Sub-Processors
Last updated: August 25, 2026 · Compiled from a review of deployed code and deployment configuration.
A sub-processor is a company we use to store, transmit or process data on our behalf in order to run Vetra. This page lists them, what each one does, what data reaches it, and what country it is in. It covers the Vetra product; it does not cover the tools a subscribing business chooses to connect on its own.
This list is referenced by our Privacy Policy and forms part of it. We will update it when a provider is added or removed. If you want to be told about changes, email legal@usevetra.com and ask to be added to the notice list.
Where a provider is outside the United States, it is marked in red. See Privacy Policy Section 7A.
1. Infrastructure, hosting and storage
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| Supabase | Primary database, authentication, file storage | Substantially all business and end-user records we hold | United States |
| DigitalOcean | Hosts our application servers | All data in transit and in memory while the service runs | United States |
| Amazon Web Services (S3) | Image and media storage | Receipt and tip-slip photographs; recorded call audio | United States |
| Amazon Web Services (Textract) | Reads the text out of a photograph | Receipt and tip-slip photographs | United States |
| Cloudflare | DNS, content delivery, and hosting for the customer dashboard | Request metadata for traffic to our sites | United States |
| GitHub | Hosts our public marketing site; source control | Visitor IP and request metadata for the marketing site | United States |
| Sentry | Error and performance monitoring | Error reports and traces. Configured to suppress personal data by default, with additional scrubbing applied | United States |
2. Communications
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| Telnyx | SMS delivery, telephone numbers, call transport | Phone numbers, message content, call routing data | United States |
| Resend | Transactional email (receipts, alerts, notifications) | Email addresses and message content | United States |
| Customer.io | Lifecycle and promotional email to business owners | Business-owner email addresses and engagement data | United States (an EU region exists and is not currently in use) |
| Google Workspace | Our own company email | Mail we send and receive, and its recipients | United States |
3. Voice — the AI phone service
If you telephone a business that has enabled our AI phone service, these providers handle your call. See Privacy Policy Section 1, “Callers”.
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| LiveKit | Carries the live call and runs the voice agent; writes recordings to our own storage | Live call audio and room metadata. LiveKit states it does not retain this or train on it | United States |
| Deepgram | Converts your speech to text, and generates the agent’s speech | Raw caller audio | United States |
| OpenRouter | Routes the conversation turns, and the post-call summary/extraction, to a language model | What you say, as text, turn by turn, and the full call transcript for post-call lead extraction and (where enabled) grading, insights and intelligence | United States (routes onward — see section 4) |
| Vapi | Our previous voice platform. No live telephone number is served by it, but it remains connected to our systems and is not yet removed | Historic call audio and transcripts | United States |
Call audio and transcripts are not used to train models
Vetra does not use call audio or call transcripts to train or improve any model, and does not pool one business’s calls into analytics for another. That part is our own commitment. The part that is harder — and the part this section is really about — is what our providers are able to do with the same audio, because a promise we make about someone else’s systems is only worth what their systems actually enforce.
So where a provider offers a technical no-training control, we set it on the request itself rather than relying on a contract term:
- Deepgram runs a model-improvement programme that is opt-out, not opt-in — its default is to include a share of submitted audio. We send the opt-out parameter on every speech-to-text and text-to-speech request the phone service makes. Deepgram states that data from opted-out requests is kept only for as long as it takes to process the request.
- OpenRouter lets a request specify that it may only be routed to model providers that do not train on submitted content and do not retain it. Every request that carries call audio or a call transcript specifies both.
- DeepSeek offers no such control — no per-request setting and no account setting (see Section 4). Call audio and call transcripts are therefore not sent to DeepSeek at all. This is enforced in code, not by policy: the phone service refuses to process a call rather than route it there, which is a deliberate choice to lose the feature instead of the guarantee. Until August 2026 a documented fallback would have sent call transcripts to DeepSeek if OpenRouter were unavailable. That fallback has been removed.
- LiveKit, which carries the call itself, commits in its terms not to train or fine-tune on the content it carries and requires the same of the providers it uses. There is no control to set because it is the default.
The honest limit: setting a provider’s no-training control is the strongest technical step available to us, and it is not the same as proof. We can show you the setting in our code and quote the provider’s published behaviour; we cannot audit their systems. Where a provider gives no warranty at all — OpenRouter expressly does not, for the providers it routes to — that remains true no matter what we set on the request, which is why Section 4 names those providers individually rather than stopping at the gateway.
4. Artificial intelligence
Most of our AI work is sent through OpenRouter, which forwards the request to whichever model provider is selected. OpenRouter states in its own terms that it makes no warranty about a downstream provider’s data handling, retention or training practices, so we name the providers we route to rather than implying that routing through a gateway settles the question.
