Selling One-of-a-Kind Items Online: Automation When Every Product Is Different

Most e-commerce automation relies on repetition. This client had none: no product ever appeared twice.

The business is a vintage furniture brand. The founder buys old pieces, restores them and sells them online. Every piece is unique, with its own dimensions, its own wear and its own history. There is no replenishment. Once a piece is gone, it is gone.

When we started, the bottleneck was not sales. The pieces sold well. The bottleneck was the stretch between "piece is restored and standing in the workshop" and "piece is online". That stretch took a good 40 minutes per item, and it was the reason finished goods sat around unsold for weeks.

I am writing this up anonymously because we do not have permission to use names. The approach transfers anyway, and that is what matters here.

Why unique items follow different rules

In a normal shop you create a product once and sell it a thousand times. The effort per sale goes to zero.

With unique items it is the other way round. The effort per sale is the effort of creating the product, every single time. At a few hundred pieces a year, that is a full-time problem.

On top of that come three quirks that make the usual toolkit useless:

There is no master data. No item master, no variants, no reusable copy. Every piece needs its own dimensions, its own condition description, its own photos.

There is no inventory management. Stock is always one. A reorder setup like the one for purchasing makes no sense here, because there is nothing to reorder.

Shipping is a case-by-case decision. A sideboard 1.80 metres wide does not go out as a parcel. The price depends on dimensions, weight, destination and whether somebody has to carry it up a flight of stairs.

The starting point

We spent the first two weeks doing nothing but measuring. No tools, no proposals, just a tally of everything that cost time. We use this approach in every project, see How an automation project actually runs.

The result in rough numbers, per week:

Together that was more than half a working day, every day, for a founder who wanted to be restoring furniture. The enquiries kept pulling him out of the workshop.

One number that surprised us: around half of all enquiries were about information that could have been on the product page. Dimensions in centimetres, the state of the surface, whether a drawer sticks, whether the piece fits through a standard door.

What we built first: product creation

The biggest lever was product creation. The old process looked like this: take photos, pull the photos onto the computer, cut them out, measure with a folding rule and write it down, type a description, invent a title, upload everything to the shop.

Today's process has three steps:

• Photograph to a fixed pattern. Six shots per piece, always the same angles, always against the same wall. Plus a seventh shot: a sheet of paper with the handwritten dimensions and the purchase note.
• Upload into a folder. One folder per piece, and that is all the founder does.
• The rest happens on its own. A workflow picks up the images, crops them, reads the dimension sheet, generates title and description and creates a product draft in the shop.

The decisive trick was the photographed dimension sheet. We first tried collecting the dimensions through a form. The founder never used the form, because in the workshop he has no appetite for typing. A photo of a scrap of paper he takes anyway.

The copy comes from the images plus the keywords on the sheet. We work with fixed text blocks for the category and a free section for the specifics of the piece. How we build copy like this in general is in Product descriptions with AI.

Important: the workflow creates a draft, not a live page. The founder looks at every draft, corrects one or two sentences on average and publishes it. That takes around 5 minutes instead of 40.

What the condition part of the copy has to do

One point we had to fix twice: describing signs of wear.

On the first attempt the copy read too smooth. A scratch became "charming patina", a missing trim piece an "authentic detail". That produced complaints, because customers expected something other than what arrived.

We turned that part of the copy around. Condition details are now written plainly and with a location: "Scratch, roughly 4 cm, on the right side panel." Selling happens through the photos, describing happens through facts. Complaints about condition mismatch dropped noticeably afterwards.

The same lesson applies to every AI-generated product description: what the machine does not know for certain, it must not assert. More on that in How to avoid AI hallucinations.

Enquiries: the answer belongs on the product page

The second building block was enquiries. We could have built an assistant to answer them. We solved it the other way round.

First we went through 200 past enquiries and sorted them into groups. Four groups covered most of the volume: dimensions and doorway clearance, the condition of specific spots, shipping cost to a particular address, availability.

For the first two groups the answer was better product pages rather than automation. Every piece now has a fixed dimensions table, including the narrowest clearance width, plus a condition list with photos of the spots mentioned. That roughly halved the enquiries on those two topics.

