A small online seller should hand AI one job first, and the best candidate is usually answering buyer questions, because every reply stays a draft you approve. Measure that job for a week, run the AI version for 30 days, and keep it only if time per task falls and customer replies stay accurate.
Most sellers who say AI did nothing for them tried five things at once and measured none of them. A single job with a before and an after tells you what worked. The best first job is usually one where you already handle 20 or more real cases a week, because those cases show the assistant what a good answer looks like and give you enough volume to measure a change.
Four jobs to choose between
The jobs that suit a solo seller are the repeatable ones that touch a customer's decision to buy: product listings, buyer questions, abandoned-basket emails and post-purchase emails. Each has a different risk and payoff. Score them honestly against your own week.
| Job | Time it can save | Risk if wrong | Needs first |
|---|---|---|---|
| Product listings | High when you add many items | Medium: a wrong fact becomes a refund | A tidy sheet of product facts |
| Buyer questions | Medium to high | Low if you approve each reply | Your last 30 real questions |
| Abandoned-basket emails | Low (set once) | Low | The store's built-in automation switched on |
| Post-purchase emails | Low to medium | Low | Order confirmation, shipping and review-request wording |
A quick scoring method: give each job a mark from 1 to 5 for hours per week now, a mark from 1 to 5 for how repetitive it is, and subtract a mark from 0 to 3 for how bad a mistake would be. Highest score goes first. A seller who adds two products a month will score listings low even though everyone writes about it. A seller who answers 30 sizing and delivery questions a week will score buyer questions near the top.
The main example: a one-person online bookshop
Take an illustrative independent bookshop that sells secondhand and signed editions through its own online store and ships about 70 orders a month. The owner works alone. A normal week looks like this, measured over seven days with a phone timer:
- Writing new listings: 12 books a week at about 15 minutes each, so 3 hours.
- Answering buyer questions by message and email: about 25 a week at about 4 minutes each, so 100 minutes.
- Abandoned baskets: about 40 a month, handled by a store automation she has never edited.
- Post-purchase emails: default templates, never touched.
On the scoring method, listings get 4 for hours, 3 for repetition and minus 1 for risk, a score of 6. Buyer questions get 3 for hours, 5 for repetition and minus 1 for risk, a score of 7. Questions win by a nose, and there is a second reason: she already has 25 real examples a week, which is exactly what a good prompt needs. Listings come second.
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 nowWeek 0: measure before you change anything
Without a baseline you will not know whether the next 30 days improved anything. Spend one week collecting four numbers for the job you picked:
- Volume: how many questions arrive in the week. The bookshop count was 25.
- Time per task: start a timer when you open a question and stop it when you hit send. Average of 4 minutes.
- Response speed: hours between the question arriving and your reply. Say 7 hours on average.
- Outcome: how many questioners ordered within 48 hours. Say 9 of 25.
Also copy every question into a single document, with your actual reply beside it. That document becomes the raw material for the assistant.
Weeks 1 to 4: the plan for the buyer-questions test
Week 1: build the knowledge sheet and the first prompt
Write one page of facts that answer most questions: postage costs and times, packaging, return rules, how you grade condition (Fine, Very Good, Good, Fair), what "signed" means in your shop, and your opening hours for replies. Add your tone in two or three lines. Store the page in a ChatGPT Project or a Claude Project, a saved workspace that keeps your instructions and files for every chat. Claude lets free accounts create up to five projects, and the paid plans include them.
The instruction block:
You help me answer customer messages for my online bookshop.
Use ONLY the facts in the attached "Shop facts" page.
If a question is not covered, write "NOT IN FACTS" and tell me
what you would need. Never promise a delivery date, a discount
or a refund; say I will confirm. Keep replies under 90 words,
friendly and plain, sign off "[your name]". Start every reply with
the customer's first name if known, otherwise "Hello".
After each reply add one line: CHECK: followed by anything I
should verify in the order before I send.Week 2: draft, edit, send, and log the edits
Paste each incoming question, read the draft, fix it and send. For every reply, note whether you sent it unchanged, lightly edited or rewritten. After a week you will have a rewrite rate, which is the single most useful number in this test.
