The five jobs an existing online store should hand to AI first are support replies (as drafts), product copy, review replies, ad text and reorder alerts. Rank them by weekly hours spent and cost of an error, then start with the one that has high hours and a human check before anything reaches a customer.
The order matters more than the tool. Owners who begin with the exciting job, usually ads, often begin with the one where a mistake costs real money within the hour. The scoring below puts your own hours into a simple table, so the first job is chosen by your store's numbers rather than by what a vendor demonstrates. Everything here is tool-neutral: ChatGPT, Claude, Gemini or the assistant built into your store platform can all do these tasks.
The store we will use: an online shop selling HVAC filters and parts
Imagine a three-person online shop that sells air filters, thermostats and spare parts for heating and cooling systems. It lists about 400 products, ships roughly 70 orders a day on a busy day, and has two part-time people on customer messages. Everything below is illustrative. Its owner timed one ordinary week and wrote down where the hours went:
| Job | Hours a week | Share AI could draft | Hours saved | Error risk (1 low, 5 high) |
|---|---|---|---|---|
| Support replies | 9 | 50% | 4.5 | 4 |
| Product copy | 4 | 60% | 2.4 | 3 |
| Review replies | 2 | 70% | 1.4 | 2 |
| Ad text and variants | 5 | 40% | 2.0 | 3 |
| Reorders and supplier emails | 3 | 40% | 1.2 | 4 |
The "share AI could draft" column is a guess, so write it as a range if you prefer. The "error risk" column answers one question: if this goes wrong and nobody notices for a week, how bad is it? A wrong fitment claim on a filter means returns and refunds. A slightly stiff review reply means nothing. Score honestly and score before you start.
A simple score that picks the order
Divide hours saved by the risk score. That gives a rough "value per unit of worry":
- Support replies: 4.5 / 4 = 1.13
- Product copy: 2.4 / 3 = 0.80
- Review replies: 1.4 / 2 = 0.70
- Ad text: 2.0 / 3 = 0.67
- Reorders: 1.2 / 4 = 0.30
The ranking: support, copy, reviews, ads, reorders. The score is a prompt for a conversation, not a law. Reorders rank last partly because the AI saves little there: stock logic belongs in a spreadsheet, as forecasting stock and reordering with AI explains, and the AI's role is drafting the supplier email. The ordering also shifts with your own facts: a store with 2,000 products would put product copy first.
One rule overrides the score. Anything that sends money, changes a price or speaks to a customer without review waits until the same job has run for a month with a human approving every output. The first month of every job below is "AI drafts, person approves".
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowBefore the first job: put the store's facts in one place
Every job goes better with a short briefing document the AI can read each time. Write it once, about 500 words, and keep it as the instruction page of a project in your assistant of choice (ChatGPT Projects, Claude Projects and Gemini all offer a way to attach standing instructions and files; custom GPTs are being retired in December 2026, so avoid building on them). It should hold:
- who you are and who you sell to, in three sentences;
- tone rules, with two or three real sentences from your own writing;
- the policies the AI may quote: returns window, shipping times, warranty, with the exact wording;
- things it must never do: promise delivery dates, quote prices not in the sheet, claim a part fits without the compatibility list.
An hour spent on this page saves more than any prompt trick, and the same page serves all five jobs. Keep any customer personal data out of it.
Job 1: support replies, drafted and approved
This is the job with the largest saving and the most risk, so it is set up as drafts only. Take twenty old emails you have already answered and test before you let it near a live inbox.
You draft replies for a small online shop selling heating and cooling parts. Use the shop notes attached for policies and tone.
For each customer email below: (1) say in one line what the customer wants, (2) draft a reply under 100 words, (3) list anything I must check before sending, such as an order number or a compatibility claim.
If the answer depends on information not in the notes, do not guess: write "NEEDS ANSWER: " and the question I should answer.
Email: {{customer email}}Illustrative output on a customer who asks whether a 16x25x1 filter fits a unit marked 16x25x4: the draft says the thicknesses differ and asks for the model number to confirm, with a check-list line: "Do not confirm fit; thickness mismatch." The "NEEDS ANSWER" instruction is the one that earns its place: it turns a confident-sounding guess into a question for you.
Measure the first month with two numbers: minutes per reply (time it for ten replies before and ten after), and the share of drafts sent without edits. A typical shift, in this example, is from about 6 minutes to about 2.5, with half of drafts sent as written. For a longer plan, the first 30 days of AI customer support for an online shop sets out week by week what to switch on.
The failure to look for
Drafts that sound right and are wrong. The classic is a confident statement about a warranty or return window the model has half-remembered from training instead of reading from your notes. Spot it by searching each draft for numbers (days, dollars, percentages) and confirming each against the policy text. Anything numeric that is not in the notes is a fault in the prompt, and the fix is a line forbidding it.
Job 2: product copy from a supplier sheet
Copy is the best job for a store with many products and a single supplier spreadsheet. The AI's work is to turn rows of specifications into readable listings. It is not to decide what the specifications are.
Write a product listing from the data row below. Format: a title under 70 characters, three short bullet points, and a 60-word description.
