI agent works for the merchant and their team. It is separate from Aura Chat, which talks to shoppers on the storefront, and from the AI built into the admin screens. The agent is yours, from your own AI provider, and it works across the whole store.

How does an AI agent work? Model, skills, tools and knowledge

An AI agent has four parts: a language model that reasons, skills that tell it how to work, tools it acts with, and knowledge it answers from. With CloudCart only the model is your choice (Claude, GPT in Codex or Gemini). The rest is ready-made and the agent installs it itself: the CloudCart AI Toolkit skills, the CloudCart CLI that runs requests, the Dev MCP that checks every request to the Admin GraphQL API before it goes out, and the CloudCart reference wiki. The build process business guides describe (architecture, integrations, memory, testing) comes down to one sentence and one key.

Which AI agents work with a CloudCart store?

Six AI agents work with a CloudCart store: Claude Code and Claude Desktop from Anthropic, Codex from OpenAI, Gemini CLI from Google, Cursor and VS Code. They all connect with the same sentence and a personal access token. The difference is what the agent installs and whether it gets the CloudCart skills with built-in rules for bulk changes.

AI agent	From	What it installs for CloudCart	Built-in rules for bulk changes
Claude Code	Anthropic	The cloudcart-plugin plugin: skills, both MCP servers, the reference wiki	Yes
Cursor	Cursor	A plugin: skills and both MCP servers	Yes
Gemini CLI	Google	An extension: skills and both MCP servers	Yes
VS Code	Microsoft	A plugin: skills and both MCP servers	Yes
Codex, Claude Desktop	OpenAI, Anthropic	The Dev MCP, a copy of the reference wiki and four standing rules	No, you ask for the preview yourself

For day-to-day store management, pick an agent with the plugin: Claude Code, Cursor, Gemini CLI or VS Code. The sentence for each agent is on the pages for Claude Code, ChatGPT Codex and Cursor.

Can an AI agent run an online store?

Yes. An AI agent runs a CloudCart store through the Admin GraphQL API, which has more than 1,100 operations for products, variants, prices, stock, orders, customers, discounts, the blog, marketing and settings. Whatever the API offers, the agent can do. Questions about how the platform works it answers from the CloudCart reference wiki.

The tasks fall into six kinds: guidance ("How does the abandoned cart email work?"), catalog edits, bulk changes ("Hide everything that is out of stock"), store data such as orders and low stock, copy written from your real products and sales, and development of a custom Nitrogen storefront.

If you sell across Europe with a separate store for each country, each with its own language and currency, every store gets its own token, and you can ask the same agent to compare them: "Compare October revenue and average order value in the German and Italian stores."

The limits: nothing the API does not offer, no files on your computer unless you give them, and actions that need a two-factor code (such as a customer export, when that step is on) wait for you. The API accepts up to 300 requests a minute per store address, so a big job takes time.

How many hours and how much money does an AI agent save? A worked example

In a sample cosmetics store with 400 products, an AI agent saves about 19.3 hours of manual work a month, which at €20 an hour is about €386. The calculation below is an example. The assumptions are spelled out so you can swap in your own.

Assumptions: 400 active products, a supplier price list twice a month, 40 new products a month, a weekly report, two promotions a month, an hourly cost of €20. The time "with the agent" is the time you spend writing the sentence and reviewing the result.

Task for the month	By hand in the admin	With an AI agent	Saved
New prices from a price list, 2 × 400 products (about 20 seconds per product)	4.4 h	0.5 h	3.9 h
Weekly report: sales, average order value, channels, 4 times	4.0 h	0.7 h	3.3 h
Reorder table for the supplier (low stock), 4 times	3.0 h	0.3 h	2.7 h
Descriptions and SEO titles for 40 new products (10 min by hand, 2 min to review)	6.7 h	1.3 h	5.4 h
Two promotions: discount, code, segment, email draft	3.0 h	0.7 h	2.3 h
Tidying: out-of-stock items, labels, categories	2.0 h	0.3 h	1.7 h
Total	23.1 h	3.8 h	19.3 h (~€386)

Subtract the AI assistant's subscription from the savings. The CloudCart connection costs nothing; you pay the subscription to the provider: Anthropic for Claude, OpenAI for Codex, Google for Gemini, Cursor, or GitHub for VS Code. For example, Claude's Pro plan, which includes Claude Code, is $17 a month billed annually (claude.com/pricing). At the ECB reference rate of 1.1165 dollars to the euro, $17 is about €15, so the net gain in the example is about €371 a month.

