How Will AI Affect Retail? The Jobs a Small Shop Hands Over First

The first jobs a one-shop owner can hand to AI are the desk jobs done after closing: building the weekly reorder list from till sales, writing product descriptions, drafting replies to routine customer messages, and checking prices against costs. Set up a shop-facts document, a weekly sales export and a named person who approves every draft.

Nothing below depends on a forecast of where retail is heading. Each job works with tools sold today and can be tried for a month at little cost. The order matters more than the tool: start where the data is already yours and a mistake shows up as a number rather than as an upset customer.

Follow me on Instagram@sagnikteaches

Why reorders go first and prices go last

Reorders go first because the mistake stays private. If the sheet suggests eight bags and you know the shop sells six, you change a number and nobody else is involved. Product copy comes second because every claim can be checked against a label. Customer messages come third because errors reach people, so they stay as drafts a person sends. Prices come last because they change revenue the day you apply them, and the assistant sees margins but not what your regulars think of the shop. That order is a judgement, not a law. If your biggest headache is a flood of customer messages rather than empty shelves, start with the week-three job instead, but keep every reply as a draft until the shop-facts document has been tested on a fortnight of real questions.

Connect on LinkedInSagnik Bhattacharya

The set-up list: five things before the first prompt

1. A shop-facts document. One page that says what the assistant may state as fact. A filled-in example is below.
2. Somewhere to keep it. ChatGPT Projects hold up to 5 files on the Free plan and 25 on Go and Plus. Claude's Free plan allows up to five projects, each with its own instructions. Google's Gemini Gems are becoming skills from November 2026. Pick the tool your team will actually open; the document matters more than where it lives.
3. Three exports from the till. Units sold by item by week for at least eight weeks, stock on hand, and the supplier cost list with case sizes. If your till offers no CSV export, ask the vendor now, because every job except product copy depends on it.
4. An approval rule. Each kind of draft has a named approver. Nothing reaches a customer or a supplier unseen, and unapproved text carries the word DRAFT.
5. A mistake log. A sheet with three columns: date, what the assistant got wrong, and the rule you added. After a month it shows which job to trust and which to drop.

SHOP FACTS (version 3)
Opening hours: Mon-Sat 9 to 5.30, Sun 10 to 4. Closed on public holidays.
Returns: unopened items with receipt within 28 days. Opened pet food
cannot be returned unless faulty.
Delivery: local delivery Tuesday and Friday, free over $40, otherwise $4.
Click and collect: ready two hours after ordering; we text when ready.
Voice: friendly and plain. No exclamation marks, no emojis. Sign off
with the first name of the person replying.
Never: give pet health advice (refer to the customer's vet), promise
stock or delivery dates, discuss other customers.

On cost, the free plans are enough for a month's trial, because a weekly routine needs a couple of uploads at most and ChatGPT's Free plan allows three a day. ChatGPT Plus and Claude Pro are each listed at $20 a month if you outgrow them. Give every person who uses the assistant their own login; both providers' terms forbid sharing credentials. A pet shop with an owner and two part-timers, about 1,400 lines and a till that exports CSV files will carry the example through the next four sections. The figures are illustrative.

Subscribe on YouTube@sagnikteaches

In that shop the approval rule fits in a small table, pinned up by the till so nobody has to remember it:

Two details make it work. The person who checks a draft is never the person who prepared it, and every job has a deadline, so a draft cannot sit unread in a chat window until the next order is already late.

Week one: a reorder list for 38 lines

The owner exports eight weeks of sales for her 38 fastest-moving lines and asks the assistant to build a reorder sheet. The rule she gives it is cover: lead time to the next delivery, plus the days to the one after, plus a few days of safety stock, all multiplied by daily sales. For a 12 kg bag of dry dog food that sells 6 a week (about 0.86 a day), an order placed Friday arrives Tuesday, 4 days away. Deliveries are weekly, so the cover runs for 4 + 7 days, and she adds 3 days of safety stock. That is 14 days of sales, or 12 bags. Five are on the shelf and in the stockroom, so the sheet says 7. The supplier sells in cases of 4, so she orders 8.

