Best AI for Ecommerce Operations: Orders, Stock, Returns, Suppliers

For the back office of an online store, use your platform's assistant (Shopify Sidekick) for order and stock questions, a helpdesk or chat agent (Gorgias or Tidio Lyro) for returns and order-status queries, and a general assistant for supplier emails and reorder checks. Keep a person on refunds and purchase orders until the error rate is low.

This is the operations side, not marketing. The tools here are judged on whether they get the right answer from your order and stock data, how they fail, and what each costs per month at a store sending a few hundred orders.

Follow me on Instagram@sagnikteaches

Four back-office jobs, four different tools

Operations tasks differ in what data they need, so one tool rarely fits all.

Connect on LinkedInSagnik Bhattacharya

Notice that the sequence in each row is the same: set the rules, let AI prepare, a person approves. The tools differ in how much of the preparation they can do alone.

Subscribe on YouTube@sagnikteaches

Orders: what Sidekick can and cannot do

Shopify's help page describes Sidekick as able to help with tasks such as editing orders, to build Shopify Flow workflows that automate repeated work, and to write and run code to forecast, compare or summarise store data; it also surfaces recommendations through Sidekick Pulse on Shopify Home. It adds that in some stores Sidekick cannot run code, in which case it uses a skill only when you select it. Sidekick is included in every Shopify plan, but Shopify says features and usage limits vary by plan, so test the jobs below on your own store.

Practical uses: "show unfulfilled orders older than 3 days", "which products had the most refunds last month", and "build a flow that tags orders over $200 for review". Each answer should be checked against a filtered order list, because it is cheap to verify. How to use Shopify Sidekick to run your store faster has more prompts.

A first test for any order assistant

Ask it three questions to which you already know the answer: the number of orders yesterday, your top seller this month, and the order value of a named recent order. If it gets one wrong, you have learned how far to trust it, at no cost.

Stock: reorder checks that a person can verify

Stocky, the inventory app Shopify used to offer, was withdrawn from the App Store and switched off for existing stores at the end of August 2026; purchase orders are now handled in Shopify admin itself. There you save a draft, export it as a PDF to send to the supplier, mark it as ordered once the supplier confirms, and receive the goods through an inventory transfer. Shopify itself does not write or send the covering email, and that gap is where a general assistant helps: it prepares the email, and you send it with the PDF attached.

The formula to give the assistant

Do not ask any AI "what should I reorder?" without giving it a rule. The standard rule is a reorder point:

Reorder point = average daily sales x (supplier lead time in days + safety days)
Days of cover = units on hand / average daily sales
Order quantity = average daily sales x days you want to cover - units on hand - units on order

An illustrative example for a pet-treat line that sells 6 a day, with a 14-day supplier lead time and 4 days of safety: reorder point = 6 x (14 + 4) = 108 units. With 120 on hand and nothing on order, days of cover = 120 / 6 = 20 days, so the product is still above the reorder point, and will cross it in 2 days (120 to 108 is 12 units). If you want stock on hand plus stock on order to cover 30 days from today, order quantity = 6 x 30 - 120 - 0 = 60 units. Note what that means on arrival: after 14 days of selling, about 36 of the 120 remain, so the delivery leaves roughly 96 units, or 16 days of cover; if that feels thin, raise the days you want to cover, not the safety days. Put these formulas in a sheet so the arithmetic is exact, and let the assistant do the part it does well, which is reading the list and writing the summary.

A prompt for a weekly reorder check

Here is a CSV export of my stock: SKU, name, on hand, on order, avg daily sales (last 30 days),
supplier, lead time days. Safety days = 4.
For each row, calculate nothing yourself; I will add the formula columns.
Instead: list SKUs where the data looks wrong (negative stock, zero sales with high stock,
lead time missing), then group the SKUs by supplier so I can build one order per supplier.

The illustrative result flags three oddities: one SKU with -3 on hand (a stock count error), a seasonal item with 0 sales over 30 days but 90 units in stock, and a supplier with no lead time recorded. That is the sort of cleaning a tired owner misses, and it makes the reorder sheet trustworthy.

