AI Tools for Cataloguing Stock: From Photos to Item Records

Photograph each item against a plain background from several angles with a ruler in frame, upload the set to a chat assistant that reads images, and ask it to fill a fixed record: title, category, materials, visible condition, measurements to confirm and search terms. Then check provenance, condition and dimensions against the item itself, because the tool only describes what the photos show.

That last clause is the whole difference between a useful assistant and a risky one in this trade. An AI that sees a brass candlestick can say "brass, turned stem, domed base, light tarnish". It cannot say the base is weighted with plaster, that the dent is a repair, or that the piece is a reproduction. For a dealer, the value of the process is in the speed of the first draft, and the integrity of the record comes from you. For the question of dating and valuing, which this tutorial deliberately leaves alone, see what AI can and cannot do on antique stock.

Follow me on Instagram@sagnikteaches

Why a dealer's records are not a library's

Library cataloguing is about classification: where an item sits in a system so it can be found. A dealer's record has other jobs. It has to support a price, back up a description to a sceptical buyer, remember what you paid, say where the item is on a shelf, and feed listings on your site or a marketplace. That changes what you ask a tool to do:

Connect on LinkedInSagnik Bhattacharya

• Describe, do not classify. You want accurate words for what is visible, not a catalogue number scheme.
• Keep facts and guesses apart. A record that mixes "Georgian" (a guess) with "oak" (visible) is more dangerous than a blank.
• Carry the money columns. Cost, date acquired and source are for you. They never go near a public listing.
• Remember uniqueness. One-off items need one record each, with stock numbers you assign.

The record layout: decide before you photograph

A fixed layout keeps every item comparable and makes the spreadsheet usable for filters. Here is a practical set of columns for a small antiques and vintage dealer:

Subscribe on YouTube@sagnikteaches

The "Filled by" column is the safeguard. Anything a tool might plausibly guess, such as materials and faults, is labelled as a draft. Anything that is knowledge only you hold, such as provenance and cost, is yours. If you are unsure whether a spreadsheet will carry you as stock grows, the comparison in AI inventory tools for small retailers is a useful next read.

Take photographs the tool can actually read

Poor photos make poor records, whether a person or a tool reads them. A routine that takes about two minutes an item:

1. Plain, even background. A sheet of white or grey card, a window to one side, no direct sun.
2. Six shots. Front, back, both sides, top or underside, and one close-up of any mark, signature, hallmark or fault.
3. A ruler or a scale card in frame in at least one shot. The tool will then guess dimensions from the ruler, but you will still measure.
4. The stock tag in the first shot so the item and the number cannot be separated later.
5. File names with the stock number. 0412_a to 0412_f, so a folder sorts itself.

Maker's marks and labels deserve their own close-up. A tool reading a hallmark from a blurry photo may produce a confident but wrong reading. If a mark matters, read it yourself with a loupe and type what you see, then ask the tool for possible reference sources to look up, never for the answer.

A main worked example: a mixed shelf of six items

Illustration only. Say a two-person vintage and antiques shop has bought a mixed box at a house clearance and wants to record six items before anything goes on the shelf: a brass candlestick, a stoneware jug, an oak stool, a pair of enamelled tin signs, a ladies' wristwatch, and a leather suitcase. They have six photos each, stock numbers 0412 to 0417 on tags, and the dealer's own measurements.

The first prompt sets the rules once, per batch:

You are helping a small antiques and vintage dealer create stock records from photos.
For each item I upload, return one row of a table with these columns:
StockNo (from the tag in the photo, or UNKNOWN), Title (60-80 characters, plain description, no value words
like "rare", "valuable", "antique" or "Victorian" unless the photo itself shows a marking that says so),
Category (choose from: Lighting, Ceramics, Furniture, Signs, Watches, Luggage, Other), Description (2-3 sentences, visible features only),
Materials (mark each as VISUAL ONLY), VisibleFaults (list every mark, chip, crack, stain, repair or missing part you can see,
and say which photo shows it), Questions for the dealer (things you cannot tell from photos).
Do not estimate age, maker, value or origin. If you think something might be relevant, put it under Questions.
Do not guess measurements. I will measure.

