It’s Sunday night. The sale closes Tuesday. You’ve got three hundred lots photographed and a description written for maybe forty of them, and you’re staring down the retyping of the other two hundred and sixty before you sleep. That’s the job AI is honestly good at — taking the first swing at the boring part so you’re editing at ten o’clock instead of starting from a blank box at midnight.
But here’s the thing that should give you pause, and it’s the whole reason this page exists: these tools never sound unsure. A vision model will call a Weller vase a Roseville, or a plain Pyrex bowl a piece of Fire-King, with the exact same cheerful confidence it uses to tell you the sky is blue. It doesn’t know the difference between a fact and a good guess, and it’ll hand you both in the same friendly sentence. Paste that into your catalog and the one bidder in the room who does know catches it — and now it’s your name, not the robot’s, that took the hit in front of people who came because they trust your word.
That’s exactly why our AI is built the way it is. Every draft it hands you — title, category, condition, value range — comes stamped with a confidence score, so you can see at a glance where it’s sure and where it’s reaching. A second AI reads the first one’s work and pushes back before any of it reaches you. And on the things it has no business guessing at — coins, jewelry, fine art, anything in a thin market — it’s built to say “I don’t know” instead of filling the blank. You approve every lot before it goes live. It saves the typing, not the deciding.
So here’s the plainspoken list of what to keep away from the robot — and, on each one, how a tool that’s honest about its own limits keeps you out of the ditch. The through-line never changes: AI drafts the boring parts. Your judgment is still the job.
1. Authenticity and grading on anything worth real money
This is the big one, so it goes first. A model looking at a photo is pattern-matching against a lot of pictures it saw once. It’s genuinely useful for “this looks like a mid-century Danish teak sideboard, here are search terms.” It gets dangerous the moment money rides on the model being right instead of in the neighborhood.
It cannot tell you:
- Whether a signature is authentic or a later addition
- Whether a coin is MS-63 or MS-65 — two grades apart can mean several times the price — or whether it’s been cleaned in a way that guts its value
- Whether that’s a period piece or a 1970s reproduction of it
- Whether the maker’s mark is real, transferred, or fantasy
- Whether a “Tiffany” lamp is Tiffany or a shade someone married to a base last week
You already know grading firearms, coins, cards, and glass is a trained eye and a loupe, not a phone. Nothing about that changed. Where AI earns its keep is before the expensive part: pointing you at what to research.
That’s the exact line our AI holds on purpose. Show it a coin, a diamond ring, a signed canvas — and instead of handing you a grade or an “attributed to,” it tells you it doesn’t know and points you at what to check. It runs the identifier lookups it’s actually good at — VIN, ISBN, UPC, a maker’s mark it can read — and stops where the loupe begins. A tool that refuses to guess on the expensive stuff is the only kind you can safely put your name behind.
“I have a pressed glass compote with a pattern of thistles and a scalloped rim. What patterns and makers should I look into, and what details would tell them apart in hand?”
That’s a good use. It gives you a reading list. Then you or a specialist confirm it in hand. What you do not do is paste the model’s answer into the catalog as fact. If it matters and you’re not sure, the honest catalog line is a description with no attribution — “pressed glass compote, thistle pattern” — and you let the room decide, or you pay an expert. “Attributed to” is a promise. Don’t let a chatbot make it for you.
2. Hidden condition from a photo
A photo shows the side facing the camera in the light you gave it. AI can only work with what’s in the frame, and condition problems live exactly where the frame isn’t: the hairline you only catch when you tilt it, the repair under the base, the smell of smoke or mildew, the drawer that sticks, the veneer lifting at the back corner, the crack you feel with a fingernail before you ever see it.
Ask a model “what’s the condition of this vase” from one photo and it will confidently describe the condition of a vase. Whether that description matches your vase is luck. It has no way to see a flaw it can’t see.
Our AI will draft you a condition line off the photos, because that’s a useful starting point — but it comes with a confidence score that’s honest about what it is: a read of one picture, not a report from your hands. The score is there to remind you the last step is always the same. Pick it up. Turn it over. Run a light across it at an angle. Tap the porcelain and listen. That’s the part that protects you from the “this wasn’t disclosed” email three days after the sale.
And the flip side, which matters just as much: fixing a photo’s lighting, background, and crop is fair game. Editing out a chip, a stain, a crack, or a repair is fraud. It doesn’t matter that the tool made it easy. If the bidder can’t see the damage that’s there, you’ve misrepresented the lot — and one refund plus one ruined reputation costs more than every photo you ever cleaned up. Straighten it, light it, crop it. Never launder it.
3. Pricing thin markets and rare pieces
AI is at its worst exactly where you need help most: the odd item with no comparable sales. A model can be genuinely handy when there’s a deep, recent market — a common truck, a current-production rifle, a mass-market collectible that sells every week. There are real numbers to point at.