Returns: write the rules before choosing a tool

Shopify's returns help page covers the building blocks: start a return or exchange from the order, send shipping instructions with an optional label, inspect and refund (in full or in part), and set return and cancellation rules saying which items are eligible and how long customers have. It also describes self-serve returns, where customers submit requests from their order status page. Restocking sits in the processing step: when you process a return you can choose a location to restock each item, which is optional, so decide in your rules whether a returned item goes back on sale or is written off.

An AI agent is only as good as the rules behind it. Before connecting anything, write one page: the return window, what condition items must be in, who pays postage, which products are final sale, and what evidence you need for damage. The method is set out in how to automate returns and refunds with clear AI rules.

Gorgias or Tidio Lyro for returns

Gorgias's pricing page lists helpdesk plans from Starter at $10 a month for 50 tickets and Basic at $60 a month ($50 on annual billing) for 300 tickets. Its AI Agent goes on top, priced by the automated interactions you commit to, where an interaction is a request resolved without a person: the comparison table lists $1.00 each on Starter and $0.90 on Basic and Pro, and each interaction past your commitment costs $1.50. Tidio sells Lyro from $32.50 a month on yearly billing for 50 conversations, and your first 50 Lyro conversations are free, once.

Choose Gorgias when you already live in a ticket queue, need order data beside each message and want refunds and edits from inside it. Choose Lyro when your main need is a chat agent that answers from your policy pages. A broader comparison, with costs per resolved chat, is in what each ecommerce chatbot platform charges.

A worked sum: the cost of returns handling

Say a small store ships 750 orders a month with an 8% return rate: 60 returns. Each takes about 6 minutes of human time (reading the email, checking the order, replying, issuing a label), so the total is 6 hours a month. If an agent handles the policy questions and label requests for 40 of those at about $1.00 each (the top of Gorgias's listed range, provided you commit to that volume rather than paying $1.50 overage), the fee is about $40 and the saving about 4 hours; at a value of $20 an hour that is $80 of time. The margin is thin, which is the point: returns automation pays when volume is higher or when the clock matters to customers. Run your own numbers before subscribing.

A returns triage prompt

Policy: 30 days from delivery, unworn, tags on, customer pays postage unless faulty.
Final sale: underwear, personalised items.
Customer message: [paste]. Order data: [paste date delivered, items, price].
Decide: eligible / not eligible / needs a person. Give the reason in one sentence,
quote the policy line you used, and draft a reply with [first name] and [your name].

An illustrative outcome: a customer writes on day 34 about a pair of boots that split at the seam. The assistant returns "needs a person: outside 30 days but the reason is a fault" and drafts a polite request for a photo. The "needs a person" category is the important one. A system that forces every case into eligible or not will make mistakes on exactly the cases that damage reviews.

Suppliers: chasing, ordering and queries

Supplier emails are repetitive and tone-sensitive, ideal for drafting. The assistant needs the purchase order, the promised date and the history.

Write a short, firm, polite email to a supplier. Facts: PO 1042, 6 cases of leather dog leads,
promised delivery 12 November, not received on 19 November, no tracking sent.
Ask for tracking today and a firm date. Mention that two customer orders are waiting.
Tone: plain, no threats. Sign off as [your name]. Under 100 words.

The illustrative reply is a tight paragraph with the PO number, the missed date, the two requests and a sentence about the waiting orders. What you fix: any invented detail (it might say "as discussed on the phone") and any statement about penalties you have not agreed. Never let it quote a price or promise payment terms from memory. For comparing offers rather than chasing, see how to compare supplier quotes side by side with AI.

Cost of the general assistant

A general assistant costs $20 a month on ChatGPT Plus, or $20 monthly ($17 on annual billing) for Claude Pro. For stock exports with supplier names and prices, a business plan whose terms say content is not used for training by default is the safer choice; ask whether your supplier agreement allows you to share their price lists with a third party.

A small online optician's lens-and-frame store

An online seller of glasses frames (illustrative) with 90 SKUs and four suppliers puts the pieces together in one weekly routine of forty minutes:

1. Export stock and last-30-day sales on Monday morning.
2. Paste into the assistant with the data-cleaning prompt above; fix the three or four oddities it flags.
3. Calculate reorder points in the sheet and filter to SKUs below them.
4. Group by supplier and ask the assistant to draft one purchase email each, from the sheet rows.
5. Read each email, correct quantities, send. Log the PO numbers.

