The AI engine room of a donated-goods retailer.

Retail where nothing has a barcode twice.

A normal store sells the same product thousands of times. A donated-goods retailer almost never sells the same thing twice, so Swedemom's small crew had to figure out, for every single item, what it is, what it honestly sells for, and whether it's worth listing at all. That per-item judgment was the constraint capping the whole operation.

The best year in Swedemom's history.

Sales slid for five straight years before 2024. Then 2025 became the best year Swedemom has ever had, and 2026 is on pace to beat it by about 30%, with the holidays still ahead.

Output grew even faster. Swedemom now lists 3.1x as many new items a month as in 2023, and a new item goes live a median of 5 days after intake, down from 20.

Donations were never the limit; labor was. Since 2024, the first year Swedemom tracked it, the labor cost to put one new item up for sale has fallen by more than half, even after hiring.

Photograph an item. Know what it's worth in seconds.

A worker photographs a donated item, and in seconds the platform identifies it from vision models, live marketplace comps, and Swedemom's own sales history, then says what it sells for and whether to produce, store, donate, or recycle it. For items worth listing, it drafts the whole listing. People decide; the system does the tedious part.

The most tedious jobs in the building, handed to agents.

Unsold listings used to sit still. Now an always-on relisting agent picks candidates and reposts them every day without anyone touching it: 60,000+ listings and $150,000+ in sales since early 2025.

Before anything goes live, someone has to give it a final check, about six minutes of staring per item. Now an agent does most of that staring: 79% of checks in August, work that would take about 500 hours by hand, for about $60 of compute. By September its share reached 82%, and the cost of a check was down 92% from late 2024.

The part we're proudest of: nobody was laid off. The hours moved to judgment work, sourcing better items and handling the cases that genuinely need a human eye. Counting both agents and the per-item speedups, automation now does $36,000+ a month of work that would otherwise be done by hand.

Every dollar treated as an experiment.

Marketplace ads report inflated “attributed sales.” We ignore them and run randomized tests on ad rates, judged on one number: net profit after ad fees. The winning rates doubled net profit per listing in key segments, and ads that lost money were cut to zero.

The same discipline picks the AI: every model earns its place in blind-graded evaluations, and one test alone found a cheaper model that cut AI costs by about $900 a month at the same quality.