> 1. Ask your phone to buy {name of item}

Outside theory, I don't see that working very well in the real world currently.

On amazon fresh website, there are currently

229 results for "green apples"

8 results for "granny smith apples"

2 results for "1 granny smith apples"

Same with ham, eggs, or any other products really. Sure, the user could be more specific, but then it gets a lot more work and problematic than simply scan the damn barcode for this "4-in-1 pack of organic spanish strawberry yoghurt march special offer".

First - what's simple about "simply scan the damn barcode"? Simplicity is telling your phone that you ran out of pasta sauce, not stocking a barcode scanner in your pantry. The hardware may be sleek, but amazon's solution most certainly isn't. For a company with amazon's resources, I'd expect more.

Next - presuming that you bought something from Amazon, why do you need the barcode? If I recently bought and/or regularly buy granny smith apples from Fresh, it's quite easy to figure out the most probable buying targets for queries of "apples." Further, if I regularly buy 5 different types of granola bars, now I can use one query and a checkbox instead of scanning each box individually. (Sidebar: how do you use Dash to scan produce that you've already consumed? Do you have to keep the containers around?)

In the big scheme of things, figuring out which product you probably want to buy based on your buying history (and that of others) is a trivial problem compared to the other problems that had to be solved for "always listening" and such. Amazon's job ends up being the easiest in that stack.

You're not considering that a lot of people might just prefer to be explicit and specific about purchases. If I see I've run out of pasta sauce, I might want to scan the empty one to reoder it, because it's precise. I definitely wouldn't want to risk vagueness and inaccuracy just for the privilege of using an abstraction like saying "I'm out of pasta sauce" into a phone. What would that achieve anyway? Outsourcing the bit of your brain that says "we need more pasta sauce", so you've only got to worry about thinking "oh there's no pasta sauce"..? Scanning a barcode is simple and precise.

  risk vagueness and inaccuracy

You're scanning a barcode of an item you recently purchased from amazon. Where is the vagueness in querying for that item by type?

In most cases, your pantry has 1 type of each item, which means you've got a canonical mapping from "pasta sauce" to an item in your fridge. That also means you've got a very high probability of mapping the recent purchase of pasta sauce to the query of "pasta sauce." Where is the vagueness?

In the worst case, you've bought two or three different types of pasta sauce, and you'd like to replace just one. Let's be clear - this is most definitely the exception to the rule. In this case, your phone gives you a list of options to choose from -- the list of your N most recently purchased pasta sauces (backfilled by prior orders or popular choices).

"Phone, order green apples"

"Same type as last week?"

"No, the ones I bought before that"

"OK, ordered XXX apples"

Steps 2 and 3 may be better in the hands of the consumer -- in other words, while you can get away with probabilistic guesses for an initial query, you'd be better off leaving the consumer with an interface they can pick & choose from instead of trying to pull a Siri.

Where is the barcode on a granny smith apple?