The first week took 90 minutes. By the fourth it took 40, and one stock-out was caught before it happened: a frame with 9 days of cover and a 14-day lead time, which the old process would have found when customers complained.

Order exceptions: the daily trickle that eats an hour

Most stores spend their back-office time not on normal orders but on exceptions: an address typed wrongly, a payment flagged, a customer who wants to add an item after paying, a parcel marked delivered that has not arrived. These are good candidates for prepared replies and checklists, because the steps repeat even though each case looks different.

A before and after for "wrong address"

Before: a customer emails "I put the wrong flat number", and the owner opens the order, checks if it has shipped, edits the address, emails back and notes it, about eight minutes. After: the owner keeps a one-page checklist and a reply template, and uses the assistant to fill the template from the order data. The checklist reads: has the order been fulfilled (if yes, contact the carrier, do not edit); is the new address in the same delivery zone (if not, flag for a person); reply with the confirmed address, repeated back, so the customer can spot a second typo. An assistant given those three lines and the order record produces a two-sentence reply in seconds, and the owner reads it before sending. The saving is a few minutes a case, which matters when it happens ten times a week.

Parcel marked delivered, not received

The prompt is simple and the output must be checked closely:

Customer says order 5531 was marked delivered on 3 November but not received.
Carrier tracking says: delivered, left in a safe place, 11:42.
My policy: ask the customer to check with neighbours and around the property for 2 working days,
then open a carrier claim; resend or refund after the claim, not before, for orders over $50.
Write a calm reply that follows the policy. Do not promise a refund or resend.

An illustrative draft thanks the customer, quotes the delivery time from tracking, asks them to check for two working days, and says what happens next. What you fix: any sentence that sounds like blame, and any promise the policy does not make. If your policy is to refund quickly for low-value orders, put that in the prompt; the assistant has no way of guessing.

Using returns data to fix the cause

The most valuable operations job is not handling returns; it is reducing them. A monthly export of return reasons, product and size, pasted into an assistant with the question "group these reasons into five categories and show the products appearing most often in each", gives a ranked list in minutes. A small store (illustrative) with 60 returns finds that 21 relate to one pair of trousers described as "regular fit" that runs small. The fix is a line in the product description and a size note in the photos, not a better refund process. The assistant's categories need checking against the raw reasons, but the pattern is easy to confirm in a pivot table.

Backorders and delayed stock

When a line sells out, customers still email. An assistant can draft a standard backorder message from the stock sheet: the product, the expected restock date from the supplier's last email, and the option to cancel. The rule to enforce is that the date comes from the supplier's confirmation, not from the assistant. A good habit is to put "expected restock (from supplier email of [date])" in the sheet, so the draft carries its source. If a date has slipped twice, say so honestly in the message.

For a ready template that fills from stock data, see a backorder notification email template, which suits stores that also sell to trade.

A weekly operations review in fifteen minutes

Once the pieces exist, a short review stops them drifting. Each Friday, the owner answers five questions with the help of the assistant and the exports already to hand. How many orders were delayed past the promised date, and why? Which three SKUs are closest to their reorder point? How many returns arrived, and what were the top two reasons? Which supplier is late on a purchase order? Which customer message did the chat agent hand over, and was the hand-over right? Writing the answers in a running note, one line each, turns scattered fixes into a record. After a quarter the note shows which supplier is unreliable, which product page causes returns and which agent rule needs tightening, and that evidence is what justifies, or rules out, paying for another tool.

Where back-office AI goes wrong

• Stale data. An agent answering from last night's export tells a customer an item is in stock when it sold at noon. Connect live data or add a line saying stock is confirmed at checkout.
• Policy invention. A chat agent without firm boundaries will occasionally promise a refund you do not offer. Restrict it to published pages and ask it to hand over when unsure.
• Silent failures in flows. A Flow that tags orders can stop when a condition changes. Check the run history weekly.
• Over-ordering. A reorder suggestion based on a promotional spike will tell you to buy too much. Exclude promotion weeks from the average or cap the order at a number you approve.
• Customer data in the wrong place. Never paste names and addresses into a consumer chat when a stripped order ID will do.

Which to start with, by store size

The thresholds are rough. A store selling high-ticket items with many questions per order needs help sooner than one selling low-priced basics. For the cost of phone calls specifically, the cost per order-status call with an AI voice agent is a separate calculation.