The shape of the output matters. A typical, illustrative row for the candlestick:

0412 | Brass candlestick, turned stem, domed base, 22 cm tall (to confirm) | Lighting |
Single brass candlestick with a turned stem, ringed collar and circular domed base. Surface shows an even
dark tarnish with brighter wear on the stem. | Brass (VISUAL ONLY) | Light scratches on base underside (photo d);
small dent near rim of nozzle (photo e) | Is the base weighted? Any maker's mark under the base? Is the nozzle
original or replaced?

Several things stand out. The assistant has not guessed an era. It has pointed to photos for each fault, so the dealer can look at exactly the right spot. And its "Questions" column has turned its uncertainty into a to-do list. The dealer checks the base, finds it is weighted, adds that to the record, and measures the height, which is actually 21.5 cm, not the 22 cm the tool read off the ruler despite being told not to guess measurements. That small correction is the reason dimensions stay in the "Filled by: You" column.

Total time for the six items, illustrative: photographing 12 minutes, upload and prompts 8, reading and correcting the table 25, measuring and condition grades 20. About an hour for six items, compared with perhaps 90 minutes by hand, with a more consistent description style as the main gain.

Other examples across different stock

A stoneware jug. The tool describes "salt-glazed stoneware jug, cylindrical body, strap handle, impressed band decoration". It notes a hairline in the rim on photo c, which the dealer confirms with a fingernail. The draft says nothing of a possible repair. In the Questions column, the tool asks "Is there a repair under the handle? Tap test for cracks." The dealer taps it, hears a dull note, and grades it C with "old crack to handle, stable" as a fault.

An oak stool. The tool writes "oak" and the dealer marks it VISUAL ONLY, then checks the underside and sees pine in the stretchers. The record becomes "oak top, softwood stretchers". A buyer who later reads the description will not be surprised. This is why the Materials column is always flagged until you have checked.

Enamelled tin signs. The tool transcribes the lettering from the photos. Dealer's check: read each word against the sign, since optical reading of lettering in an old typeface is a common place for small errors. A single wrong letter in a transcription can send a buyer's search to the wrong place.

The wristwatch. The tool describes the dial and the case. It cannot say whether it runs, whether the case back is original, or whether the dial has been refinished. The dealer adds "not tested; sold as seen" to the record. A rule I would keep: if a statement would change what a buyer pays and you have not verified it, the record says "not verified".

The suitcase. Smell, lining condition and hardware function are invisible to the camera. The tool lists "lining visible in photo f, water staining" and asks about smell and locks. The dealer opens it, notes a musty smell, and adds "airing needed" to the internal notes.

Condition and provenance: the human checks

These two fields carry the most risk and the most trust. Here is how I would handle each.

Condition

Create your own grading scale and write it down, so you and any assistant use the same words. A simple version: A is excellent, with no visible faults beyond light age wear. B is good, with minor faults listed. C is fair, with significant faults listed. D is for restoration or parts. The grade is yours, assigned with the item in your hand. The tool's role is to list faults it can see in the photos, which you then confirm, add to and sign off. A realistic mistake: relying on photos to grade a glazed ceramic, missing a hairline crack that only shows in raking light. Hold the item up to a lamp or tap it, as you would anyway.

Provenance

Provenance is the item's history: where it came from, who owned it, what paperwork exists. It is yours, because only you know the house clearance, the receipt, or the previous owner's story. Record the facts you hold with their source ("bought from a house clearance, receipt 2291") and leave out anything told to you that you cannot support. If you suspect something, such as a possible reproduction, record it as a suspicion in your internal notes, not as a fact. Never let an AI-suggested origin enter this column.

Prompts for the next steps: listing text and search terms

Once the record is verified, the same assistant can turn it into public text. Feed it only the verified, public columns, not your cost or source:

Using only the verified record below, write a listing for my website: a title of up to 80 characters,
a 90-120 word description, and 8 search terms. Do not add facts that are not in the record. If the record says
"not tested", say "sold as seen". Include the fault list in plain words.
Record: [paste the row without CostPrice, DateAcquired, Source]

The result is predictable and safe because the inputs are fixed. For longer marketplace text, see AI for Etsy titles, tags and descriptions and, for your own site, Shopify product descriptions with AI. A before-and-after for the candlestick: before, "Lovely old brass candlestick"; after, "Brass candlestick with turned stem and weighted domed base, 21.5 cm high. Even tarnish; small dent near the nozzle rim; light scratches under base. Not polished." The second version sells better because it answers questions before a buyer asks them.

