An AI agent can already look up stock and sales in your till or Shopify data, flag lines about to run out, draft reorder emails to suppliers and reply to routine customer questions, with you approving anything that spends money. Counting stock, handling people at the counter, negotiating terms and deciding refunds still need a member of staff.
Most coverage of retail agents describes chains with data teams and thousands of stores. A shop with one till has a different problem and one advantage. The advantage is a small, knowable set of products, suppliers and customers, so someone who knows the shelves can check the agent's work. The limit is the stock file: an agent can only be as right as your counts, so fixing the counts comes before buying anything.
What an agent adds when the system is one till and a Shopify store
A report tells you what happened. An assistant answers a question about it. An agent goes one step further: it reads your sales and stock, works out what needs doing, prepares the action (a purchase order, a reply, a price change) and waits for you to approve it. The difference between an agent, a chatbot and a plain automation matters less to a shop than one practical test: where does the approval step sit, and what can the agent reach before it gets there?
Two built-in examples show the range. Shopify's Sidekick comes with Shopify plans and answers questions about your store's own data. Stocky, Shopify's old inventory app, stopped working on 31 August 2026. Purchase orders now live in Shopify admin, and Shopify's guidance points merchants to Sidekick for the reorder suggestions Stocky used to give. Two limits show up in daily use. Purchase orders can't be emailed to suppliers directly from Shopify, so a person still sends the order. And Stocky's history, meaning old orders, received quantities and supplier records, can't be imported. Sidekick's custom app generation is limited to the Grow, Advanced and Plus plans. The day-to-day side is in how to use Shopify Sidekick to run your store faster.
Lightspeed AI, the assistant inside Lightspeed Retail, started as a way to ask plain-English questions such as which items sold best last weekend. In March 2026 Lightspeed announced a beta tool that reads supplier invoices and drafts the purchase orders, to cut manual stock entry when deliveries arrive. Its October 2026 announcement goes further: agentic features that prepare actions such as draft purchase orders and stock transfers between locations for you to review and approve, offered first to selected customers through an opt-in beta. Transfers matter only if you run more than one shop, and the announcements don't say which plan includes which feature, so confirm what yours includes before you plan around it. Lightspeed Retail's list price starts at $89 a month on annual billing; check the pricing page, because it changes.
No built-in assistant? A general one still does the drafting half. ChatGPT, Claude and Gemini all accept spreadsheet uploads, so you can export the sales and stock report from your till and ask the same questions. You lose live data and gain independence from the till vendor. An individual plan for any of them costs about $20 a month at list price.
Nine shop jobs: what the agent does and what staff do
| Job | The agent can | A person still has to |
|---|---|---|
| Stock look-ups ("is the olive mug in?") | Answer from stock data with an "as of" time | Check the shelf when the count is old or the item is reserved |
| Low-stock flags | List lines below target cover each morning | Decide whether a line deserves its place in the range |
| Reorder drafts | Work out quantities and draft one order per supplier | Adjust for deals and pack sizes, approve and send |
| Supplier price and lead-time checks | Compare each invoice price with the last one and flag rises | Phone the rep and negotiate |
| Customer messages | Answer hours, delivery cut-offs and click-and-collect status; collect order numbers and photos | Refunds, damaged goods, complaints and price matches |
| Product listings | Draft descriptions from supplier data | Check specifications, safety wording and photos |
| Slow-seller markdowns | List lines with no sale in 90 days and suggest a markdown | Choose the price and where the stock goes in the shop |
| End-of-day summary | Post takings, top sellers and exceptions | Read it and ask why |
| Stocktake | Prepare count sheets and variance reports | Physically count and recount big differences |
The pattern holds in every row. The agent works on data and text; the person works on shelves, people and money. When something goes wrong it is almost always in the right-hand column, where a job that was meant to be "review the draft" quietly turned into "press go".
Five questions that separate a working agent from a renamed report
Retail vendors have been relabelling dashboards as agents, so put these to any product before you pay for it.
- Which actions can it prepare, and which can it take alone? A drafted purchase order is a different thing from an order that has been sent.
- Where is the approval step, and can I make it mandatory for anything involving money? If the answer is "in the settings somewhere", find the setting before the demo ends.
- Can I see the numbers behind each suggestion? Weeks of cover, the sales window used and the pack size are the minimum. An agent that can't show its working can't be corrected.