The hours are the smaller part of the gain. The bigger part is speed: the promotion goes live the day you decide on it, and a sold-out product is hidden before it takes an order you cannot fulfil.

AI store automation: 6 reasons to run your store through an agent

Running the store through an AI agent pays off when a task spans several screens, touches many products or has to happen today. For a single change, such as one price, the admin is just as fast.

1. One question instead of three screens. Reports, catalog and stock in one answer, for example which best sellers run out within 10 days.
2. A rule instead of a spreadsheet. "15% off the Winter category" needs no export, formulas and import.
3. A new task without a new scenario. Rule-based automation does only what was planned; the agent does what you thought of this morning.
4. Copy from real data. Ads, emails and descriptions come from the products, prices and sales in your store.
5. Decisions backed by numbers. Ask for revenue, conversion rate and average order value before and after every change.
6. Control stays with you. The key decides what the agent can do, and with the CloudCart skills it waits for your "yes" on 10 products or more.

Which metrics to track, and how to get each one with a single sentence

The agent reads the same reports as the admin's Analytics screen, 35 boxes from sales to traffic sources, refreshed every hour.

The seven metrics that say most about an online store's health (how each is calculated and what a normal value looks like is in the article on ecommerce KPIs):

Metric	What it shows	Sentence for the agent
Sales and orders	Revenue and order count for the period, against the previous one	"Show sales and orders for the last 30 days against the 30 before."
Average order value (AOV)	What one order brings in on average	"What is the average order value by week since September?"
Conversion and funnel	The share of sessions that reach an order, and where people drop out	"Show the conversion funnel for October and the step with the biggest drop."
Sales by source	Which channel brings in money: search, Facebook, Instagram, email, ads	"Sales by source and medium for the last 30 days."
Sales by country	Where your orders come from when you ship across borders	"Orders and revenue by country for the quarter."

How conversion rate is calculated and raised is covered in the article on online store conversion rate.

Important: reports only need read access. Create a separate key at the Read Only level for them, so the agent that runs reports cannot change anything in the store.

Email marketing with an AI agent in Marketing Suite

The AI agent prepares an email campaign in CloudCart end to end: the customer segment, the campaign draft and the email template, written from your real products. Marketing Suite is the Marketing → Marketing suite screen in the admin, where you see the results of every channel: email, SMS, Viber and Web push. Sending email is included in every plan, and the number of subscribers in Marketing Suite depends on the plan (see prices and limits by plan).

1. Segment. "Create a segment of customers with at least 2 orders who have not ordered in 90 days."
2. Draft. "Create a campaign draft for this segment." It is saved as a draft and does not go out on its own.
3. Email. "Write the email with our three best sellers from last month and a 10% code."
4. Review and send. You check the copy and recipients in Marketing Suite and send it yourself.
5. Result. A week later: "How much revenue did the campaign bring in?"

Before you send, check that the segment includes only people who have agreed to marketing emails. Across the EU the ePrivacy Directive requires prior consent for marketing email, with one exception: you may email existing customers about your own similar products if they can opt out easily when you collect the address and in every message (Directive 2002/58/EC, Article 13). Which 7 automated emails are worth setting up is in the article on email marketing for an online store, and the screen step by step is in the guide to an email campaign in CloudCart.

Meta, Google Ads, TikTok and Search Console with an AI agent

The agent writes ads from your real products and sales, and matches ad results against store sales. It reads CloudCart data itself; data from Meta Ads Manager, Google Ads, TikTok or Search Console you give it as a CSV export or through the agent's own connection to those services.

Typical sentences: "Write 3 Instagram ad variants for our three best sellers, with price and benefit." "Here is the Google Ads export. Compare spend by campaign with sales from google / cpc in the store." "Here is the Search Console export. Which product pages get many impressions and few clicks? Suggest new SEO titles and save them after my yes."

Pixels are set up once in the admin (Meta Pixel, Conversions API and TikTok Pixel). Channel strategy is in the articles on Meta ads, Google Ads, TikTok Shop and TikTok ads and Google Search Console for an online store.

Why the mobile version is the first thing to check with the agent

The mobile version comes first because across Europe 54.45% of web page views in October 2026 came from phones, against 43.46% from desktops (StatCounter). The agent gives you the exact share for your own store with one sentence: "Sessions by device type for the last 90 days, and the conversion rate for each." If phones bring many sessions and convert worse than desktop, start with the mobile cart and page speed.

