Sort the donations first, then photograph only the items nobody can price at a glance: one picture of the item and one close-up of its label or mark. Ask AI what it is and which price band fits, check real sold listings, and write the ticket. Aim for 90 seconds per researched item; price the rest from a band card.
Most donated goods do not need a chatbot at all. A card that says paperbacks are $2 and high-street jumpers are $4 prices them in ten seconds. The AI earns its place on the minority of items where a volunteer would otherwise guess (27 of 180 in the illustrative day below), and its job is to turn a guess into evidence. The danger is that it sounds just as sure when it is wrong about a brand as when it is right, so the workflow below puts a sold-price check you do yourself between the AI's answer and the ticket.
Sort into four piles before anything is photographed
The speed comes from not researching things that do not need it. Four piles, defined by the shop manager, decide how each item is priced.
| Pile | What goes in it | How it is priced | Time target |
|---|---|---|---|
| A: Band card | Paperbacks, mugs, plates, everyday clothing, most toys and DVDs | From the band card, no photo | 10 seconds |
| B: Check | Branded clothing and shoes, kitchen sets, small furniture, anything with a maker's mark nobody recognises | Two photos, the AI prompt, a sold-price check | 90 seconds |
| C: Specialist | Jewellery, watches, art, antiques, designer bags and shoes, signed items, collectables, instruments | Set aside in the specialist box; the manager or a specialist volunteer prices it | 3 minutes to photograph and box |
| D: Not for sale | Anything your safety policy says the shop must not sell | Removed before pricing starts | None |
What belongs in pile D is a matter for your charity's safety policy and the rules where you trade, such as electrical goods, helmets, child car seats and upholstered furniture without labels. It is not an AI's call, and the sorting volunteer should not have to think about it twice.
Here is an illustrative day. Six bags and boxes arrive, containing 180 items. Sorting sends 144 (80%) to pile A, 27 (15%) to pile B and 9 (5%) to pile C. Pile A takes 144 x 10 seconds, which is 24 minutes. Pile B takes 27 x 90 seconds, which is 40.5 minutes. Pile C takes 9 x 3 minutes, which is 27 minutes. The total is 91.5 minutes. If a volunteer had priced all 180 from memory, asking colleagues and searching the odd one on a phone, at an average of 50 seconds an item, the same stock would have taken 150 minutes. The difference is 58.5 minutes, or 39%. Your split will differ, so count one day's donations and time a sample with a phone stopwatch, because the whole case rests on how many items land in pile B.
Pile C matters more than the minutes saved. A watch or a signed print priced at a guess is where a charity shop loses the most, and an AI cannot remove that risk. Its only job there comes afterwards, describing the item for a listing once a person has set the price.
The 90-second target for pile B breaks down like this.
| Step | Seconds | What happens |
|---|---|---|
| Photograph | 15 | One picture of the item, one close-up of the label or mark |
| AI answer | 20 | Upload both photos with the standard prompt and read the reply |
| Sold check | 35 | Search the AI's suggested terms, filter to sold items, read the last few prices |
| Ticket | 20 | Choose the price within the band and write the ticket |
| Total | 90 |
If an item has passed three minutes, stop. Price it at the bottom of the band, mark it for review, and move on. The speed target only works if the slow items are allowed to leave the queue.
Two photos, and what to keep out of the frame
Take the pictures in the ChatGPT app on a shop phone, or upload them from the camera roll. Photo one shows the whole item on a plain surface by a window: clothes flat or on a hanger, not on top of a pile. Photo two fills the frame with the mark that identifies it: the neck label, the care label, a base stamp, a hallmark, a model plate. Hold it steady and keep the text the right way up.
Those habits come from the limits OpenAI lists in its own image guide. It says small text can be hard to read, and that enlarging it helps. It says rotated or upside-down text and images may be misread, that descriptions or captions may sometimes be wrong, that file names and metadata are not processed, and that object counts may be approximate. In practice that means three rules. Fill the frame with the mark. Type what a photo cannot show (the size, "set of six", whether a game is complete), rather than hoping the AI works it out. And never ask it to count a pile of items; count them yourself and say the number.
