It’s Sunday night. The sale is Tuesday. You’ve got three hundred lots photographed and not one of them written up, and you already know who’s writing them, because it’s you, same as always. So you sit down and start typing. Around lot eighty you begin cutting corners. By lot two hundred you’re describing things you can barely remember shooting. That’s the grind everybody who’s run a sale knows in their shoulders.
Handing that off is the easy wish, and it’s the honest one: AI is genuinely good at the typing. Give it a decent photo and a few facts and it’ll draft you a clean title and description faster than you can. That part is real, and it’s here today.
But there’s a second thing buried in the same job, and it’s the one that can actually hurt you — the machine sounding certain when it’s wrong. You feed it a photo, it comes back “Fire-King jadeite bowl, $65–$90,” sure right down to the dollar. Except it’s a plain Pyrex bowl worth about six bucks, and now that number is sitting in your catalog. A bidder who knows the difference just decided you don’t. That’s the real cost. Not the wrong price — your name next to it.
So there are two completely different things one of these tools does when you ask it about a lot, and they look identical on your screen. One saves you an evening. The other embarrasses you in front of a consignor six weeks later, when the lot brings a tenth of what you told him. The whole skill is telling them apart — and using tools that tell you which one you’re looking at.
Two answers that look the same — one saves your night, one costs your name
When a model answers from memory alone, it’s running a very fancy version of “what word usually comes next.” It has read a lot of text, so it has a vague sense that a Griswold #8 skillet is worth “some money,” and it’ll hand you a figure — $120, say — that sounds like an answer. But it didn’t look anything up. It has no idea what one actually sold for last month. That number is a guess wearing a suit.
The other kind of answer comes when the tool actually goes and looks first — searches, reads real listings and sold-price pages, and then tells you what it found, with links and a range built out of specific sales. That’s the one you can work with: a real spread to set an estimate and an opening bid against.
The difference between them isn’t the AI being clever. It’s whether it bothered to look — and whether it’s honest with you about which it did. Your entire job here is to force it to look, and refuse to trust it when it won’t.
Make it show its work — and grade its own certainty
The tool will hand you a bare number with a straight face, and at eleven o’clock on a Sunday a confident figure is awfully easy to just believe. Don’t. Make it prove where the number came from. Two quick tells that you got research instead of a guess:
- It gives you links. A grounded answer cites where it found each sale. No links, no trust.
- It names actual sales. “Three sold on [site] in the last 60 days for $95, $110, and $140” is research. “This is typically worth around $120” with no source is a guess. Throw it out.
If it hands you a number with no sources, ask it flat out: “Did you search for actual recent sold listings, or is that an estimate? Show me them.” Half the time it’ll admit it guessed and go look. That one question saves you from publishing fiction.
Better still is a tool that grades its own certainty before you have to ask. That’s the whole idea behind how the AI works inside Hammerwerks: every draft it puts in front of you — title, category, condition, value range — carries a confidence score, a second AI reads the first one’s work before it ever reaches you, and on the categories where guessing goes bad — coins, jewelry, art, anything with a thin market — it’s built to say “I don’t know” instead of inventing a number. Not because candor is a nice touch. Because the confident wrong number is the exact one that costs you, and a tool that flags its own shaky guesses is a tool that protects your credibility with buyers instead of spending it.
A prompt that actually works
Vague questions get vague, invented answers. Describe the item the way you’d describe it to another auctioneer, and tell it exactly what you want back. Something like this:
Search the web for recent completed/sold listings of this exact item. I need comparable sales, not a guess.
Item: Griswold #8 cast iron skillet, large block logo, heat ring, marked 704. Smooth cooking surface, light seasoning, no cracks or warp that I can see.
Find me 3–5 actual recent sold prices from the last few months. For each one give the price, the date, the condition described, and a link. Then tell me the range. If you can’t find real sold comps, say so — do not estimate.
Three things are doing the work in that prompt. You said sold, not “asking” — asking prices are wishful and often double what things bring. You gave it the identifying marks a collector would search on. And you gave it explicit permission to come back empty, which is the single most valuable instruction you can give an AI, because otherwise it will invent something rather than disappoint you.