Her prompt asks for formulas rather than answers, so the sheet keeps working after the chat is closed:

I run a small pet shop. The attached CSV has one row per product:
item, case_size, on_hand, on_order, then units sold in weeks 1-8.
Build a reorder sheet for Google Sheets:
1. Average daily sales from weeks 1-8.
2. Cover target = daily sales x 14 (4 days to the next delivery,
   7 days to the one after, 3 days of safety stock).
3. Needed = cover target - on_hand - on_order, never below zero.
4. Order = Needed rounded up to whole cases.
Write formulas that fill down and explain each in one sentence.
Flag any row where the order is more than four weeks of sales, and
any row with a blank or zero case size. Do not invent missing data.

Here are five rows of the sheet the assistant builds for her, with the cover rule worked through (14 days of sales is twice the weekly figure):

The last three rows are why a person stays in the loop. For the hay, a case of 6 leaves her a unit short of the target and two cases overshoot, but she has the room, so she takes two. For the fish flakes, the sheet needs one tub but the supplier sells 12, which is twelve weeks of sales for this line; the four-week flag in the prompt is what brought it to her attention. She asks the supplier for a part case and, if refused, drops the line when the stock runs out. A sheet can show that the arithmetic and the pack size disagree; deciding what to do about it is shop-keeping.

She checks five lines by hand before trusting the other 33 and finds two problems. A discontinued chew toy still appears in the list because its sales history is in the export, and the cat litter shows a case size of 6 when it ships in 4s. With that error, an order for two cases would have been 8 units where the sheet intended 12, a third short. Both are fixed in the source data, and the lesson goes into the mistake log: audit five lines whenever a sheet is new or a supplier changes packaging. The tutorial on forecasting and reordering for small shops goes further into the method, and the pet shop tutorial on stock, subscriptions and advice covers the pet-specific lines.

The same export can answer a second question, which is where your cash is stuck. Ask for any line with no sales in the last four weeks, and any line whose stock on hand covers more than eight weeks of sales. In this shop that list names a small-animal bedding line (14 bags on hand, selling one a week) and two discontinued toys. Put those on a clearance shelf, and leave them off the next order.

One more trap sits inside the averages. Eight weeks of history suits steady lines and misleads on seasonal ones. Wild-bird seed shows it well: say a 2 kg bag sells 4 a week through late summer and 10 a week once the weather turns cold. Eight summer weeks set the cover target at 8 bags, which lasts about five and a half days at cold-weather sales instead of fourteen, so the shelf is empty before the next delivery. Mark seasonal lines in a column of their own and tell the sheet to use the higher of two figures for them: the same weeks last year, or the last two weeks alone. Promotion weeks need the opposite treatment. Leave them out of the average, or a one-off spike inflates the next two months of orders.

A note for anyone who sold online through Shopify and relied on its Stocky app. Shopify's help page says Stocky is no longer available to install from the App Store. The app stopped working on 31 August 2026, purchase orders now live in the Shopify admin, and Shopify points merchants to its Sidekick assistant for reorder suggestions. The tutorial on Stocky alternatives compares the options. Time for the whole week: 90 minutes to set up and about 30 minutes each week after, against roughly 90 minutes of manual ordering.

Week two: 40 product descriptions from one supplier sheet

A new range of toys and accessories arrives with a supplier spreadsheet: name, material, size, colour, intended dog size and features. The prompt gives the assistant one rule above all, which is to use only the facts in the row.

Write a product description for a pet shop website, 60 to 80 words,
plain and friendly, no exclamation marks. Use ONLY the facts in the row
below. If a fact is not in the row, do not mention it. End with a line
starting "Unknown:" listing anything a customer would normally ask that
the row does not answer, such as washing instructions.
Row: Rope tug toy | cotton rope | 38 cm | blue and white | medium dogs |
three knots | no care information

Without the "only the facts" rule, a typical first draft reads: "A tough 38 cm cotton rope tug toy with three knots, in blue and white. Machine washable and gentle on gums, it is ideal for medium dogs and endless games of fetch." Count the problems: "tough" is not in the row, "machine washable" and "gentle on gums" are invented, and "endless games" is hype. With the rule it becomes: "A 38 cm cotton rope tug toy with three knots, in blue and white, made for medium dogs. Unknown: washing instructions and whether it is suitable for unsupervised chewing." The Unknown line is the useful part. It turns each gap into a question for the supplier instead of a made-up claim.

In the example shop it takes about 4 minutes a line including the check, against 12 minutes writing from scratch. Over 40 lines that saves 320 minutes, or 5 hours 20 minutes, once. If you also list products online, the tutorial on taking a local shop online with product listings covers photos and titles.