- What does it read? Live till data, a nightly export, or only what you paste in? The answer decides how stale its stock figures can be.
- Can I switch off one feature without losing the rest? A shop often wants look-ups on and customer-facing replies off for the first month.
Demos usually answer the first question with a drafted order on screen. Ask to see what happens between that draft and the supplier's inbox, because that gap is where a shop's own judgement has to sit.
Stock look-ups: quick answers from numbers that may be stale
Ask a stock assistant whether you have the olive green enamel mug and it answers in a second. The question to put to yourself is how old that number is. This is the shape of answer worth insisting on (illustrative):
Olive enamel mug, 350 ml
On hand: 4 Reserved for click-and-collect: 2 Available to sell: 2
Last counted: 9 weeks ago (figure unverified)
Sold in the last 4 weeks: 36That answer is useful because it admits its age. Without the last two lines, a clerk would tell a customer "yes, we have four" and be wrong. Put a standing instruction in place: always show the last count date, and if it's older than 60 days, call the figure unverified.
Then test the data before you trust anything built on it. Choose 20 lines at random, count them, and compare. If fewer than about 18 match exactly, reorder drafts will wobble, because the agent trusts your figures completely. That 90% bar is my own rule of thumb, not a vendor threshold. If the test fails, the fastest fix is a tighter counting routine, and running a stocktake faster with a phone scanner covers how to do it without closing the shop.
A Monday reorder run, line by line
Take a four-person homeware shop with about 1,350 active lines and 38 suppliers, nine of which it orders from every week. The Monday run used to take 90 minutes of scrolling reports and writing emails. The rule given to the agent is simple: hold three weeks of cover (one week of supplier lead time, one week until the next order, one week of safety). The suggested quantity is weekly sales times three, minus what's on hand and on order, rounded up to the supplier's pack size. Here is the ceramics supplier's draft and what the owner did with it (illustrative figures).
| Line | Sold in 4 weeks | On hand | Agent suggests | What the owner did |
|---|---|---|---|---|
| Olive enamel mug | 36 | 4 | 24 (four packs of 6) | Accepted |
| White milk jug | 8 | 5 | 4 (one pack) | Skipped: 5 on hand is 2.5 weeks of cover |
| Stoneware side plate | 20 | -3 | 18 | Held. A shelf count found 9 returned plates never booked back, so 6 were on hand; ordered 12 |
| Candle gift set | 24 | 7 | 12 | Skipped: 20 of the 24 sold in a two-week promotion that has ended |
| 1.2 litre teapot | 12 | 2 | 8 | Skipped: the supplier is discontinuing the line, so mark down the rest |
| Linen napkins, set of 4 | 14 | 3 | 0 | Nothing to do: the agent counted 12 already on order, arriving Thursday |
The agent suggested five lines worth $421.80 at cost (mugs $4.80, jugs $6.20, plates $3.10, candle sets $9.50 and teapots $14). The owner ordered two of them, worth $152.40. A first run is meant to look like this. Nearly every correction came from something an agent can't see unless you tell it: a negative stock figure, a promotion that distorted four weeks of sales, and a supplier's plan. Each became a standing rule for the next run. Block lines with negative stock until they are counted. Flag any line where more than half of four weeks' sales fell in two weeks. Mark a product as discontinued the day a supplier says so.
The email the agent drafts for that supplier shows the same discipline. This is what the owner sent after the review (illustrative, with placeholders where a real one carries names and codes).
To: [supplier orders address]
Subject: Order for delivery on [day] - account [number]
Hi [rep first name],
Please can we order the following for delivery on [day]:
- MUG-OLV olive enamel mug 350 ml 24 (4 packs of 6)
- PLATE-SIDE stoneware side plate 12 (2 packs of 6)
Please confirm stock, price and the delivery day.
Thanks, [your name]Three checks happen before it goes. The codes and pack sizes are compared with the supplier's own price list, because agents sometimes mix your internal code with theirs. The email carries no prices, and that is deliberate: an agent that writes a price from memory is guessing, and a supplier who confirms it back gives you a proper check against the last invoice. And the delivery day is tested against the supplier's cut-off, which lives in your head rather than the data. If a draft ever says "as agreed last week", delete the phrase. The agent has no record of any agreement.
