How AI-assisted acceptance rules actually work

"Just use AI to check donations" sounds simple until you notice that no two op shops accept the same things. One store can't take electrical goods without a tester on site. Another has a furniture buyer three days a week and none the rest. A third has plenty of shelf space for kids' clothes but is drowning in adult formalwear. None of that is a fixed, universal rulebook — it's specific to the store, and it changes.

So the AI's job was never "decide what's acceptable." It's narrower and more useful than that: identify what's in a photo, and check it against your rules.

What the model actually does

When a donor photographs an item, three things happen in sequence:

1. Identification. The model looks at the photo and works out what the item is, roughly what condition it's in, and what category it falls into — using the same shared category list every store draws from, so "kids' bike" means the same thing everywhere.
2. Rule matching. That category is checked against the receiving store's own acceptance rules — the ones a store manager has actually configured, including any per-store overrides (no furniture at one location, no electrical without testing at another).
3. A confidence-aware decision. If the match is clear, the donor sees an accept or reject, with the store's own reason attached. If the model isn't confident — a worn item where condition is genuinely borderline, say — the result is "conditional": bring it in, staff decide in person. Nothing gets a hard rejection purely because the model wasn't sure.

Where store managers come in

None of this is meant to run unsupervised forever. Every decision a donor sees is recorded, and store managers can correct any of them — wrong category, wrong condition call, wrong price estimate. Those corrections aren't just a complaint box; they're the training signal that improves accuracy for that specific store over time. A charity with unusual rules (say, no glassware at all) doesn't need to wait for someone to hand-code an exception — a manager corrects a few decisions and the pattern holds going forward.

The result is closer to "a volunteer who's seen every donation this store has ever received" than to a generic image classifier. It's tuned to one store's shelves, one store's rules, and it keeps learning from the people who actually run that store.