One of the quietly useful things we’ve built at Sparkle is a proprietary voice agent platform that calls real places of business and gathers information the way a friendly intern might — politely, patiently, and without ever needing a lunch break. This week we pointed it at car washes in the DFW area to collecting pricing data.
Why car washes, why pricing, why now
Most car washes don’t publish their prices on their websites. A surprising number don’t really have a website at all, and the ones that do tend to show a photo of the building, a hours-of-operation block, and not much else. If you actually want to know what a wash costs, you either drive by and read the sign at the entrance, or you pick up the phone and ask. That’s it — those are the two options.
For the Car Wash Intelligence Platform we’ve been building, pricing information is one of the single most important layers in the dataset. Reviews tell you sentiment; census data tells you demand; traffic counts tell you visibility. But none of those tell you where an operator sits in the market — whether they’re the cheap option, the premium option, or the one that quietly raised their monthly membership last quarter. With a complete pricing layer, a car wash operator or a PE firm evaluating the space gets something they can actually act on: a live view of what competitors in a given trade area are charging, where the pricing gaps are, and which operators have room to move.
So we’re dialing. Our agent introduces itself, asks for current pricing on the standard tiers, notes any membership promotions, thanks the person on the other end, and hangs up. It takes about ninety seconds. It’s still running as we write this, and the dataset is filling in steadily.
The strange part
We expected to hear a lot of teenagers with headsets, a lot of background noise from vacuum stations, and the occasional owner who wants to know who’s asking and why. We’ve gotten all of that. What we didn’t expect is how often the voice on the other end of the line is, well, not a person at all.
Sometimes it’s obvious — the cadence is just a little too even, the greeting a little too complete, the pauses a little too symmetrical. Sometimes it takes a few exchanges before we notice. And occasionally it’s only in the transcript afterward that we realize our agent spent a minute and a half in a perfectly polite, perfectly useful conversation with another AI. Two voice models, calmly exchanging menu prices, with nobody human in the loop on either side.
Sample calls
Here are two examples from this week’s run — just normal calls from the batch, where both sides happen to be synthetic.
What it means (probably)
The obvious read is that AI voice agents for small and mid-sized businesses have crossed some threshold in the last twelve months. A year ago, calling a small brick and mortar business meant talking to a human close to 100% of the time. Today, in our DFW run, it’s noticeably less than that — and will almost certainly continue to decline.
From a pure data-quality standpoint, this isn’t a problem. The pricing we collect is just as valid whether the voice on the other end belongs to a staff member or to a voice agent running on an operator’s receptionist stack. In some respects the AI-handled calls are actually more reliable: the receptionist models tend to read from a current, structured menu, whereas human staff sometimes have to guess at a tier they don’t handle every day. In our batch so far, the AI-to-AI calls are among the cleanest rows in the dataset.
It does raise a question worth taking seriously. When both endpoints of a phone call are language models, the traditional framing of a “customer call” starts to break down. What’s actually happening is closer to an unstructured, voice-mediated API exchange between two systems — one asking for pricing, the other serving it — with the phone network as the transport layer. That has real implications for how businesses should think about their phone lines, how voice agents should identify themselves, and how data collected this way should be treated downstream. We don’t think the industry has fully worked through any of those questions yet.
What’s next
We’re still running the DFW batch, so the numbers will firm up over the next week or two. When the collection is complete we’ll fold the pricing layer into the Car Wash Intelligence Platform. From there, operators and investors will be able to see, in one place, where each wash sits on price relative to its neighbors — something that today requires a windshield tour or a string of phone calls to piece together.
If nothing else, our agent seems to be enjoying the conversations. We asked. It said they were going well.