Week three: twelve questions customers keep asking

The owner lists the twelve questions that arrive most often by message or at the counter: opening hours, returns, delivery days, click and collect, whether a food is in stock, which size to buy, gift cards, bulk discounts, loyalty offers, special orders, parking, and whether the shop can recommend a food for a health problem. She asks the assistant to draft a reply to each using the shop-facts document, pastes in five past replies she liked, and edits the twelve until they sound like her.

Two rules come out of the first week. The assistant never states stock or delivery dates, because it cannot see them. A message such as "Do you still have the 12 kg salmon one in?" gets a draft that says: "Hi [first name], thanks for asking. I'll check the shelf and the stockroom and message you back by [time]. If it's here, I can put a bag aside for you for 24 hours. [your name]". The bracketed gaps are filled in by a person, and the 24-hour hold is a policy that sits in the shop-facts document. The second rule is that any question about a pet's health gets the same reply: "That's one for your vet, as they know your pet's history. I can show you the ingredient list on any food we stock if that helps." The assistant drafts it, a person sends it. The tutorial on retail chatbots and stock questions explains why live stock answers need a connection to your till.

Policy questions show why the shop-facts document has to be pasted in every time. A customer writes: "My dog won't touch the new food. Can I bring the open bag back?" Run without the document, the first draft read: "Of course! Just pop it back in with your receipt and we'll give you a full refund." That promises a refund the shop does not offer on opened food, and it opens with the exclamation mark the voice rules ban. With the document included, the draft became: "Hi [first name], sorry the new food hasn't gone down well. We can only take back opened food if it's faulty, so if anything seems wrong with the bag, bring it in with the receipt and we'll take a look. If it's simply not to your dog's taste, ask us about a smaller bag before switching next time. [your name]". The owner checked one thing before sending: whether that food comes in a smaller bag at all. It did not, so the last sentence went.

Some messages never get a draft at all. A complaint that an animal fell ill after eating something from the shop, a request for a refund beyond your usual limit, a threat, or a message that names another customer goes straight to the owner. Add those categories to the shop-facts document under "Never", so a member of staff pasting a message in knows when to stop and pick up the phone instead.

About 30 messages a week arrive this way. At roughly 2 minutes saved on each, that is an hour a week, and the drafts are more consistent than replies written in a hurry at 8 p.m.

Week four: nine price changes from sixty-one new costs

A supplier sends a new price list with 61 changed costs. The owner pastes the old and new costs, the shelf prices and her margin targets (28% on food, 55% on accessories) into the assistant and asks which lines now fall below target and what price would restore it, using price endings of .49 or .99 on accessories and .99 on food. It flags 14 lines. She changes 9 and leaves 5 where a rival's price or a loyal customer base makes her hold.

Attached: item, category, old cost, new cost and shelf price for 61
lines. Margin = (shelf price - new cost) / shelf price.
Targets: food 28%, accessories 55%.
1. List every line now below its target, with old and new margin.
2. For each, find the lowest price that reaches the target, then round
   up: accessories to a .49 or .99 ending, food to .99 only.
3. Show the margin at that price and the change in dollars.
4. Leave out every line still at or above its target.
Reminder: food prices end in .99, never .49.

Four of the 14 rows it sent back, illustrative but typical:

Three rows are right. Take the chew toy: its cost rose from $3.80 to $4.30, so the margin at $8.99 fell from 57.7% to 52.2%. Reaching 55% needs at least $9.56 ($4.30 divided by 0.45), and the next permitted ending, $9.99, restores 57.0% for a $1.00 rise. The dog food's cost went from $38.40 to $40.20, so its margin slipped from 30.2% to 26.9%, and $55.99 brings it back to 28.2%. The wet cat food is wrong twice, even with the reminder at the end of the prompt: $14.49 breaks the .99 rule for food, and 26.2% does not reach the target anyway. The lowest compliant price is $14.99 (28.6%), a $1.00 rise on a line regulars buy every week, so she checks the competitor before deciding. Recalculating two rows by hand each week is how slips like this get caught.

She holds the dog food at $54.99 this week because a competitor has an offer on, and the log records why. Holding has a price worth writing down: at 6 bags a week, a dollar less margin a bag is $6 a week, or $18 over the three weeks the rival's offer runs. That is a small cost against the risk of regulars trying the competitor and not coming back, and the note in the log stops a temporary hold from quietly lasting all year. The assistant proposes; the owner decides, using what she knows about the street. The tutorial on prices and promotions for a small shop covers the wider questions.

