Returns that sell again · bol
What if every returned product on bol got a fair second chance?
A concept for the returns section of bol's Partnerplatform. AI reads the buyer's reason and photos, suggests whether an item goes back on sale, gets repaired, is sold as Tweedekans or recycled, and the seller approves with one click.
Self-initiated concept. Not affiliated with or endorsed by bol.
- Role
- Self-initiated concept, from interviews to tested prototype
- Company
- bol (concept)
- Timeline
- 2024 · 8 weeks, evenings
- Team
- Solo, with 9 bol sellers who joined interviews and tests
- Platforms
- bol Partnerplatform (web), bol app (buyer return flow)
1Reason and condition read from the buyer's text and photos
2The seller approves or changes each suggested route
3Share of returns resold this month, shown at the top
11 → 4 min
Time for ten returns, moderated tests
87%
Suggested routes testers kept unchanged
52 → 81
Usability score (SUS), current flow vs concept
9 of 9
Sellers who preferred the concept
The problem
The bol sellers I spoke to handle returns one by one, and decide what to do with each item on gut feeling. Small sellers often write off items they could have resold, because checking every return takes too long.
What I did
I interviewed nine bol sellers, analysed 600 anonymised return reasons they shared with me, and designed a concept in bol's own visual language: AI reads the reason and photos, suggests a route, and shows sellers which listings cause avoidable returns.
What changed
In moderated tests with the same sellers, ten returns took 4 minutes instead of 11, testers kept 87% of the suggested routes, and all nine preferred the concept to how they work today.
01 · Context
Why I picked bol.
bol is where most Dutch people shop online, and tens of thousands of independent sellers sell through it. At PackBack I spend my days on reuse, so returns caught my attention: every returned product is a small decision between selling it again, repairing it, or throwing it away.
This is a self-initiated concept. I don't work for bol and had no access to their data. Everything here comes from interviews with sellers, data they chose to share, and bol's public Partnerplatform and design language.
- Sold↓
Shipped to the buyer
Seller
- Returned↓
Reason and photos from the buyer
Buyer, bol app
- Assessed↓
AI suggests condition and route
Partnerplatform
- Routed↓
Restock, repair, Tweedekans or recycle
Seller
- Sold again↺ back to 01
Often as Tweedekans
bol
02 · Problem
Decisions made in a hurry, item by item.
Every seller I met had their own system: a spreadsheet, a shelf of 'to check' boxes, or simply a rule of thumb. The buyer's reason and photos arrived in one place, the decision happened somewhere else, and nobody looked for patterns across returns.
The question I worked on
How might a bol seller decide what to do with each return in seconds, and send more of them back into use?
03 · Research
Nine sellers, mostly in their stockrooms.
I found sellers through LinkedIn and seller communities, visited five of them, and spoke to four over video. Four sellers also shared anonymised return reasons from the last three months.
The reason text says more than the reason code
Buyers chose 'damaged' for anything from a dented box to a broken product. The free text and photos told the real story.
“Half the 'damaged' ones just have a scratched box.”
Sellers want help deciding, and the last word
Every seller wanted suggestions. None wanted bol to relist their stock without asking.
“Tell me what you'd do, but let me press the button.”
Many returns are predictable
A few products caused most returns, often for one fixable reason, like a size chart that ran small.
“A third of my jacket returns were the same size issue. I only found out by accident.”
04 · Process
Eight weeks of evenings.
I rebuilt the returns screens in bol's visual language, from the navy Partnerplatform bar to the heavy slab headings, so sellers tested something that felt like their own tool. The AI suggestions in the prototype came from a small language-model classifier I ran over the shared reasons.
Automatic relisting
Fastest on paper. Sellers refused, and a wrongly relisted item hurts their rating.
Better return reasons only
Improved the data, but sellers still had to judge every item themselves.
Suggested route, seller approves
AI does the reading and suggests a route with its reasons. The seller approves or changes it.
05 · Design decisions
Read the evidence, suggest a route.
Reason and condition come from what the buyer actually wrote
The concept highlights the phrases in the buyer's text and marks the spots in their photos that led to the suggestion. The condition grade comes with its evidence, so a seller can check it in seconds.
Tweedekans is a first-class route
bol already sells returned items as Tweedekans. The concept makes it one of four routes, with a suggested discount based on similar items that sold. In tests it was the route sellers were most surprised by and most likely to use.
Like new
Unopened, or packaging damage only.
Sell as second-hand
Light signs of use. Suggested price −15%.
Fix first
A fixable fault, sent to a repair partner.
End of life
Broken beyond repair, to a certified recycler.
Insights that prevent the next return
Return reasons are grouped per product with a suggested fix. For one seller's rain jacket, 34% of returns said 'too small', and the concept suggests adding a size note to the listing.
A buyer flow that collects better evidence
Better suggestions need better input. In the bol app, the return flow asks for the reason in plain words and a photo for damage claims, with one line on why it helps.
06 · Results
What the tests showed.
I tested the prototype with the same nine sellers, using ten real returns from their own shops. Each seller did the task twice: once the way they work today, once with the concept.
“If this was in the Partnerplatform tomorrow, I'd use it tomorrow.”
“I didn't know Tweedekans could be this easy.”
07 · Reflection
What a concept can and can't prove.
Nine sellers and ten returns each is enough to test a direction. Promising numbers would need more. The next step would be real return data at bol's scale to tune the routes and the price suggestions.
It confirmed something from my reuse work: people happily give products a second life when the effort is low and the money makes sense.
Tools on this project
Figma, FigJam, Maze, Python and a small LLM classifier, Dovetail
Names of the people quoted have been changed to protect their privacy.
One screen for 'why is my bill higher?' →