Quick update on where Neptune Voice stands and where we are heading.

We switched our voice AI to Grok.

For the past week we were hitting real problems with our previous model (GPT Realtime). The agent would get stuck repeating the same phrase over and over without actually doing anything. It would start actions before the caller finished talking. It could only handle one call at a time. And tool calls would silently stall and never reach the dealership system.

We were patching and patching but it kept fighting us. So we made the call to switch to Grok Voice Think Fast 2.0, and the difference was immediate.

Why Grok is better for us

Grok actually thinks while it talks. That means when you ask it to look something up, it starts the lookup before it even finishes its sentence. No awkward pauses, no canned phrases, no getting stuck in loops.

Independent benchmarks put Grok at the top of speech-to-speech quality right now. It responds in under a second, handles tools better than any other voice model, and just sounds more natural. Shorter sentences, one question at a time, no filler.

Sources: xAI announcement ... Independent benchmarks by Artificial Analysis

What we fixed

• No more repeating. The agent was saying "Got it, thanks" over and over without doing anything. That is gone. Now it actually responds like a person would.

• Tools actually work. When the agent needs to look up a customer, book an appointment, or check availability, it does it smoothly in the same breath. No more silent stalls.

• Multiple calls at once. We can now handle more than one caller simultaneously without calls interfering with each other.

• Natural greetings. The agent opens the conversation based on who you are and where you are calling, instead of playing a prerecorded greeting.

Dealership integration changelog

Full audit of our XTime connection. Every tool was tested live against Genesis of Norman with real data.

Fixed

• Customer lookup by phone. Callers could not be found because of a formatting issue on our side. Fixed.

• Existing appointments. The agent could not see a customer's prior appointments. They were there the whole time, just not being read. Fixed.

• Repair services booking. Over half the service catalog was silently unbookable. The agent would offer them but the booking would fail. Fixed.

• Appointment slot filtering. The agent was offering time slots the dealership would actually reject at booking. Now it only shows slots that will go through.

• Greeter capacity. Slots now respect how many service staff are actually available, not just whether the slot exists.

Added

• Appointment cancellation. Mapped and tested with cleanup.

• Note clearing. Mapped and tested with cleanup.

• Customer deactivation. Mapped and tested with cleanup.

• Faster customer lookup. New shortcut tool that finds customers in a fraction of the data and time.

• Confirmation number lookup. Callers can now read their confirmation number and the agent pulls up the appointment in one step.

8+ hours of red team testing

Before anyone touches this, I wanted to make sure it is bulletproof. I spent over 8 hours running adversarial tests trying to break the agent. Trying to trick it into giving out customer info without proper verification, trying to get it to skip the phone confirmation, trying to get it to book things it should not. It held every single guardrail.

Honestly, working with Grok has been so much more pleasant. The conversations just flow. It sounds like someone actually thinking ahead instead of reading from a script.

I will attach a voice example from the testing convos so you can hear the difference yourself.

What is next

The next milestone we are working toward is a playground soft opening so Jon can start his work before we open it up to everyone. That is the hope, not a promise. The playground lets you call the voice agent yourself, try it out, and see how it handles real scenarios. It runs on the live Genesis of Norman system with your own test profile, so you are seeing real data and real integration, not a mock demo.

We are keeping the old model (GPT Realtime) available alongside Grok in the playground, so you can switch between them and hear the difference for yourselves.

A larger update is coming soon with more detail as we get closer to that milestone.

Quick summary

• Switched to Grok Voice Think Fast 2.0 (GPT Realtime still available for comparison)

• Repeating phrases, single-call limit, and tool stalling: all fixed

• Full dealership integration audited and repaired

• 8+ hours of adversarial testing, every guardrail held

• Next milestone: playground access for Jon, then staff-wide

• Bigger update coming soon