Jobs I would leave alone for now

• An unsupervised chatbot on the shop's messages. Wrong answers about stock, delivery or pet health land directly on customers. Drafts first; a chatbot only after months of clean drafts.
• Automatic reordering. The sheet suggests, a person places the order. A supplier's pack change or a delisted line is exactly the kind of thing a sheet cannot see.
• Automatic repricing. A weekly proposal with a reason is easier to defend than a price that moved overnight.
• Anything about animal health. Food suitability for a sick pet, medication, symptoms: refer to a vet, every time.
• Customer personal data. Keep names, addresses, phone numbers and loyalty history out of the assistant. Replace them with first names or a code in drafts, and add the details yourself afterwards.

What a month of the mistake log looks like

The log sounds like admin until you see what it collects. These are the entries from the four weeks above, and each one became a rule in the shop-facts document, the prompt or the weekly routine.

Four entries in four weeks, each fixed once, is a healthy pattern. If the same kind of mistake keeps coming back, the job needs tighter rules or should go back to a person. If the log is empty after a month, you probably are not checking closely enough.

Checking that it paid for itself

At the end of the month, line up three numbers. First, minutes spent a week on ordering and replies before and after. In this shop, ordering falls from 90 to 30 minutes, replies fall by 60 minutes, and the new weekly price check costs 20, a net saving of about 100 minutes a week, plus the one-off 5 hours 20 minutes on product copy. Second, empty facings on your top 50 lines at the same time each Saturday morning. Third, the count in the mistake log. If minutes fall, empties fall and the mistake count drops each week as your rules settle, keep the jobs. If a job's mistakes do not fall, drop it or give it tighter rules.

Take the "before" readings on the Saturday ahead of week one, or there is nothing to compare against. The pet shop's scorecard, filled in from the stopwatch, the shelf count and the log:

Read it across rather than down. Ordering time fell after the set-up week and stayed down, so that job stays. Empties fell from 11 to 6 once the sheet started driving orders, which is the result that matters to customers. Replies only began in week three, so one more month is needed before judging them. A steady one mistake a week is normal for a new routine; in month two it should fall, and if it does not, the job with the repeat mistakes needs tighter rules.

The money sum is short. A $20-a-month plan is repaid by one hour of your time a month if that hour is worth $20 to you; a hundred minutes a week is far more than that. What matters is whether the hours go on something that earns, such as serving customers or buying better, and whether you checked the output.

What a one-shop owner asks before starting

Which AI plan should a one-shop owner start with?

Start on a free plan for a month, because the weekly jobs here need few uploads. Move to a paid personal plan, listed at about $20 a month for ChatGPT Plus or Claude Pro, when file limits or message caps get in the way. Business plans need at least two seats, so they suit a shop where a manager also uses the assistant.

Should AI answer customers directly?

Not at first. Drafts that a person sends keep promises about stock, delivery and pet health under your control. If you later add a chatbot, restrict it to opening hours, returns and delivery rules, and test it on 30 real past questions before it goes live.

What should never go into the assistant?

Customer names with addresses, phone numbers, loyalty-card data, card details and anything about a customer's circumstances. Item-level sales, supplier costs and product facts are fine. Give each member of staff their own login, because the providers' terms forbid sharing credentials.

How do I know whether it is saving time?

Time two things for a month: the minutes you spend on ordering and replies each week, and the number of drafts you had to correct. If minutes fall and corrections fall too, keep going. If corrections stay high, tighten the shop-facts document or drop that job.

Further reads

• Examples of AI in Retail: Eight Jobs a Small Shop Can Copy — A longer list of eight jobs once these four are running.
• What AI Costs a Small Retail Shop in Its First Year — What a small shop actually spends in its first year.
• AI Agents in Retail: What an Independent Shop Can Actually Run — What an independent shop can run on AI agents today.
• AI Software for Retail Stores: Five Checks Before You Buy — Five checks before you buy retail AI software.
• Can a Small Shop Use AI Without Having an Online Store? — How a shop with no online store can still use AI.
• Help ChatGPT Discover Your Products: A Checklist for Small Shops — A checklist for being found when shoppers ask ChatGPT.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's help pages on file uploads and Projects, Anthropic's help page on Claude projects, ChatGPT and Claude plan pages, Shopify's help page on Stocky, checked 29 October 2026. The shop, its figures and the sample outputs are illustrative.