01The direct answer
No. A restaurant's calls, its menu, its transcripts, and its orders are not used to train a model that another restaurant's calls run against. That's a flat answer, not a qualified one, and it holds for every restaurant Dohos serves, not just some of them.
Training a general model on that content — or on a caller's content — is a use the governing policy currently rules out by default, alongside a short list of other things treated the same way: building a Voiceprint from a caller's voice, inferring emotion or health from how someone speaks, and using call content for advertising.
02What "training" means here, precisely
Worth being exact about what "training" refers to, because a nearby but different thing does happen with your data, and conflating the two would either overstate the protection or understate what actually happens. Training means adjusting a model's underlying weights — the thing that changes what the model does for every user of it, everywhere, permanently, until it's retrained again. That's the thing that isn't happening with restaurant or caller content.
What does happen is configuration — feeding specific, current information into a call at the moment it happens, the same way giving someone directions doesn't change how they'd give directions to someone else tomorrow. The next section is specific about what that looks like.
04Third-party model providers
Where part of a call is handled by a speech or language model built by another company, that provider's use of the audio and text passing through it is bound by contract, not left to whatever that provider's own consumer product might do by default. Before a model provider is used at all, its account settings for training and improvement, its retention window, its logging, and who at that company could access the content are reviewed and locked to an approved configuration — not left on a vendor's default setting.
What's sent to a model in the first place is also deliberately reduced: a restaurant's full content, complete payment details, and anything legally sensitive stay out of what gets transmitted, cut down to what the specific task in that moment actually needs.
05How the system gets tested without your data
Improving the ordering system's accuracy — handling an unclear address, a menu item that sounds like another one, a caller changing their order mid-sentence — is done against representative and deliberately difficult test cases built from synthetic and approved aggregate data, not by pulling a live restaurant's real calls to make the product better for everyone else.
That evaluation data is governed the same way production data is — it doesn't quietly become training data just because it started as a test case, and it isn't published or shared as a benchmark without its own separate review.
06What would have to change first
Using a restaurant's or a caller's content to train a general model is a use the current policy rules out, not a switch that happens to be off today. If that were ever proposed, it would need its own legal review, its own basis for asking a restaurant or caller to agree to it, and a rewritten version of this page describing it accurately — a product update by itself can't make that change quietly. Nothing described here changes without that review happening first.