An illustrative question and draft:
Customer: Hi, is the hardback copy of Middlemarch in good
condition and can it arrive by Friday? It's a birthday gift.
Draft: Hello, thank you for asking. The copy is graded Very Good:
the slipcase has light rubbing at one corner and the pages are
clean and bright. I can post it first class on the day I receive
your order. I cannot promise it will arrive by Friday, but if you
order today I will confirm the posting day by email.
[your name]
CHECK: confirm the grading note and that first class is offered.What you would fix: the shop's grading note must match the actual book, so open the listing. "Cannot promise Friday" is correct because the prompt forbids promises, but you might add, "If it helps, I can send a gift note." Good drafts rarely need more than a sentence added.
Week 3: add listings if the question job is steady
If at least 7 in 10 drafts go out lightly edited or unchanged, begin a second, smaller job while the first one runs: write listing descriptions from your own product sheet. The next section shows how.
Week 4: score the test
Compare the same four numbers with week 0. The bookshop's illustrative results: volume unchanged at 25, time per task down from 4 minutes to 1.5, average response down from 7 hours to 2, and 11 of 25 questioners ordering within 48 hours instead of 9. Time saved: 25 times 2.5 minutes, about 62 minutes a week. Over a month that is just over four hours. Two more orders at an average of $34 is $68, though two orders in a month can be noise, so treat the orders figure as a hint, not a proof.
A listing prompt that works from your own facts
Listings reward a spreadsheet of facts and a rule against invention. For a florist, a photography studio or a bookshop alike, the pattern is the same: one row per product, with columns you fill in once.
Here is one row from my product sheet:
Title: Middlemarch, illustrated hardback, 1999, slipcase
Condition: Very Good (light rubbing to slipcase corner; pages clean)
Size: 24 x 16 cm; weight 1.4 kg
Notes: illustrated by (name on spine), numbered edition not signed
Price: $38
Write: (1) a listing title under 70 characters, (2) a 90-word
description in my voice (warm, specific, no hype), (3) five bullet
points a buyer would scan, (4) three search phrases a buyer might
type. Do not add any fact that is not in the row. If something
important is missing, ask.Before (hand-written in a hurry): "Middlemarch hardback, nice condition." After (edited AI draft): "Middlemarch, illustrated hardback with slipcase, 1999. Very Good: light rubbing to one slipcase corner, pages clean and bright. Illustrated edition, 24 by 16 cm, weighs 1.4 kg, so it ships in a box rather than an envelope." The second is longer, but each sentence answers a buyer's likely question, which in turn cuts the number of questions you have to answer. The product descriptions tutorial goes further on wording, and the Shopify Magic comparison is useful if you sell on that platform.
Abandoned baskets and post-purchase emails use the store's own tools
These two jobs sit in your shop's email settings, so you need AI only for the wording. On Shopify, automatic abandoned checkout emails are set up under Apps, then Messaging, then Automations. A checkout counts as abandoned when it stays incomplete for more than ten minutes after the customer enters an email address. You choose the Send after delay and the Send to group, and you can edit the template. No email goes out if the customer completes the order first, or if they entered a phone number instead of an email. Other stores have equivalents; check yours.
Give the assistant the template and your tone and ask for three versions of a short reminder. An illustrative draft for the bookshop:
Subject: Still thinking about Middlemarch?
Hello [first name],
You left a book in your basket. I only ever have one copy of each
edition, so I am holding yours as long as I can. Finish your order
here and I will post it the same day.
If you have a question about the condition, just reply to this email.
[your name]Check two things before using it. "Holding yours" is a promise, so only keep it if your shop does hold stock; otherwise cut it. And one-off stock is the reason this wording works, so a seller of mass-produced goods needs a different line. For a full set-up, see the abandoned basket emails tutorial. On Shopify, Messaging includes up to 10,000 emails a month free and then bills $1 per 1,000 more, which is far more than a small shop sends, but confirm on your own plan.
Post-purchase emails follow the same pattern: ask for a shipping confirmation that sets expectations, and one review request a week after delivery. Keep each to 80 words and avoid inventing a delivery window; copy the carrier's real range.