Where to point it — free sources worth checking
The model is only as good as what it can reach, and for a lot of categories you can do the same lookups yourself in a couple of minutes. Often the smartest move is to have AI point you at the right database, then read the source with your own eyes:
- Completed/sold listings on the big marketplaces. The gold standard for common collectibles and tools. Filter to sold, not active. What people actually paid, not what they’re hoping for.
- Vehicles and equipment: the free NHTSA VIN decoder (search “NHTSA VIN lookup” or vPIC) confirms year, make, model, engine, and trim before you write a word of the catalog. Get the specs right and the value research gets easier.
- Books: free ISBN databases and used-book listings will tell you fast whether a box of hardbacks is a $5 lot or has one valuable first edition hiding in it. Most old books are worth the box. The AI can help you spot the exception, but confirm it on a real bookseller listing.
- Anything with a barcode: a UPC lookup gets you the exact product name and model, which is half the battle on modern tools, electronics, and household goods.
- Photo-first ID: Google Lens or your phone’s camera search is genuinely good at “what is this thing” — a maker’s mark, a pattern, a piece of pottery. Use it to identify, then research the value separately. Identification and valuation are two different jobs.
The part where the machine will lie to you
You have to know exactly where these tools fall on their face, because they fall confidently.
They cannot judge authenticity. An AI cannot tell you a signature is real, a mark is genuine, or a “Rolex” isn’t a $40 fake. It’ll cheerfully price it as real. That’s your eye and your specialist’s eye, not the software’s.
They cannot grade or see hidden condition. Condition is most of value on a lot of items, and the model is working from your photos and your words. It doesn’t know about the hairline crack you didn’t mention or the repair under the base. Coins, cards, glass, furniture — grade is everything and the AI can’t do it.
They mix up makers constantly. Ask about a piece of art pottery and there’s a real chance it’ll call a Weller vase a Roseville, or vice versa, and then price the wrong one. It sounds just as sure either way. When the answer hinges on the exact maker, go verify the mark yourself.
None of this means the tool is useless. It means the tool drafts, and you decide. Every single time. The AI is the eager new hire who did some legwork and brought you three listings — helpful, worth having, and absolutely not the person who sets the estimate.
The most dangerous number is the one with no comp
The hardest lots are the ones with no comparables — the genuinely rare piece, the oddball, the regional thing nobody else is selling. Here’s the trap: when there are no comps, a guessing AI produces its most confident-sounding numbers, because there’s nothing real to anchor it and nothing to contradict it. That number is pure fantasy, and it’s the one most likely to end up in front of a consignor.
When the honest answer is “no recent sales found,” the right move isn’t to invent an estimate. It’s to set a sensible opening bid and let the room find the price. That’s what a live auction is for. “No estimate — opening at $50” is honest and protects you. A made-up estimate of “$800–$1,200” on a thing that’s never traded just sets your consignor up for a letdown and makes you look like you didn’t do your homework.
A good rule: if you can’t point at a real sale, don’t print a number. An opening bid commits you to nothing. A fantasy estimate commits you to a disappointment. This is also why a value-drafting tool worth using will hand that one back to you unpriced and flagged, instead of filling the blank with something that sounds good.
A workflow you can run tomorrow
- Identify it. Photo search or barcode/VIN/ISBN lookup to nail down exactly what it is, marks and all.
- Ask the AI to search for sold comps — using the prompt above, demanding links and giving it permission to come up empty.
- Open the links and read them yourself. Check the sold ones actually match your item’s condition and variant. This takes two minutes and it’s the whole ballgame.
- Set your range from the real sales, not from the AI’s summary. If the comps say $95–$140, that’s your range.
- No comps? Set an opening bid and skip the estimate. Say so plainly.
- You sign off. The auctioneer decides. Always.
Used this way, AI isn’t doing anything mysterious. It’s a fast research assistant that reads listings quicker than you can and hands you the sources to check. It takes the typing off your Sunday night. It does not take the deciding — that’s still your eye, your room, your name on the estimate. Keep it on a short leash, never let it publish a number it can’t back up, and it earns its keep.
One honest aside: we build a lot of this into Hammerwerks so it lives right next to your catalog — photo-to-draft with a confidence score on every line, a second AI checking the first, VIN and ISBN and UPC and maker’s-mark lookups plus NHTSA and CPSC recall checks on the items that have them, and you approving every lot before a soul sees it. But you don’t need us to start. Everything above works today with the free tools already in your pocket.