The saving has a defined scope
OpenAI’s September 23 Ringg customer story describes multilingual agents spanning voice, chat, WhatsApp and web. It reports an approximately 90% model-cost reduction for selected workloads moved from GPT-4.1 to GPT-5.6 Luna, alongside a platform-level claim of up to 65% routine-request resolution without a human agent.
The same account says GPT-4.1 still handles most real-time traffic and describes different models for different tasks. It should not be read as a complete platform migration or a 90% reduction in the total cost of customer service. These are vendor-reported results, not independently verified outcomes.
A return request is more than a conversation
For an ecommerce team, our recommended trial would start with a narrowly defined request, such as locating the published return policy. Separate that information task from actions that alter an order, promise a refund or disclose account details. A fluent answer is not sufficient evidence that the agent performed the right action.
Use test cases that include an unclear request, an unavailable item and a customer switching languages. Ask the reviewer to check the final state as well as the conversation. Did the request reach the right person? Was any promise made before the relevant information was available? These are proposed ecommerce tests, not claims about a particular Ringg deployment.
Measure the handoff to a person
A partial automation can still help if it transfers the customer’s context cleanly. It can also make service worse if the person taking over has to repeat every question. Include that handoff in the trial and measure the customer’s time to a correct resolution.
Review permissions and customer-data handling separately from the model comparison. Keep actions limited to those the team has explicitly approved, and make exceptions visible to a responsible operator. A case study is not a complete specification for account access or retention.
The commercial question is what happens after routine cases have been automated. Track repeat contacts, corrections and human follow-up alongside model charges. That gives a brand a more useful basis for adoption than a savings percentage detached from the requests its customers actually make.