A small mistake that cost a refund
An illustrative case from a florist who tried the same approach: the assistant drafted a reply to a customer asking whether tulips ordered on Monday would arrive for a Saturday party. The draft said "yes, definitely". The knowledge sheet said nothing about Saturday cut-off, so the model filled the gap with something that sounded helpful. The florist read the draft, caught the line, replaced it with the true cut-off, and added an instruction: "never say definitely". That one extra instruction fixed the pattern across later drafts.
The lesson is that approval is the safety net for the whole test. If you switch to automatic replies, which many chat tools offer and Shopify Inbox does through its opt-in Inbox agent, do it only after weeks of drafts that rarely needed changing, and keep a human route for anything about money or timing. The tutorial on wrong AI replies covers the repair.
The same test in a florist's week
The method is not tied to books. Picture an illustrative florist with a small online shop for same-week bouquet delivery. Her biggest time sink is not listings, which change with the season, but a Monday pile of questions about delivery slots, vase sizes and card messages: about 40 a week at three minutes each, so two hours. Her knowledge sheet differs from the bookshop's, with delivery cut-offs, how substitutions work when a flower is out of stock, and what she writes on a card, yet the instruction block is nearly identical. She adds two more rules: never name a flower variety that is not on the sheet, and always offer a photo of the finished bouquet before dispatch.
Her week-0 numbers: 40 questions, 3 minutes each, a 10-hour average reply time because she answers in batches on Mondays and Thursdays. After four weeks with drafts: 1 minute each, 3-hour average reply time, because it is now easy to reply between jobs. The two hours become about 40 minutes, a saving of 80 minutes a week. The risk was in substitutions: the assistant once offered "peonies" that were out of season. The fix was one line on the sheet listing what is in stock this month, updated every Monday, and she built that into her ordering routine.
When a question needs the order, not the facts page
Some questions cannot be answered from a shared sheet because they depend on a specific order. "Where is my parcel?" is the usual one. The draft must either ask you to look it up or it will invent a status. In the instruction block, make the rule explicit: for any question about a particular order, reply "NEEDS ORDER LOOKUP" and nothing else. Then handle those by hand, or paste the tracking line into the chat and ask for a short reply.
A filled-in example: the customer writes, "Ordered a week ago, nothing yet." You paste the tracking line "Collected by carrier 3 days ago, in transit, last scan 18 hours ago" and the assistant drafts: "Hello [first name], thank you for waiting. Your parcel was collected by the carrier three days ago and was last scanned 18 hours ago, so it is still moving. If it has not arrived by the end of tomorrow, reply to this email and I will start a trace." You confirm the carrier's usual range before saying "end of tomorrow". If you want to hand order tracking to a chatbot entirely, the order tracking tutorial explains what that takes.
Protect customer details while you test
Buyer questions contain names, addresses and sometimes order numbers. For a one-person test you do not need any of them. Replace names with "[first name]" and leave addresses out; the draft only needs the question. If you paste whole emails, check the privacy setting on your plan that controls whether chats are used to improve the model, and consider a business plan if you will rely on this long term. Business plans of ChatGPT and Claude do not train on business content by default, though each needs at least two seats.
A check-in table for each week of the test
| Week | What you do | What you record |
|---|---|---|
| 0 | Time the job as you do it today | Volume, minutes per task, response time, orders within 48 hours |
| 1 | Write the facts page and the instruction block; run ten messages as a dry run | Questions the page could not answer |
| 2 | Use drafts for every question; mark each as unchanged, edited or rewritten | Rewrite rate |
| 3 | Fix the gaps the rewrites revealed; add the listing job if rewrites are under 3 in 10 | New rules added to the sheet |
| 4 | Repeat the week-0 measurements | Change in each of the four numbers |
Thirty minutes at the end of week 4 is enough to compare and decide. If you are unsure whether a change counts, a shift of a third in time per task over 25 or more messages is a reasonable bar, and anything smaller is probably normal week-to-week variation.
What to do with the second and third jobs
If the first job passes, resist the urge to switch on everything. Add one job a month, in the order your scoring table gave. The listing job is the natural second step, because the same knowledge sheet already holds your grading, sizing or material notes. Abandoned-basket wording is third because it is set once and checked a month later against the recovery figure in your store's reports. A rough sum for the bookshop: with 40 abandoned checkouts a month and, say, 3 recovered at an average of $34, the email earns about $102 a month. If your reminder recovers none in two months, rewrite it or turn it off; a bad reminder is worse than silence.