The checklist for the phone (speed, buttons, cart, payment) is in the article on the mobile version of an online store. The agent can also find products with no images or very long names, which look poor on a small screen, and fix them after you confirm.

Is it safe to give an AI agent access to your store?

It is safe when you limit the key: in CloudCart an AI agent has exactly the permissions of the personal access token you gave it, and nothing more. A personal access token (PAT) is a private access key that only the store owner can create. It starts with cc_pat_, is shown once, and the platform keeps only a fingerprint of it.

The token is Read Only, Full Access or Custom Permissions across eight resources (products, orders, customers, inventory, discounts, blog, settings, webhooks). You can add an expiry date and up to 20 allowed IP addresses or networks, and switching the key off stops it at once on every computer. Every option is explained in the guide to personal access tokens.

The CloudCart skills for Claude Code, Cursor, Gemini CLI and VS Code add rules for bulk changes. For 10 products or more, the agent shows the count and the first 5 and waits for your "yes". It always asks before deleting. For 100 rows or more it first saves the current values to a CSV file and does a dry run, and above 1,000 rows that file is never skipped, because it is your way back.

Changes through the API have no undo button, just like changes in the admin, so CloudCart offers daily backups as a paid add-on (Settings → Backup & Restore), from which you restore the whole store, or individual data with the partial restore add-on. If the agent reads customer data, remember that under the GDPR personal data must be limited to what the purpose needs (Article 5(1)(c)): a key for ad copy does not need the Customers resource.

How to connect an AI agent to your CloudCart store: 7 steps

1. Store. Sign up for a CloudCart store. New stores pay €1/month for the first 3 months on any plan.
2. Key. In the admin, open the user menu at the bottom left → PAT Tokens → Create Token. Give it a name, for example "Reporting agent", and choose the Read Only level (marked as recommended).
3. Copy. Optionally set an expiry date and IP address under Advanced Settings, click Generate Token and copy the key: it is shown only once.
4. The sentence. The admin dashboard has a "Connect your store to an AI agent" card. Copy the instruction from there, or take the sentence from the CloudCart AI integration page.
5. The agent. Paste the sentence into Claude Code, ChatGPT Codex, Gemini CLI, Cursor or VS Code. It asks for the profile (choose store management), checks for Node.js 20 and git, and installs the CloudCart CLI.
6. Sign-in. The agent asks for the key and the store address, for example mystore.cloudcart.net or your own domain, and confirms the connection.
7. First task. Start with a report: "Show sales by source for the last 30 days." When you are ready for changes, create a second key with custom permissions, for example products and inventory.

Each agent has its own guide: Claude Code for your online store and an online store with ChatGPT and Codex, and for the others there are the pages for Gemini CLI, Cursor and VS Code.

A working rhythm with an AI agent: sentences for the day, week and month

An AI agent gives the most when you ask it on a schedule, not only when something goes wrong. Here are 8 sentences, grouped by when you write them. Each takes under a minute, and the answer comes from your own store's data.

Every morning

1. "How many orders and how much revenue did we have yesterday against the same day last week?"
2. "Which products are below 5 units? Make a reorder table for the supplier."

Every week

3. "Give me a report for last week: sales, orders, average order value and top 5 products against the week before."
4. "Which products have no description or SEO title? Write them and save them after my yes."

Before a promotion

5. "Cut everything in the Winter category by 15% until Sunday and show me the list before you save."
6. "Create a segment of customers who bought in spring but not in autumn, and an email draft for them."

Every month

7. "Here is the new price list. Update the prices by SKU and show me which SKUs are missing from the store."
8. "What is the abandoned cart rate this month against last month, and how many orders did the abandoned cart email bring in?"

Twelve more jobs AI does in the store outside the agent (chat with shoppers, search, product photos) are in the article on AI for an online store.

Mistakes to avoid when working with an AI agent

• Full Access from day one. Start with Read Only and widen the permissions once you have seen how the agent works.
• Changes without a preview. Write "show me first". With agents that lack the CloudCart skills (Claude Desktop, Codex) this is your only safeguard.
• One key for everything. A key per agent and task shows who did what and can be stopped on its own.
• The key in a chat or a file. Treat it like a password; the CloudCart CLI keeps the sign-in on your computer.
• No backup before a big bulk change. Switch on daily backups or at least keep the CSV file with the current values.

Frequently asked questions

Create a store for €1 and connect your AI agent to it. Questions go to support@cloudcart.com.