Keep the following out of every shot: faces, because volunteers and donors should not appear in photos on a chatbot; letters, cards, name labels and photographs left in bags and pockets; and anything medical. Empty pockets before an item is photographed. Use a shop phone rather than a volunteer's own, and delete the photos at the end of the shift. If a bag contains personal papers, pass it to the manager, and keep donor names and addresses out of every prompt. The rules on donor data and ChatGPT cover what else a charity must never paste in.
The standing instructions and the per-item prompt
A ChatGPT Project holds your rules once, so every chat inside it starts with them. OpenAI's help page says project instructions apply only inside the project and override your global custom instructions, and that the Free plan allows five files per project, with Go and Plus allowing 25. You need only two files: the band card and a "never" list of the things the AI must not claim. Put these instructions in the Project.
You help volunteers in a charity shop price donated items. Files in this
project: the band card and the never-list. Rules:
1. Identify the item from the photos. Say what you can read (label text,
marks) and how sure you are: Sure, Probable or Guess.
2. Never say an item is genuine, authentic, rare or valuable. If a brand or
mark suggests value, write SPECIALIST BOX.
3. Never give a sold price from memory. Suggest search terms for a sold-price
check instead.
4. Choose a band from the band card for the condition I state (A, B or C).
5. Finish with a ticket line of 40 characters or fewer.
If a photo is unclear, say what is missing instead of guessing.Each item then needs only a short message with two photos attached.
Item: women's coat. Condition: B. Size 12.
Photo 1: the item. Photo 2: the neck label.
Reply with: IDENTIFIED AS / BASIS / CERTAINTY / BAND / SOLD-CHECK SEARCH /
TICKET.An illustrative reply looks like this.
IDENTIFIED AS: Women's wool-blend coat, size 12, single-breasted
BASIS: Neck label shows a mid-range high-street name; care label reads
60% wool, 40% polyester; buttons and lining look intact in photo 1
CERTAINTY: Probable
BAND: Adult clothing, premium (check): $12 for condition B
SOLD-CHECK SEARCH: "[brand] wool blend coat size 12", sold items only
TICKET: Wool-blend coat, size 12, $12Read it as a volunteer would. The AI has done the slow part, which is turning a label into search terms and a candidate band. It has also made a judgement the volunteer must test. Your band card might list that high-street label under ordinary clothing at $4 for condition B, and the model has leaned towards "premium" because it saw "60% wool". The sold check settles it, and the next section shows how.
The sold-price check: 35 seconds that replace the guess
Take the AI's search terms to a marketplace search and switch the filter to sold items. Only sold listings count. Asking prices show what sellers hope for, and a coat that has sat at $60 for six months is evidence that $60 is too high, not that the coat is worth it.
Take the coat from the previous section. Seven recent sold listings for that label, in a similar size and condition, show $18, $22, $20, $35, $16, $19 and $21. In order they run $16, $18, $19, $20, $21, $22, $35, so the median is $20. The mean is $21.57, pulled upwards by the $35, which was a new coat with tags and does not describe yours. The median is the figure to use. The shop's own rule (illustrative) is that an in-store price is about 60% of the median sold price, so 60% of $20 is $12, which matches the premium price for condition B on the card and confirms the ticket. An online listing would aim for about 90% of the median, which is $18. Had the median come out at $6, 60% would have been $3.60, which is the ordinary band's $4, and the AI's "premium" would have been the error.
Where to find sold data depends on how much you sell online. The sold-items filter in eBay's search results is the quick route. For the shop's regular online lines, eBay's research tools page says its Product Research tool sits under the Research tab in Seller Hub and is free to all sellers, while Sourcing Insights needs a Basic Store subscription or above. Google Lens, which Google describes as a way to look up who makes something, how much it costs and where to buy it, helps with identification. The cost it shows is usually a new price, so use it as a clue about the item, never as the number for a second-hand ticket.
The AI can speed this up only so far. If web search is switched on, ask it for links to sold listings and open them. A figure with no listing behind it is a guess, and a page showing an asking price is not a sold price. The 60% and 90% ratios are settings, not laws: if coats priced at 60% sit for a month, lower the ratio, and if they go in two days the shop is probably underpricing. Set an online threshold too. An illustrative rule is that an item whose sold median reaches $40 goes to the specialist box for the manager to consider listing online, because below that, platform fees, postage and the volunteer's time for the listing eat the gain.