Keep a single page called "What AI does in our shop" with one line per job, the date it started, the baseline and the latest figure. When a platform changes a feature or a tool is retired, that page tells you what to re-test. It also gives you something concrete to show an adviser if you later decide to spend money on a chatbot or a dedicated tool.
Decide at day 30: keep, change or drop
| Result at day 30 | What to do |
|---|---|
| Time per task down by a third or more; under 3 in 10 drafts rewritten | Keep. Add the second job (listings). |
| Time down but half the drafts rewritten | Fix the knowledge sheet; run two more weeks. |
| Time unchanged | The job was not the bottleneck. Try the next job on your score list. |
| A reply promised something false | Tighten the rules, then re-test; do not automate yet. |
What this costs
The test can run on free plans. A paid chat plan, ChatGPT Plus at $20 a month or Claude Pro at $17 a month billed annually ($20 monthly), is worth it once you rely on saved Projects and uploaded sheets. Business plans for both need two seats, so a solo seller is better on an individual plan. Time is the real cost: about three hours in week 1 to build the facts page, then ten to fifteen minutes a day. If you would rather map the right first job with someone, that is what my AI implementation consultation covers, and you can reach me through the contact page.
Whichever job wins, write the result down. A one-line record such as "Buyer questions, 25 a week, 4 minutes to 1.5, 2 drafts in 25 needed a rewrite" is the evidence that decides your next move. For a wider view of where a small shop can use AI, the small online shop tutorial is a good next read.
Questions from sellers about to start the test
Should I start with listings or buyer questions?
Start with whichever eats the most of your hours each week and carries the least risk if the wording is slightly off. For most solo sellers that is buyer questions, because you can keep every reply as a draft you approve. Listings are the next job once your product facts are in one tidy sheet.
Do I need a paid AI plan for this?
Not to begin. A free plan is enough to test the idea on a handful of messages and listings. A paid plan, such as ChatGPT Plus at $20 a month, becomes worthwhile once you rely on saved instructions and longer uploads. Your store may also include free AI features worth trying first.
Will AI-written listings hurt my search ranking?
Search engines and marketplaces care whether a listing is accurate, specific and useful, not which tool typed it. Generic copy that could describe any product does poorly whoever writes it. Feed the assistant your real product facts and edit the output so it sounds like you.
What if the test fails after 30 days?
A failed test is cheap and useful. Drop the job, note what went wrong, and try the next candidate on your list. Common reasons are weak input facts, too many edits needed per reply, or a job that was never a big time sink in the first place.
Further reads
- Is AI Worth It for a One-Person Bookkeeping Business? — Weighs AI for another one-person business.
- How to Answer Etsy Customer Messages Faster With AI — Answers marketplace messages faster, with checks.
- How Much Does AI Cost a Small Business in 2026? — Full cost picture before you pay for a plan.
- How to Run Your First AI Pilot Project in a Small Business — A general pilot method you can reuse for the next job.
- Your First 30 Days of AI Customer Support for an Online Shop — Week-by-week plan if support becomes your first job.
- How to Write eBay Listings With AI That Sell Faster — Marketplace listing prompts and checks.
- How to Measure Customer Reaction After Introducing AI — A fuller way to track how buyers react to AI replies.
- What to Check in an AI Tool's Privacy Policy and Terms — What to read in a plan's privacy terms before pasting emails.
- Amazon Business Reports: A Weekly Review in ChatGPT or Claude — Download two Business Reports CSVs, run one prompt, check the totals, and decide what to change, in about 30 minutes a week.
- Shopify Profit Report With Claude: Add Ad Spend, COGS and Returns — Join four exports in Claude to see profit per order after ads, costs, shipping and returns, with a full worked month.
- Shopify RFM Analysis: Segment Customers With ChatGPT From an Export — Score customers on recency, frequency and spend from a Shopify export, then write one email per segment, keeping personal data minimal.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Shopify Help Centre (Recovering abandoned checkouts; Assigning your Inbox agent), shopify.com Messaging page, openai.com ChatGPT and Business pricing, claude.com/pricing (all checked 5 November 2026).