Drive through seems like a great proving ground. Record every drive through customer / cashier interaction. Match each recording up with the transaction entered into the register. Train a model by having the model "listen" to the recording to predict what the order should look like, then match it to the items on the transaction receipt.

Then, phase 1 of implementation is to use the model in real time by listening to the live conversation at the drive through, predicting what it thinks the order should be, then prompting the cashier to double-check the order to see if the human made a mistake entering the order if the prediction doesn't match.

Phase 2 is human-supervised, where the order taking system interacts directly with the customer to take the order, the human checks the result, and is able to step in / take over if there's a mistake or a special case the order system can't handle.

Phase 3 is "fuck your entry level employment" and no human is monitoring the system.

All 3 phases seem completely doable to me at this point, depending on how much backlash MCD is willing to deal with.