This article is based on a Q&A with VoiceBit Co-Founder and CTO Jay Patel and Greg Kihlström for The Agile Brand Guide’s Expert Mode series.
For most of the last decade, “owning the customer” has meant something pretty specific for restaurant and retail marketers. Get people off DoorDash and Uber Eats and onto your own site or app, where you keep the margin and the data. We’ve built loyalty programs and app-download campaigns around that idea, and handed out a lot of direct-order discount codes along the way.
That definition gets shaky once the order starts and finishes inside an AI assistant. If a customer asks ChatGPT for dinner and pays without ever opening your app, it’s fair to ask what “owned” means now. Jay Patel, co-founder and CTO of VoiceBit, argues that ownership was never about the interface. His company builds AI phone and web ordering for independent restaurants, along with an ordering app inside ChatGPT, so he has an obvious stake in the answer. His claim is that the customer belongs to whoever holds the record behind the order, and that you can test whether that’s you.
The Export Test
Asked whether moving discovery and checkout into ChatGPT just swaps one intermediary for another, Patel separated where the conversation happens from where the transaction lives. In his framing, ChatGPT is “the front door,” not the house. The order, the payment and the customer record can still sit in the restaurant’s own systems. Whether they actually do is something a marketing leader can check.
The test for a CMO is simple. Export your customers. If you can pull a list of names, contact details, and order history tomorrow and use it in your own email or SMS campaigns, you own the relationship. If the channel turns off and your customers vanish with it, you never did. The other thing to watch is pricing. If a channel can change your prices, your fees, or your position without asking, it’s a landlord, not a doorway.
It’s a refreshingly blunt test, and most of us could run it this week. List every channel that takes orders today, from the marketplace and the call center to any new assistant integration, and ask each vendor what a full export of customer-level first-party data looks like. Not a dashboard, but a file you can load into your own CRM. Channels that can’t produce one are acquisition you rent, and they should be measured that way. The pricing half of the test matters just as much. A partner that can reprice your menu without asking belongs on the risk register.
Your Menu Is a Dataset
Patel described two very different paths into an assistant’s answer. In one, the customer has an ordering app connected, and ChatGPT hands the request to it without being told to. In the other, the assistant goes looking on its own.
What decides whether a restaurant shows up is menu data. Gluten-free has to be a tag on the item, not a note in a PDF. Structured items can be shown; unstructured ones don’t exist. … Without an ordering app installed, ChatGPT recommends on its own from the web, reviews, and whatever structured data it finds. A marketing team controls far less there: an accurate website, consistent menus, reviews that mention what people ask about. … Treat your menu as a dataset, not a document.
In most organizations the menu, or the product catalog, belongs to operations. Patel’s point is that it’s now a visibility asset, and “unstructured ones don’t exist” is a line worth sending to whoever owns your product data. What he’s describing is basically product information management applied to a restaurant: dietary attributes, portions, modifiers and prices kept accurate in one place. When we asked what a multi-location brand should fund next quarter, he gave the same answer in budget form. Fund “one source of truth for every item,” feeding the website, the phone agent, the POS and ChatGPT alike. His observation that most chains are running four versions of their menu, “and none of them is right,” will sound uncomfortably familiar to anyone who’s audited one.
The organic path calls for something closer to answer engine optimization: an accurate site, consistent listings, and reviews that actually mention the things customers ask about. Neither path shows up in a media plan. That’s probably why both get underfunded.
Apply the Test to the Vendor, Too
VoiceBit’s ChatGPT app only surfaces restaurants that use VoiceBit, so it puts a network of its own between the restaurant and its customer. DoorDash and Uber Eats are building into ChatGPT too. We put that to Patel directly, using his own line from an earlier interview about not depending on a vendor you can’t leave.
The question is fair, and I’d apply the test to us first. A vendor you can leave gives you three things: your data on the way out, no lock-in contract, and a price it can’t change under you. Restaurants can export their customers and order history. … And our app surfaces restaurants that use VoiceBit, so yes, there’s a network. The difference from a marketplace is that we don’t take a commission on the order, we don’t set the restaurant’s prices, and we don’t rank one restaurant against another to sell placement.
That’s a fair answer. It’s also a sales argument, and buyers should weigh it as one. It still makes a usable checklist for any vendor in this category, VoiceBit included: get export capability in writing and ask what happens to pricing at renewal. And “no commission” isn’t the same as free. Restaurants still pay for the software, so the comparison worth running is total cost per order across every channel, not the headline fee on one of them.
On ChatGPT itself, Patel was candid. No restaurant can leave ChatGPT any more than it can leave Google, he said, which is exactly why it shouldn’t be the only way in. He expects several: the phone, the website, the chat agent and whatever comes next. The goal is for all of them to feed one customer record the restaurant controls, so that if ChatGPT changes its rules, “the restaurant should lose a channel, not its customers.”
Channel Shift Isn’t Growth
Patel’s example came from New York Pizza & Pasta in Pleasant Hill, California, where VoiceBit’s phone agent now recognizes returning callers and offers them their last order. The feature exists, he said, because a regular once called and asked for “my usual,” and the agent had no idea what that meant. According to Patel, regulars now mostly just confirm details, and the owner came away with a sharper picture of how much of his business a fairly small group of loyal customers accounts for.
That’s exactly where attribution gets messy. A loyal customer who moves from the phone to an assistant isn’t a new customer, and Patel was clear about how to tell the difference.
On attribution, don’t trust the channel’s own report. Match ChatGPT orders against your customer record. A new phone number or email that has never ordered through your site, your app, or your phone line is new demand. A customer who ordered from you three weeks ago is a channel shift, which is still worth something if the margin is better, but it’s not growth. Run the match for a quarter, and you’ll have an honest number.
For anyone heading into a budget review, this is the most useful idea in the exchange. Every new channel arrives with a dashboard claiming the revenue it “drove.” Patel’s method is plain incrementality work, and it doesn’t need the vendor’s cooperation. Split assistant orders into customers you’ve never seen and existing customers who changed channels, and track that split for 90 days. Then compare average order value and order timing against your app and site. If assistant orders arrive at hours or basket sizes your own channels never saw, Patel argues, that’s new behavior. If they look just like your app orders, you’ve moved revenue around, maybe at a better margin. You’d want to know which before the renewal conversation, not after.
We’ve spent years treating owned channels as the goal. Patel’s argument is that the asset was always the customer record behind whatever channel took the order. If he’s right, the job right now is less about picking an AI partner and more about unglamorous groundwork. Run the export test on every channel that takes orders. Give your product data an owner who treats it as a marketing asset. And set up the match-back before the first assistant dashboard lands on your desk.