Use ONLY facts in the row. Do not add ratings, certifications, compatibility or performance claims that are not in the row. If a useful fact is missing (for example the airflow rating), list it under MISSING at the end.
Row: Pleated air filter | 16 x 25 x 1 in | MERV 11 | pack of 6 | synthetic media | 3-month change interval stated by supplierIllustrative result: a clean title, three bullets (size, rating, pack size), a description that mentions the supplier's three-month interval as "the supplier suggests", and "MISSING: frame material, airflow resistance". The check is simple: read every bullet against the row. The usual slip is a "helps with allergies" line, which the row did not say. Delete it.
Batch ten products first and read them as a customer would. If the voice is flat, add two examples of your best listings to the project notes and run the ten again. When the style is right, process the rest in batches of 25, spot-checking one in five. A walkthrough for a specific platform is in writing Shopify product descriptions with AI.
Job 3: replies to product reviews
Reviews take few hours and carry little risk, so this is a good place to learn the pattern. Give the AI the review and the star rating, ask for a short public reply in your voice, and set rules by rating. For four and five stars, thank them for something specific. For one to three, apologise for the specific problem, say what you will do, and move the conversation to email. Never ask a reviewer to change a rating.
Reply to this product review as the shop owner, in under 60 words. 5 stars: thank them for one specific thing they mentioned. 1-3 stars: apologise for the specific issue, offer to fix it by email, no excuses, no mention of refunds in public. Do not copy phrases from the review back at the customer.
Review (2 stars): "Filter was the right size but arrived with a bent frame. Fitted anyway but not happy."Illustrative reply: "Thank you for telling us about the bent frame. That should not happen, and I am sorry. Please email us your order number and a photo and we will sort out a replacement. [your name]". Read for two errors: an offer you did not mean to make, and any claim about what caused the damage. Review replies are also one of the safest jobs to let a junior team member approve.
A before and after from the filter shop, in illustrative figures: replying to 30 reviews took the owner about an hour and a half a month, mostly because she kept putting it off. With drafts, it took 25 minutes, and the share of reviews answered rose from about a third to nearly all. The saving in hours is small; the gain is in consistency.
Job 4: ad text and variants
Ad copy sits fourth because the cost of a mistake is paid in spend and in claims made in public. Use the AI to produce variants, not to run campaigns. Give it the product row, the audience in one line and the ad format limits (character counts for headlines and descriptions on the platform you use: Google Ads, for example, allows 30 characters per responsive search ad headline and 90 per description; check the current limits on the platform's help page, as they change).
Write 8 headline variants (max 30 characters each) and 4 descriptions (max 90 characters) for the product below. Audience: homeowners replacing a furnace filter. Use only facts from the row. No claims about health, savings or "best". Mix angles: size match, pack value, delivery speed only if I confirm it below.
Row: {{product row}}
Delivery claim allowed: noThen count the characters yourself, because models miscount them. Paste the variants into a spreadsheet with a LEN formula, and discard any over the limit. Run the variants against one another in the platform's own test feature for two weeks; the AI is a supply of options and the data decides. Ad platforms also have rules about what claims can be made, and about health and savings claims in particular, so read the policy page for the platform you use.
The money rule: no budget, bid or targeting change is made by an AI without your approval. If you use the platform's own automated tools, treat their suggestions as proposals and keep a weekly cap in your own spreadsheet. A store that spends $600 a month on ads can lose an unexpectedly large part of that in a week through one over-eager automation.
Job 5: reorders, as alerts and drafts
The fifth job pairs a spreadsheet rule with an AI-drafted email. The rule: for each item, reorder when stock on hand falls to weekly sales times the supplier's lead time in weeks, plus a buffer. The AI drafts the supplier email, summarises a week's low-stock list in plain words, and explains an unusual jump in sales. It does not choose quantities.
A short worked check for a filter that sells 40 a week, with a 3-week lead time and a one-week buffer: reorder point 40 x 3 + 40 = 160. If 150 are on hand, the alert fires and the order covers the next lead time plus a margin, perhaps 200. A person approves the quantity. The AI's draft reads: "Hi [first name], please can we reorder 200 packs of the 16x25x1 MERV 11 filters, same spec as last order. Can you confirm stock and the ship date? Thanks, [your name]". See how to automate orders, support and restocking for the connected build.
How to time your own week before you score anything
The table above only helps if its hours are real. Most owners guess, and guesses lean towards the jobs they dislike. Spend one ordinary week with a note on the desk or a running note in your phone, and every time you finish a block of work in these five areas, write the minutes and the job. Count a block only if it is over five minutes. At the end of the week, add up each job.
Two things usually surprise people. First, support takes more than expected because of the small gaps between emails: finding the order, opening the courier page, copying a tracking number. Those gaps are exactly where a draft that already contains the order details saves time. Second, product copy takes less than expected in a settled store, since most listings were written long ago, but a new supplier range can turn it into forty hours in a month. Score it on a normal week and again on a launch week, and use whichever matters more to you.
Ask the assistant to tidy the notes into a table if you wish, but read the totals yourself. A run of fifty small time entries is the sort of list where a model will happily drop two lines without telling you.