But the thin verticals — regional folk art, an obscure maker, a one-off piece of farm equipment, a local historical item — are the ones where a raw chatbot will invent a number that looks authoritative and rests on nothing. It doesn’t say “I don’t know.” It fills the blank. And that made-up estimate is worse than no estimate at all, because it anchors you, it anchors the consignor, and it can blow up a reserve conversation six weeks before the sale.
This is the single place our AI’s honesty earns its whole keep. When it drafts a value range, that range carries a confidence score — and when the market underneath it is too thin to stand on, it’s built to say so and leave the number blank rather than dream one up. It would rather admit the gap than protect your credibility with a guess and cost you both later.
The honest move on a thin item: skip the estimate entirely and set a sensible opening bid. “No estimate, opening at $50, let the room find it” is a defensible, professional position. A fake $400–$600 that came out of a chatbot is not.
And if you do want AI in the room on value, make it show its work — comps, not vibes:
“Don’t estimate a value. Instead, tell me what recent sold comparables I should search for, on which venues, and what search terms to use. If you don’t know of real recent sales, say so.”
A model that hands you a list of things to go verify is doing useful work. A model that hands you a dollar figure it dreamed up is handing you a liability. Cite, don’t guess — and refusing the guess is on you.
4. Running the bidding
This one’s short because it should be. Do not put an AI in charge of the auction — not calling it, not clerking the money, not deciding when the hammer falls, not managing reserves or anything shill-adjacent. The bid-taking, the increments, the “fair warning,” the read of the room — that’s the licensed, bonded, legally-accountable heart of what you do. It’s also the part with rules attached in most states.
The bidding in Hammerwerks runs on plain arithmetic, not a chatbot’s judgment: increments, proxy and max bids, buyer’s premium, soft-close — rules you can predict, not a guess dressed up as a decision. And a human is still on the block. A cheerful software bug in a chatbot is an embarrassing typo in an email. A cheerful software bug deciding your bidding is a lawsuit and a complaint to your state board. Keep a person calling the sale. Always.
5. Talking to your bidders as if it were you
It’s tempting to point a chatbot at your inbox and let it answer buyers. Resist letting it speak for you unsupervised. Your bidders deal with you because you’re a straight shooter who knows the goods. An AI answering “is the transmission good on the tractor?” will produce a confident, friendly, and possibly completely wrong answer — and now you’ve made a representation about a lot that you never actually made.
Every answer about a specific lot — condition, provenance, pickup, terms — is a promise with your name on it. Drafting is fine; sending blind is not.
Where it’s genuinely useful: drafting the general stuff you then read and approve. Your buyer’s-premium explainer. Your pickup-and-removal instructions. A polite reply to “where are you located.” Have it write the boilerplate, you read every word, you hit send. The second the question is about a specific lot’s condition or history, you answer that one yourself.
6. Anything legal, tax, or contractual
Consignment agreements. Sales tax on out-of-state buyers. Estate and probate questions. Firearm transfer rules and who can hold an FFL. Whether you owe a 1099 to a consignor. Bond and licensing requirements. Lien and abandoned-property procedures.
A chatbot will answer all of these instantly and some of the answers will be wrong, out of date, or right for the wrong state. It doesn’t know which state you’re in unless you tell it, it doesn’t know this year’s rules from three years ago’s, and it is not liable when it’s wrong — you are.
Here’s where we’re careful to stay in our lane, too. Hammerwerks will give you a consignment agreement your seller e-signs and you countersign, fingerprinted and recorded so you can later prove exactly what was agreed and when — and prove it wasn’t changed after the fact. That’s a solid record to keep. It is not legal advice, and we’d never dress it up as “legally binding” or “notarized” — that’s a conversation with your own attorney. Use AI, at most, to write down better questions for the professional you’re actually going to ask:
“I run auctions in [state]. Help me make a list of questions to ask my accountant about sales tax on online buyers from other states.”
Then you ask the accountant. The lawyer, the CPA, and your state auctioneer’s association exist for exactly this, and they carry the responsibility a chatbot never will.
The line, in one sentence
If getting it wrong costs you money, a consignor, or your license — a person decides, and that person is you. If getting it wrong just means you rewrite a paragraph — let the machine take the first swing.
That’s the whole framework. AI is a fast, tireless, occasionally-confident-and-wrong assistant that’s great at first drafts and terrible at accountability. Hand it the catalog copy, the photo cleanup, the research reading list, the boilerplate email. Keep the grading, the condition, the thin-market pricing, the bidding, the promises, and the legal questions where they belong — with the expert. That’s been you the whole time.
The difference with the AI built into Hammerwerks is that it’s designed to respect that line instead of daring you to enforce it. Confidence scores on every field. A second AI checking the first. “I don’t know” where a made-up answer would burn you. And you approving every lot before it’s live. It does the typing so you can do the deciding — and it never gets a chance to embarrass you in front of a room full of people who trust your word.