I can't share the code, but we just implemented this. We used an email plugin as a stage one script written in SS1, which pushes to a task queue for a ss2 script which sends the job to open ai as a batch job. Then we have another script that runs hourly to poll for openai responses. So far our best results have been with GPT-5 (main model, not mini/nano), and we spend 5cents/order processing it.