Starlink has quietly replaced its customer support hotline with Grok's new VoiceFast 1.0. No menu tree. No "press 1 for billing." Just a voice that picks up, takes your details, pulls your account, and walks you through your options. I called in. The entire four-minute conversation is below.
So I read this morning that Starlink had quietly replaced its phone support hotline with Grok's new VoiceFast 1.0. No menu tree, no "press 1 for billing"... just an AI picking up the call and handling the whole thing. Anyone who knows me knows I've been deep in voice agents for the last three years, building, testing, breaking, watching the field move every few weeks. So the second I read it, I thought... right. I have to call this thing.
Lucky for me, I genuinely have a Starlink account. Cancelled the service about a year ago, but the account is still there. Real customer, real reason to call. I picked up the phone, hit record, and dialled in to see what it could actually do.
What follows is the full four minutes. Unedited.
What you're about to hear
The voice that picks up is warm, conversational, slightly American... no awkward pauses, no scripted-bot tonality, no IVR. She greets me, asks who she has the pleasure of speaking with, runs me through a clean verification flow (cell number, email, code from inbox), pulls up my account, walks me through three current plan options, takes a follow-up about data caps... and at the end, I try to bait her into cracking a joke about Elon Musk.
I won't spoil the answer. Watch what she does.
The light grey waveform is me. The red waveform is her. They light up live, in time with the voice. So you can see who's actually doing what.
What I was testing for
Honestly? The same things I look for whenever I'm pressure-testing a voice agent.
Can it handle interruption? Can it recover when I trail off mid-sentence? Does it actually hear the email correctly the first time, or does it ask me to repeat? Does it stay warm under a slightly hostile question, or does it go robotic? Can it gracefully decline something it shouldn't answer without sounding like a corporate firewall? Does it know when to stop talking?
The boring boxes most production bots fail by minute two.
I asked about pricing. Asked about data caps. Slipped in a slightly off-piste question about what happens when I want to reactivate. Tried to break the tone with the Elon thing right at the end. Nothing got it off-balance. It rolled with me, stayed warm, recovered cleanly, never went into that hollow customer-service register that scripted bots default to inside twenty seconds. It was, frankly, fucking impressive.
Why this matters
Look... I've been writing about this for a while now. The shift from AI-as-tool to AI-as-collaborator. The idea that the dashboards aren't the product anymore. Most of the examples I reach for are creative ones, because that's the world I come from.
This one isn't creative. This is customer support on the front line of a critical infrastructure company, talking to a paying customer about a real account, in real time, on the phone. And it's good. Not better-than-human good... let's be honest, the warmth and the humour are comparable, not superior, the timing is comparable, not superior. But it's miles better than waiting fifteen minutes for an operator. Miles better than the scripted bots we've been suffering through for a decade. No menu tree. No transfer queue. No "let me put you on a brief hold." And on the things that do beat a human... it's faster, more accurate with details, more consistent, infinitely more patient.
The bar just moved.
It moved on a Tuesday afternoon while nobody was looking. Which is kind of how all the big shifts seem to happen now.
I'm not saying every support function should swap to AI tomorrow. I'm not saying this scales perfectly to every domain. I'm not saying the dystopian version of this isn't real. But anyone who's still going around saying things like "AI can't do nuanced human conversation yet" needs to listen to four minutes of this and update their priors. Because that take just got expensive.
The bit I keep coming back to
The line at the end. When I asked her about Elon. She didn't dodge clumsily. She didn't refuse stiffly. She understood the bait, declined to take it, and then... when I expressed mild disappointment that she didn't crack a joke about him... she acknowledged my disappointment, named it back to me, and stuck to her lane. Politely. With a tiny bit of warmth. "Fair enough, Simon. I can see why you'd expect a quip there. But I'll stick to the Internet side of things."
That is not autocomplete. I'm sorry, but it just isn't.
Anyway... that's where I'm sitting on this one. Have a listen. Tell me I'm wrong. :)