The time sum is the other half. Generating drafts for nine suppliers took six minutes and the review took 35, against 90 by hand. That saves 49 minutes a week, about 3.5 hours a month, or roughly $88 at $25 an hour. So the reorder job alone can justify paying about $88 a month for software; anything above that needs another benefit, such as fewer stock-outs. For the arithmetic behind the cover rule, setting reorder points that prevent stockouts goes further.
Customer messages: what it answers and what goes to a person
Shopify Inbox now includes a free AI agent that answers customers on its own, on the Basic plan and above for new customer accounts. It can use web search as a secondary source, which is the reason to test it on your own policies before a customer does. Ask it ten questions whose answers sit on your returns and delivery pages, and three it shouldn't be able to answer. Every wrong or invented reply tells you what to write down. The wider rules for a shop chatbot are in whether an AI chatbot can answer stock questions.
Sort incoming messages into three kinds. Answer outright: opening hours, delivery cut-offs, click-and-collect timings, gift wrapping, care instructions. Collect, then hand over: an order number, a photo of the damage, the size or colour wanted, plus a promise about when a person will reply. Hand over immediately: refunds, damaged goods, complaints about staff, price matches and trade or bulk enquiries. Here is the same message answered two ways.
Customer: Hi, the lamp I bought last week has a cracked shade and I'd
like my money back.
Over-eager draft: So sorry! I've refunded your order, you'll see it in
3-5 days.
Safe draft: I'm sorry the shade arrived cracked. Could you send your order
number and a photo of the damage? [owner first name] handles refunds and
will reply by the end of tomorrow.The first draft sounds kinder and is the one that costs you money, because the agent has no authority to refund and no way to know the shade was cracked before it left the shop. The second collects what a person needs and makes a promise you can keep, as long as someone really does reply by the end of tomorrow.
Bike shop, deli or boutique: where the first agent job differs
The homeware example suits a shop with thousands of small, steady lines. Other shops start elsewhere.
A bike shop earns most of its margin on parts, accessories and special orders, and its stock problem is the order a customer is waiting for. The first agent job is a daily list of special orders that have been on order for more than seven days, with a drafted chaser to the distributor for each. Staff keep everything on the workshop side: diagnosis, warranty decisions and any advice about brakes or other safety-critical parts.
A deli or food shop deals in short dates. Its first job is a morning "use first" list built from batch and expiry dates, with markdown suggestions for lines due to go out of date in two days. Staff keep allergen questions, quality judgement and the supplier relationships that decide who delivers on a Saturday. A chatbot answering allergen questions is a different risk altogether, and the jobs a deli or farm shop can hand over first draws that line.
A clothing boutique lives on size runs. The agent's first job is spotting a broken run (the 10 and 14 sold, the 12 left) and drafting a top-up order only where the style is still in season. Staff keep styling and fit advice, decisions on worn or marked returns, and the buying itself, since next season's range is a judgement about taste that no sales file contains.
Limits a shop meets within the first month
- The stock file. The agent can't tell that a figure is wrong. Negative stock, unbooked returns and goods sitting in the back room all look like facts.
- The shelf itself. Display pieces, damaged stock and items set aside for a customer are invisible unless your system records them.
- Events outside the data. A school holiday, a market day on your street or a price rise a rep announced by phone aren't in the sales history.
- Supplier relationships. The agent reorders at list price from a rep who would give you 5% for a bigger order. Negotiation, bundle deals and cut-off times for next-day delivery stay with the owner.
- Customer-facing authority. Refunds, goodwill gestures and exceptions need someone who is accountable to the customer.
- Customer data. Names and order histories shouldn't go into a consumer chat tool. Use the till vendor's own assistant, or a business plan whose terms say your content isn't used for training by default.
Staff hand-offs: a one-page sheet that names an owner for every output
An agent's output with no named reader is a report nobody opens. Fill in one line per output, and put the sheet where the shift lead can see it. This is a filled-in version for the homeware shop above.
| Agent output | Who acts | By when | If it's missed |
|---|---|---|---|
| Monday reorder drafts | Owner | Monday 10:00 | Orders go on Tuesday; lines under one week of cover are flagged again |
| Customer messages marked "needs a person" | Shift lead | Within 2 hours of opening | Escalate to the owner |
| Negative-stock list | Stock clerk | Thursday close | Reorder block on those lines stays on |
| Slow-seller list | Owner | Friday | Skip a week and repeat the following Friday |
| Supplier price-rise flags | Owner | Before the invoice is paid | Query the invoice first |
Four weeks to a first agent job on a single till
- Week 1, fix the data. Run the 20-line count test. Put supplier pack sizes and lead times into the product records. Clear negative stock.