For the rest, an assistant runs that knows the product data of the piece in question and answers from it. It handles about half of the remaining enquiries without a follow-up. Anything that sounds like negotiation, a special request or a complaint goes to the founder. We describe the underlying pattern in the guide to customer service automation.

The effect in rough numbers: of 5 to 6 hours per week, around an hour and a half remained.

Shipping quotes for bulky goods

Shipping was the technically most demanding part and the one with the clearest payoff.

Before, the founder requested quotes from two freight companies for every enquiry and waited a day for an answer. In that time a share of the interested buyers dropped out.

Today a workflow calculates the price directly on the product page. It needs four inputs that are attached to the product anyway: dimensions, weight, a bulky-goods category and the customer's postcode. Out of that comes a price based on a stored zone and size table, plus surcharges for upper floors without a lift and for two-person handling.

Two things mattered here:

• The table is our own work, not the freight company's price list. We built in surcharges and buffers so that one outlier does not eat the contribution margin.
• Above a certain size there is an enquiry instead of a price. Very large pieces and deliveries abroad show a button for an individual quote. A form beats a price that does not hold.

The calculation lands within the estimated range in about nine out of ten orders. The rest are cases with peculiarities no table can anticipate.

Reservations: the underestimated problem

With unique items there is a situation normal retail does not have: two people want the same piece, and one of them wants to view it at the weekend first.

That used to be handled by shouting across the room, with notes in the inbox. It happened that a piece was promised twice. That is the most embarrassing mistake in this business.

Today there is a reservation with an expiry date. The customer reserves through the product page for 48 hours, the piece disappears from sale during that time, and a workflow releases it automatically if nothing happens. Both sides get reminders, six hours before expiry.

The process is simple and that is exactly why it holds. A rule you can explain in one sentence is a rule you can also repair at eleven at night.

What stayed manual

Not everything could be shifted, and some of it should stay where it is.

• Buying. Whether a piece is worth acquiring comes down to a look at condition, provenance and the market. There is no data foundation for it, because no two cases are alike.
• Restoration. The actual value in this business is created by hand.
• Pricing. We built a proposal based on comparable pieces that had sold. It was wrong too often, because condition drives price more than category does. Today the proposal is only a range, and the founder decides.
• Taking photos. Plenty gets cropped automatically, the shutter is pressed by hand.
• Viewings and collections. An appointment on site is a sales conversation. A human belongs there.

What we switched off again

We built two things and pulled them back after a few weeks.

Automatic categorisation from photos. We wanted to derive era and style from the images. The hit rate was around two thirds, which is too low for a range with a collector audience. A wrongly assigned decade costs credibility. The founder now picks the category from a list, which takes five seconds.

Automatic price cuts for slow movers. The idea: anything sitting for 90 days gets cheaper. In practice, the pieces that sat longest were exactly the ones waiting for the right buyer. Two of them found buyers at full price after months. A time-based rule does not fit a business where patience is part of the margin.

The result, in rough numbers

I am deliberately not naming revenue figures. What follows is what is measurable about the approach.

• The time from finished restoration to being online fell from around 40 minutes to about 5 minutes per piece.
• Weekly effort for product creation, enquiries, shipping and reservations fell from the order of 18 to 23 hours to 5 to 7 hours.
• Enquiries about dimensions and condition roughly halved, without anyone answering more of them.
• Double reservations have not occurred since the switch.

The effect that mattered most to the founder appears in no table: he can buy pieces again without fearing they will sit unlisted in the workshop.

What you can take from this

Even if you do not sell furniture, three points apply to any shop with a changing range:

• Automate the capture, not the decision. The path from reality into the shop is the time sink. Judging a piece is not.
• Use the input route your user takes anyway. A photo of a scrap of paper beats any form nobody fills in.
• Answer questions on the product page before you build an assistant for them. The cheapest answered question is the one that never gets asked. We describe the same idea for shipping in Automating WISMO enquiries.

Conclusion

A range with no two identical products is no reason to rule out automation. It only shifts what you automate: the repetition in capturing data instead of the repetition in selling.

The leverage here sat in four building blocks: product creation from photos, answers directly on the product page, our own shipping table for bulky goods, and reservations with an expiry date. Buying, restoration and pricing stayed with a human, and that was the right call.

If your range is similarly awkward and you do not know where to start, drop us a line at Flowhouse. We will also tell you if the effort does not pay off at your volume yet.