Who approves what: a review rule for each job
Writing down who approves what, before the first draft arrives, stops a good habit from fading in week three. A rule that works for a small team:
| Job | First month | After a clean month |
|---|---|---|
| Support replies | Owner or lead reads every draft | Order-status replies sent after a spot check of one in five |
| Product copy | Owner reads every listing against the supplier row | One in four checked, all new ranges fully checked |
| Review replies | Any team member approves | Same, with the owner reading all one to three star replies |
| Ad text | Owner approves every variant and every spend change | Same: spend decisions never move off the owner |
| Reorders | Owner approves every quantity | Same: money leaving the business stays with a person |
Notice the right-hand column barely changes for the money jobs. That is deliberate. The time AI saves comes from drafting, and drafting does not need to be unsupervised to be worth having. Put the table where the team can see it, and change it only by a conversation, not by drift.
Five jobs switched on in one weekend, and the fortnight it cost
An illustrative case: a store owner rolled all five jobs out in one weekend. Support drafts, product copy, review replies, ad variants and reorder emails started on the Monday. By Friday he had 140 drafts to read, two incorrect fitment claims in new listings, and a headline on an ad that promised next-day delivery the shop could not offer. He turned off everything and lost a fortnight rebuilding trust in the process.
The better order is one job per week. Week one: support drafts only, reading everything. Week two: add product copy for ten products. Week three: review replies. Week four: ad variants. Week five: reorder drafts. At each step, stop and count the errors you found, because the aim is to learn each job's failure before you add another. The mistakes in week one are cheap; in week five they are cheap too, but the same mistakes made all at once are not.
The five jobs in other stores
The same five fit different shops, but the scores change. A bicycle repair shop with a small web store spends most of its time on support, since customers ask about compatibility and booking, and very little on ads. Its first job is support drafts built on a compatibility chart, and the second is reviews. A bakery selling gift boxes online has seasonal peaks, so ad text before a holiday and a reorder alert for packaging matter more than product copy, which changes twice a year. A solar installer with an online shop for small kits spends little in any category and should hand over one job only, probably product copy, then stop.
If the total weekly hours across all five come to under five, a spreadsheet and a good template may beat any AI set-up. Counting honestly is a favour to yourself.
How to tell each job is paying for itself
For each job, record three numbers on day one and again after four weeks: hours spent, errors reaching a customer, and cost of the tools involved. Write them down in one small table, because memory flatters. If a job saves under an hour a week after a month, switch it off and spend the effort elsewhere. If errors reach customers, tighten the review rule, not the prompt, since a person reading is what catches them.
On tool costs, a single paid assistant plan is usually around $20 a month, and many stores need nothing beyond that for these five jobs. Extra cost arrives with automation platforms or a chat widget; what it costs to add AI to a Shopify store breaks that down. Saved hours turn into money only if you spend them on something that earns, a point made in why AI saves time but not money.
If you would prefer to map the five with someone who has seen how they go wrong, that is what my AI implementation consultation covers, and you can book through the contact page. Whichever way you start, keep the first month small, keep the human check, and write down what you find.
Questions from store owners about the first five jobs
Which of the five should I do first?
Usually support replies as drafts, because they take the most hours and the drafts are reviewed before sending. If your support load is light, start with product copy or review replies, which carry less risk. The scoring table in this tutorial gives you an order based on your own hours.
Do I need a paid plan to do any of this?
Not to begin. A free assistant is enough for product copy, review replies and testing the support prompts on a few old emails. Move to a paid or business plan when you start pasting customer details, because business plans handle data differently from free consumer ones.
How do I stop AI inventing product details?
Give it a supplier sheet to work from and tell it to use only that text, writing 'not stated' where a detail is missing. Then read the output against the sheet for specifications, sizes, materials and compatibility, which is where invented facts do the most damage in returns.
Is it worth it for a store with only a few orders a day?
Yes, if the hours add up. Count the real hours first. Two to three hours a week saved on support and copy is worth setting up, while a task that takes ten minutes a week is not worth any set-up at all. Decide using hours, not enthusiasm.
Further reads
- How to Answer Etsy Customer Messages Faster With AI — Answer marketplace buyer messages faster.
- Best AI Chatbots for Small Shopify Stores (2026) — If you want a chat widget as well as drafts.
- How to Get Your Products Recommended in ChatGPT Shopping — Get products found when shoppers ask AI.
- How Small Amazon Sellers Use AI for Listings and Reviews — Listings and reviews on a marketplace store.
- How to Set Up AI Abandoned Cart Emails for a Small Online Store — A sixth job once these five are steady.
- How to Get Started With AI in Your Small Business: First 7 Steps — The wider starting steps for any small business.
- What Should a Small Business Automate First With AI? — If your list of candidate jobs is longer than five.
- Customer-Facing or Back-Office: Where Should AI Go First? — The same trade-off seen from the risk side.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT pricing page (Projects, training opt-out by plan), OpenAI help centre on the custom GPT retirement, Claude pricing page, Google Ads help on responsive search ad limits, all checked 5 November 2026; general ecommerce practice.