Keeping the spreadsheet honest

Records decay. Items get sold, moved or reduced, and a spreadsheet that does not reflect that is worse than none. A few habits:

• Status column discipline. Change it the minute an item sells or is reserved.
• Weekly spot check. Pick ten records at random, find the items on the shelf, and compare location and condition. Fix the process if more than one is wrong.
• Duplicate check. A filter for repeated stock numbers and a count of rows against tags printed.
• Locked columns. Protect StockNo and CostPrice so a stray paste does not overwrite them.
• Backups. A dated copy of the sheet each week.

If you would like to tie records to stocktake, the tutorial on running a stocktake with a phone scanner covers the counting side.

Batching: how to run fifty items without losing track

A single item is easy. A van-load is where records slip. Cataloguing in batches of six to ten items keeps each chat short, keeps the tool's attention on the photos in front of it, and keeps your corrections manageable. A routine I would use for a clearance of fifty items:

1. Tag first. Number every item with a tag before a single photo is taken. Nothing is photographed untagged.
2. Photograph in one session with the same background and light, item by item, in tag order.
3. Upload in batches of eight. One chat per batch, starting with the rules prompt above. Reusing the same instructions each time keeps the column order identical, which makes pasting into the sheet safe.
4. Paste into a staging sheet, not the main stock sheet. A separate tab named "Draft" holds the raw output until you have checked it.
5. Verify with the items at hand. Work through the staging rows with the physical items on a table in tag order. Fill in measurements, grade, provenance and the answers to the Questions column.
6. Move verified rows to the main sheet and mark the batch complete in a log: date, tag range, who checked.

The staging tab is the single best protection against a draft slipping into live records. Illustrative timing for fifty items: tagging 20 minutes, photographing 100, AI drafting 40 (mostly waiting), then reading the drafts and verifying them with the items about 350, or seven minutes an item. About eight and a half hours in total, in which the verification step is the longest and the one that cannot be skipped.

A before-and-after record for a harder item

Take a mid-century teak sideboard, which looks simple and hides several traps. The photos show a long low cabinet with sliding doors and tapered legs. The first AI draft reads:

0431 | Mid-century teak sideboard with sliding doors and tapered legs | Furniture |
Low sideboard in teak with two sliding doors and three drawers, brass handles, tapered legs. | Teak (VISUAL ONLY) |
Water ring on top surface (photo b); scuffs to left leg (photo d) | Is it solid teak or veneer? Are the legs original?
Is there a maker's label inside a drawer?

The dealer's checks turn it into the final record:

• Opens a drawer: a printed label gives a manufacturer's name, which goes into Provenance as "label inside top left drawer reads: [as printed]". No era is claimed, because a label is evidence of a maker, not of a date.
• Looks at the edge of the top: teak veneer on a carcase of another timber. Materials become "teak veneer on softwood carcase".
• Measures: 160 cm wide, 45 cm deep, 72 cm high. Asked for a rough width in a follow-up, the tool had said "about 150 cm".
• Tests the left leg: the scuffs hide a repaired crack. VisibleFaults gains "old repair to left front leg, stable".
• Grades it B with faults listed.

The verified row carries four corrections that the photo alone could not supply. Without the checks, the first draft would have called the piece solid teak and left out the repair, and the dealer would have heard about both from a buyer.

Spreadsheet checks that catch bad rows

Once rows arrive, simple formulas find the commonest slips. All of these work in Excel and Google Sheets. Put them in a hidden "Checks" column or a separate tab:

• Missing measurements. =IF(OR(F2="",G2="",H2=""),"MEASURE","OK") flags any record without height, width or depth, assuming you keep those three in their own columns (F to H here) rather than in one Dimensions cell.
• Duplicate stock numbers. =IF(COUNTIF($A$2:$A$500,A2)>1,"DUPLICATE","OK").
• Value words in the title. =IF(OR(ISNUMBER(SEARCH("rare",B2)),ISNUMBER(SEARCH("antique",B2)),ISNUMBER(SEARCH("valuable",B2))),"CHECK WORDS","OK"). Add your own list. The aim is to stop unverified claims getting into titles.
• Unfinished verification. =IF(COUNTIF(E2,"*VISUAL ONLY*")>0,"CONFIRM MATERIALS","OK") so that a "VISUAL ONLY" tag has to be deleted by hand before the record counts as final.
• Title length. =LEN(B2) if you list on marketplaces that cut off long titles.