Build the band card from three months of till data
A band card is a one-page table that lets a volunteer price pile A without thinking. Build it from your own sales rather than from a general guess. Export three months of till sales with category, price, date listed and date sold. For each category, work out the median price and the share that sold within 14 days. Two starting rules, which you should adjust: if more than 80% of a category sells within 14 days, move its band up a step, and if fewer than 40% sell within 30 days, move it down. An AI can do the grouping from a pasted table, but check one category by hand in a spreadsheet, because language models miscount.
Condition codes keep it quick: A is like new with no marks, B is good with minor wear, C is worn but usable. An illustrative card might read as follows.
| Category | A | B | C | Note |
|---|---|---|---|---|
| Paperback books | $2.00 | $1.50 | $1.00 | No photo needed |
| Hardback books | $4.00 | $3.00 | $2.00 | Cookery and art books may deserve a check |
| Adult clothing, high street | $6.00 | $4.00 | $2.00 | Price from the label, not the badge |
| Adult clothing, premium | $16.00 | $12.00 | $8.00 | Pile B: sold check |
| Shoes | $8.00 | $5.00 | $3.00 | Brands in pile B |
| Mugs and plates | $2.00 | $1.50 | $1.00 | Marked china in pile B |
| Children's clothing | $3.00 | $2.00 | $1.00 | Sets priced together |
| Toys, complete | $5.00 | $3.00 | $2.00 | Incomplete toys go in the bin or the bargain basket |
Keep the card to one page and fix it where the volunteers work. Review it every quarter against sell-through, and let the AI help with the arithmetic but not with the decisions: whether to raise a band is the manager's call, made from the sales figures.
What AI gets wrong about brands, and how each mistake shows up
Every example below is illustrative, and each is the kind of error that reaches a shelf when nobody checks. The pattern behind them is that an AI is good at describing and guessing and poor at verifying, so a volunteer's job is verification.
- The badge is not the brand. A polo shirt with a small embroidered crest is matched to a famous sportswear logo and sent to the premium band. The crest belongs to a local golf club, and the neck label names a supermarket range. A volunteer catches it by reading the label aloud. The guard is to price from the label in photo two, and when label and badge disagree, the label wins.
- Look-alike patterns. A bag in a repeating check pattern is described as "consistent with" a well-known fashion house. Many unbranded items borrow the same look. "Consistent with" is not "is", and any item that looks like luxury goes to the specialist box with its photos.
- Fakes look real in photographs. A quilted handbag with a gold chain and an interlocking-letter clasp is described as "appearing to be" a designer bag, with a price band for a genuine one. A photo cannot show stitching, the weight of the hardware or a serial number. The guard is the never-list: nothing is called genuine, and the bag goes to someone who can check it.
- Retail price instead of sold price. Asked what a stand mixer is worth, the AI quotes the new price of about $350. A ticket of $175 follows, and the mixer sits for six weeks, because used ones in similar condition have been selling for $90 to $120. The guard is rule three in the Project: no prices from memory, only search terms.
- Misread marks. A china base stamp is read as one factory's mark when it is a similar-looking stamp from another, and a camera lens is matched to the nearest model name because the end of its serial plate was blurred. The two models sell for very different prices. OpenAI lists small text as a known weak point, so retake the close-up until the plate is readable, and send china and metalwork with unclear marks to pile C.
- Confident "vintage" and "rare". A 1990s paperback is called a collectable first edition because its page says "first published". Establishing a first edition needs the specific printing checked, and the shop has already written "collectable" on the ticket when a customer questions it. The never-list bans "rare", "collectable" and "genuine" for exactly this reason.
Antiques and dating need more than a photo. The tutorial on AI for antique shops sets out what is and is not safe to ask a model, and the guide to the limits of AI pricing research explains why a confident number needs a source behind it.
Triggers that send an item to the specialist box
Write these on the wall by the sorting table, so no volunteer has to decide whether they are being cautious enough.
- The AI writes SPECIALIST BOX, or the sold median is $40 or more.
- Anything with a hallmark, a signature, a numbered edition or a certificate.
