Can AI reduce waste?
Only where the bottleneck is identification. Everywhere else it is theatre.
Waste is mostly a sorting problem, and sorting fails for two reasons: people do not know what an item is made of, or they know and the route is inconvenient. Only the first is a machine problem.
Identification is where a model earns its place. A camera can tell a compostable film from a polypropylene one faster than a person can read a symbol they have never been taught. That is a real, measurable reduction in contamination.
What a model cannot do is create a collection round that does not exist, or make a council accept a material its processor rejects. If the answer to 'what is this?' is followed by 'and there is nowhere for it to go', no amount of accuracy helps.
So the honest scope is narrow: reduce contamination by removing guesswork at the bin. Measure it as contamination rate per stream, not as engagement. If the rate does not move, the tool failed, regardless of how good the classification looked.