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| OpenRouter | Gateway that routes AI requests to model providers | Chat messages, call turns, text read out of photographs, extracted business data, and post-call lead extraction and (where enabled) call grading, insights and intelligence | United States |
| DeepSeek | Language model used for internal drafting assistance, some recipe and menu processing, and screening of submissions to our careers page. It is no longer wired to any part of the phone service — the fallback that could once have sent call transcripts here was removed in August 2026 (see Section 3) | Text drafted in our internal tools; images submitted to our careers page; menu item names. No call audio and no call transcripts. | China |
| Google (Gemini / AI Studio) | Reads photographs — menus, invoices, and photographs of contact lists a business uploads | Photographs that may contain customer names, phone numbers and vehicle details | United States |
| Anthropic | Language model reached through OpenRouter for chat | Chat conversation text | United States |
| Cerebras | Runs a model, reached through OpenRouter, for extracting vehicle labour data from public sources | Publicly sourced repair-pricing text | United States |
| Voyage AI | Turns text into embeddings so it can be searched | Knowledge-base content and search queries | United States |
| Tavily | Web search used to gather public reference data | Search queries | United States |
| Jina AI | Fallback web search when the above returns nothing | Search queries generated by our internal tooling | Germany / China |
Training on submitted content — what we know
We say this plainly because our previous policy said the opposite. We asked, for each provider, whether its published terms permit it to train on content we submit:
- Anthropic and OpenAI commit in their business terms not to train on customer content.
- LiveKit commits not to retain or train on call data, and requires the same of its own providers.
- Google commits not to train on content submitted through its paid tier. Its unpaid tier expressly does train, and permits human review.
- OpenRouter expressly disclaims any warranty about what a downstream provider does. It does let a request restrict routing to providers that neither train on nor retain submitted content, and we set that restriction on everything carrying call audio or a call transcript (Section 3) — but the disclaimer above still stands, and setting a routing preference is not a warranty we can offer you either.
- DeepSeek’s privacy policy states that data is used to train and improve its models, and that it is processed and stored in China. It offers no per-request or account-level opt-out; its only route is a written request by email. Call audio and call transcripts are therefore not sent to it at all.
- Voyage AI’s default terms grant it a licence to use submitted content to train its models unless an account-level opt-out is enabled.
- Deepgram reserves the right in its terms to use submitted content to improve and train its models, and its model-improvement programme is opt-out. We send that opt-out on every phone-service request (Section 3).
We are working through these and will update this page as we change providers or settings. We would rather show you an uncomfortable list than repeat a promise we cannot keep.
5. Payments, point of sale and business operations
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| Stripe | Subscription billing, card and ACH payments, payouts | Billing identity, payment method, bank details | United States |
| Toast, Square, Clover | Point-of-sale systems a business connects, where it chooses to | Check, order and payment records for that business | United States |
| Card processors — Global Payments, Elavon, NMI, Heartland, Datacap | Where a bar uses our tip reconciliation, the processor that adjusts the tip | Transaction identifiers and tip amounts. We never send a full card number | United States |
| Mercury | Our company banking and contractor payouts | Our own banking data and contractor tax data — not customer data | United States |
| Shippo | Shipping labels when equipment is returned | Business names and shipping addresses | United States |
6. Website analytics
These run on our own public marketing pages only. They are not part of the product and do not run inside a business’s dashboard or on a business’s own website.
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| Google Analytics 4 | Measures page views on our marketing pages | Page views, referrer, a Google analytics identifier | United States |
| Microsoft Clarity | Records a replay of the visit on one page (our auto repair page) | Clicks, scrolling and page interactions. Microsoft acts as an independent controller and may use this for its own purposes, including advertising | United States |
7. Providers used for our own sales and research
These receive information about businesses we research as prospects, not about a subscribing business’s customers. They are listed for completeness.
| Provider | What it does | What reaches it | Country |
|---|---|---|---|
| HERE Technologies | Business search and mapping | Business names and geography | Netherlands / Germany |
| OpenStreetMap (Nominatim) | Turns an address into coordinates | Business addresses | United Kingdom / Germany |
| Yelp | Public business listings | Business lookups | United States |
8. What this list does not yet cover
Two gaps, stated rather than hidden:
- The website chat widget. Chat messages from the widget are handed to a workflow that runs on our own server, which then calls a language model. Which model that is, is configured inside that workflow rather than in our application code, so we have not named it above. We are resolving this and will name the provider here once we have confirmed it from the workflow itself rather than guessing.
- Providers a business connects itself. If a subscribing business connects its own accounts — a point-of-sale system, a review platform — that relationship is between the business and that provider.
9. Questions
Email legal@usevetra.com. If you are asking about a recorded telephone call, tell us which business you called and roughly when, so we can find the record.