The last of these ties to the "VISUAL ONLY" label in the prompt. Because the tool writes that phrase into every material field, you can find every unverified material with one search, and clearing the phrase becomes a visible act of checking.

Privacy and ownership of the photos

Most stock photos are safe to upload, but three cases need care. Photos taken in a seller's home may show family pictures, addresses or documents in the background, so crop or retake against your own backdrop. Items with letters, diaries or papers inside should have the contents photographed separately and only if you have a right to do so. And items that may be stolen or of disputed ownership are a matter for you, your trade association or the police, not for a chat assistant. For the tool itself, use a business plan or switch off training on your chats, and delete uploaded batches when the job is done if the tool allows it. Check the vendor's current data settings, as they change.

What the time saving actually looks like

Illustrative sums only. A dealer who records stock by hand might spend about 15 minutes per item on photos, measuring, title, description, faults and tidy-up: 50 items, 12.5 hours. The batch routine above came to about eight and a half hours for the same fifty, roughly ten minutes an item, which matches the hour for six items in the first example. The saving of about four hours is real but modest. The larger benefit is consistency: every record has the same fields, so filtering by category or fault, building a listing and finding an item on a shelf get faster for as long as you trade. A shop with 800 items in the back is where the habit pays most.

What AI tends to get wrong with old objects

• Dating by style. Anything it says about a period is a guess based on look. Reproductions copy styles exactly.
• Material confusion. Brass, gilt metal, bronze and plated items look alike in photos. Silver plate and sterling are hard to tell apart in photos.
• Missing repairs. A glued crack or replaced part is often invisible.
• Misread marks. Hallmarks and maker's stamps are small, worn and easily misread.
• Too-clean language. "Excellent condition" appears in drafts unless the prompt forbids value and quality words.
• Invented measurements. Scale guesses from photos can be several centimetres out.

A final rule I would adopt: the draft can be fast and imperfect, but nothing reaches the shelf or the website until a person has held the item, compared it with the record, and initialled the sheet. That one step is the difference between a time saver and a source of returns.

Dealer questions about AI and stock records

Can AI tell me what an antique is worth?

Not reliably. It can describe what it sees and suggest search terms, but values depend on condition, provenance, maker, market and recent sales that the tool cannot see. Use its description to search sold-item records and ask a specialist for anything valuable. Never record an AI-suggested value as an estimate.

Which fields should an item record have?

At minimum: your own stock number, short title, category, description, materials, measurements, condition grade, faults, known provenance, cost price, asking price, date acquired, location in the shop, and photo file names. Add a status column so sold or reserved items do not stay listed.

Should I let AI write the condition notes?

Let it draft a description of visible wear from clear photos, then verify with the item in your hand. Condition is the field where buyers complain. Tools miss hairline cracks, repairs, smells, loose joints and replaced parts. The final condition grade should always be yours.

Is it safe to put my stock photos into a chat tool?

For ordinary stock, yes on a business plan or with training switched off. Remove customer details, house numbers and anything private visible in the background of the photos. Check the vendor's data settings before you upload a large batch, and avoid sending items whose ownership is in dispute.

Further reads

• AI Product Photos vs a Photographer: Cost and Quality Compared — When listing photos are worth paying a photographer for.
• AI Tools for Ecommerce Photography: Backgrounds, Shadows, Consistent Sets — Clean backgrounds and shadows for listing photos.
• How to Write Product Descriptions With AI That Actually Sell — Description copy once records are verified.
• How to Create Product Photos With AI on a Small Budget — Improve the photos before you catalogue them.
• Can AI Help a Small Shop Set Prices and Promotions? — Set asking prices with your own sales evidence.
• ACES and PIES Data: AI Cleans and Checks Your Parts Catalogue — What ACES (fitment) and PIES (product data) cover, the mismatches AI can find between them, and a sample audit that catches wrong fitment early.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help centre (Data controls in ChatGPT) and OpenAI business data page; Claude privacy centre pages on model training for consumer and commercial plans; cataloguing practice for second-hand and antique stock. No valuations are quoted.