- Week 2, look-ups only. Let staff ask the assistant stock questions and compare each answer with the shelf. Log every wrong one with the cause.
- Week 3, reorder drafts for one supplier. Review every line, record what you changed and why, and turn each repeated correction into a rule.
- Week 4, customer message drafts on one channel. Test the policy questions first, then let it draft replies you approve before they go.
The stop rule: if more than one reorder line in four needs correcting in week 3, stay on that supplier for another week instead of adding more. The cost side is modest. Sidekick is included with Shopify plans (Basic is $39 a month, or $29 on annual billing), Lightspeed AI sits within Lightspeed Retail, and a general assistant is about $20 a month. Set those against the roughly $88 a month the reorder job saves in time, and the sums work for a shop that does the data work first and fail for one that skips it.
What shop owners ask before starting
Do I need Shopify or Lightspeed to use an AI agent in my shop?
No. A built-in assistant has live access to your stock and sales, which is its advantage. Without one, export a sales and stock report from your till as a CSV or spreadsheet, upload it to a general assistant such as ChatGPT, Claude or Gemini, and ask the same questions. You lose live data but keep the drafting.
Can an AI agent send supplier orders by itself?
It can, but I wouldn't let it for the first two months. An order commits money, and the first runs always need corrections for pack sizes, promotions and discontinued lines. Shopify can't email purchase orders to suppliers directly anyway, so a person still sends them. Automate sending only after weeks of drafts needing few edits.
How accurate does my stock count need to be first?
Pick 20 lines at random and count them. If fewer than about 18 match the system exactly, reorder suggestions will wobble, because the agent trusts your figures. That is my rule of thumb, not a vendor threshold. Fix the counts first, starting with negative stock and lines that never move.
What should a shop never let an agent do alone?
Refunds, anything that spends money, price changes customers will see, adjustments to the stock record, and sharing customer details with outside tools. The agent can prepare each of these and explain its reasoning, but a named member of staff makes the final call.
Further reads
- Best AI Inventory Tools for Small Retailers Compared — Inventory tools compared for small shops.
- AI Software for Retail Stores: Five Checks Before You Buy — Five checks to run before buying shop AI.
- Examples of AI in Retail: Eight Jobs a Small Shop Can Copy — Eight more small-shop jobs to copy.
- How to Find Dead Stock and Slow Movers With AI — Use the agent's slow-seller list properly.
- Why AI Stock Forecasts Go Wrong: 8 Mistakes Small Businesses Make — Where stock forecasts go wrong.
- Retail Staff Scheduling: Build a Rota From Hourly Till Sales With AI — Build the rota from till data.
- How to Prepare Your Small Business for AI Agents — Get the business ready before an agent starts.
- How Small Shops Can Use AI to Forecast Stock and Reorder — A lighter-weight way to forecast and reorder.
- AI Customer Service Kiosk for a Shop Counter: Worth the Cost? — A payback test for a counter kiosk, what it can and cannot answer, and the cheaper QR code chatbot most small shops should try first.
- Retail AI Consultant: Hire-or-Not Checklist for a Small Shop — A hire-or-not test for small shops: signs you need help, what to put in scope, fixed fee versus day rate, and what to ask a retail AI consultant.
- AI for Convenience Stores: A Weekly Routine for Reorders and Rotas — One weekly run-sheet for a convenience store: a reorder list from the till export, shelf prices against two rivals and a rota draft, each step approved by you.
- How Will AI Affect Retail? The Jobs a Small Shop Hands Over First — A four-week walk-through of the first AI jobs in a small shop, with the set-up list, prompts, sample outputs and the mistakes to catch.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Shopify help and migration pages for Stocky, Sidekick and Inbox; Lightspeed press releases (March and October 2026) and Lightspeed Retail pricing page; vendor documentation for ChatGPT, Claude and Gemini file uploads.