Proving the back-office changes paid off

1. Time the weekly stock routine before and after, and write both numbers down.
2. Count stock-outs on your top 20 SKUs for a month. The target is none you did not choose.
3. Sample ten returns handled by the agent and compare each outcome with your policy.
4. Track average first-reply time on returns emails. It should fall.
5. Read twenty chat transcripts a month. Any invented policy means tightening the instructions.

If you would like help mapping which of these jobs to automate first, that is what my AI implementation consultation covers, at the contact page.

Back-office AI questions from store owners

Can AI process refunds on its own?

Technically some helpdesk agents can take actions, but most small stores should keep refunds under human approval at first. Let AI collect the order number, reason and photos, check them against your policy and prepare the case. A person clicks the refund until the error rate is low.

Which tool is best for returns, Gorgias or Tidio Lyro?

Gorgias suits a store that already handles hundreds of tickets a month and wants order data beside every message. Tidio Lyro suits a smaller store that wants a chat agent answering policy questions. Neither removes the need to set clear return rules first.

Can I let an AI order stock from suppliers?

Let it prepare, not send. An assistant can calculate which products are below their reorder point and draft the supplier email, but a person should approve quantities. Mistaken orders are expensive, and Shopify's own purchase orders are exported as a PDF for you to send anyway.

Do I need a developer to connect these tools?

Usually no. Helpdesk and chat tools connect to your store through an app install. Stock checks can start as an exported CSV pasted into an assistant. Automation tools help later, but the first version works with a weekly export and thirty minutes of your time.

How do I check the AI is not making things up about orders?

Test it with ten real past orders, including awkward ones, and compare each answer with your admin. Restrict it to data from your store and published policies, and read a week of transcripts. Any answer that quotes a date or amount should match the order record.

Further reads

• Ecommerce Customer Service Outsourcing vs AI: Costs for a Small Brand — Costs of outsourcing support against AI.
• Best AI Assistant for Shopify: Sidekick, ChatGPT or Claude by Job — Which assistant for which store job.
• AI for Ecommerce Analytics: Shopify Reports, GA4 or ChatGPT First? — Store numbers: reports, GA4 or an assistant.
• Can an AI Chatbot Handle Order Tracking and Returns? — Chatbots for tracking and returns questions.
• AI Agents for Distributors: Answer Stock and Delivery Queries Safely — Answering stock and delivery queries safely.
• Best AI Tools for a Small Ecommerce Store: 7 Worth Paying For — Which store AI tools are worth the subscription.
• WooCommerce AI Agent: Which Chatbot Fits a Small Store? — If your store runs on WooCommerce.
• How to Spot Patterns in Complaints, Returns, and Defects With AI — Spotting patterns across returns, complaints and defects.
• AI Fit Shoes Plugin: Can a Size Finder Cut Returns for a Shoe Shop? — How shoe size finders work, what data they need, and a before-and-after method to test whether one cuts your returns before you pay.
• AI Tools for Shipping Documents: Commercial Invoices and Packing Lists — A small exporter's comparison of AI tools for invoices and packing lists, with prices, a worked order and the checks that stay human.
• AI Tools for Warehousing: Which Suit a Small Warehouse or 3PL? — Stock counts, pick errors and slotting: when a spreadsheet and a chat assistant is enough and when a warehouse system's AI add-on pays.
• ChatGPT Ads Cost for a Small Store: CPC, CPM and a Test Budget — What ChatGPT ads cost a small store in clicks and impressions, plus a sixteen-day test budget with written stop rules.
• AI Tools for Ecommerce Websites: What a 50-Orders-a-Month Store Needs — A need-versus-skip checklist for a store shipping about 50 orders a month, with free tiers, break-even sums and the volume at which each skipped tool pays.
• AI Tools for Operations Management: Weekly Exceptions Report From Exports — A worked example with three exports, a review prompt, a sample exceptions list and the checks that catch a wrong join before it misleads you.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Shopify help pages on Sidekick, purchase orders, returns and return processing, the Shopify pricing page, Gorgias and Tidio pricing pages, and the ChatGPT and Claude pricing pages, read in November 2026; the reorder maths and store figures are illustrative.