- Jewellery, watches, art, musical instruments and vintage electronics.
- Any bag, shoe or accessory that looks like a luxury brand, however battered.
- Anything that arrives with its original box, papers or a story attached.
The box is checked weekly by the manager or a volunteer with relevant knowledge, and it is the only place where a photo and an AI description feed a decision made by a person with expertise. If the shop has no such person, a professional valuer or an auction house may be willing to look at photographs. Until the item is priced by someone qualified, it stays off the shelf.
Test the speed target over two shifts
Do not switch the whole shop at once. Run two shifts with the same volunteers and similar donations. In the first, price 30 pile B items the old way with a stopwatch running. In the second, price 30 with the workflow. Record the total time, how many items went to the specialist box, how many prices were changed after ticketing, and, two weeks later, how many sold. An illustrative result:
| Measure (30 pile B items) | Old way | With the workflow |
|---|---|---|
| Total time | 52 minutes | 41 minutes |
| Average per item | 104 seconds | 82 seconds |
| Sent to the specialist box | 1 | 4 |
| Price changed after ticketing | 6 | 3 |
| Sold within 14 days | 21 (70%) | 24 (80%) |
The saving is 11 minutes across 30 items, and 82 seconds a time beats the 90-second target. The figures that matter more are the other rows: three more items reached a person with expertise, and half as many prices needed correcting. One pair of shifts proves little, because sell-through also depends on season and stock mix, so repeat the test before you change the rota. If you miss the target, look at where the seconds went. The usual culprits are poor light that forces a retake, a sold check that wanders into reading listings, and long AI replies. Fix them in that order: a window or a lamp, a rule that the check reads the last five sold prices and stops, and the shorter reply format in the prompt above.
From priced item to online listing
When an item clears the online threshold, the same photos and the AI's identification feed a listing. Ask for a draft title and description that uses only the measurements and flaws you supply from your own notes, and keep the price decision with the person. The tutorial on listing donated items from a phone photo covers the listing side, and the guide to testing AI item specifics against sold data shows how to check what the tool claims against real sales. The volunteer onboarding checklist is a good place to add a ten-minute session on the four piles and the never-list.
Start with one volunteer, one table and one shift. The aim is not to replace anyone's eye for a bargain. It is to give that eye two photographs and a sold listing to look at, so the shop prices more of what comes through the door correctly and spends its spare minutes on the items that deserve them.
Questions shop managers ask before pricing with AI
Can AI tell me whether a branded item is genuine?
No. A photo cannot show stitching depth, material feel, serial numbers or the other things an authenticator checks, and a chatbot sounds equally confident about a fake. Set aside luxury-brand items for a person who can check them, and do not write 'genuine' or 'authentic' on a ticket or listing unless someone qualified has confirmed it.
Is the free ChatGPT plan enough for a shift?
It can be for a quiet shift. OpenAI's pricing page lists limited messages with uploads on the Free plan, and each researched item uses at least one upload. If volunteers photograph 30 items a shift, test a normal shift before you rely on it, and compare the Go and Plus allowances on the plan page.
Can we do this without any AI tool?
Yes. The sold-price check does the real work, and it needs only a marketplace search with the sold filter on. Google Lens can help identify an item from a photo, and the band card prices the bulk of the stock. AI adds speed on identification and search terms, so try the process by hand for a shift first and add the AI where it saves seconds.
Further reads
- How Resellers Use AI to List Second-Hand Items Faster — Once an item is priced, list second-hand stock faster.
- How to Write eBay Listings With AI That Sell Faster — Write online listings for the items worth more online.
- Can AI Help a Small Shop Set Prices and Promotions? — Price promotions and mark-downs for the shop floor.
- AI for Marketing Content: A Vintage Clothing Seller's Weekly Drop — Turn the best finds into a weekly social post.
- Charity AI Risks: Data Protection, Deepfakes, and Donor Trust — The data and donor-trust risks that sit behind volunteer AI use.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI's image-input guide (limitations section), help page on Projects and ChatGPT pricing page; Google Lens product page; eBay's research tools page (Product Research in Seller Hub); all checked 28 October 2026. Every price, band, timing and sale figure in the examples is illustrative; build your own from your till data.