Sub-processors
Last updated: October 2026
To run Verso we rely on the third parties below. Each receives only the data needed for its function. This list may change as the product evolves; we will keep it current.
| Sub-processor | Purpose | Data processed | Region |
|---|---|---|---|
| Voyage AI | Embeddings that make your knowledge base searchable | Text from your knowledge documents and queries | United States |
| OpenAI (optional) | Alternative embeddings provider, if configured | Text from your knowledge documents and queries | United States |
| ElevenLabs | Speech to text for dictation, and the live voice call (speech recognition, turn-taking, and the spoken reply) | The audio you dictate or speak during a call, the resulting transcript, and the reply spoken back to you, which may quote your knowledge base. Audio is streamed for processing and is not stored by Verso. Dictation also sends a short list of recurring technical terms taken from your own uploaded documents (for example product names or acronyms), so that speech recognition spells the words you actually use correctly. | United States |
| Soniox | Real-time speech to text for dictation, so words appear on screen as you speak | The microphone audio you dictate, and up to 25 recurring technical terms taken from your own uploaded documents (the same list ElevenLabs receives), so speech recognition spells the words you actually use correctly. Audio and transcripts are not stored and not used for training. | United States |
| fal.ai (Features & Labels, Inc.) | Generating an image or a short video, when you ask for one in Studio | The prompt you type for that picture or clip, and the shape and length you choose. Nothing about your knowledge base, your drafts or your account is sent, and nothing at all is sent unless you press Generate. | United States |
| Stripe | Payments and subscription billing | Billing identifiers and payment details (handled by Stripe) | United States |
| Resend | Transactional email (for example, team invitations) | Recipient email address and message content | United States |
| Sentry | Error monitoring, so we can detect and fix faults quickly | Diagnostic error reports and technical request context (no knowledge-base content) | European Union |
| Publishing a post when you choose to publish it, and importing your name and profile photo if you sign in with LinkedIn | The post content you publish, and the profile details LinkedIn returns. Nothing is sent to LinkedIn unless you connect the account and press publish. | United States | |
| Notion | Reading the pages you choose to connect, so Verso can ground your posts in them | A token you create yourself, and the search requests Verso makes with it. Notion sends us the pages you shared with that token, and only those. Nothing at all is sent unless you connect it. | United States |
| Readwise | Reading the highlights you choose to connect, so Verso can ground your posts in them | A token you create yourself, and the requests Verso makes with it. Readwise sends us your highlights back. Nothing at all is sent unless you connect it. | United States |
| Google Drive | Reading the Docs, Sheets, Slides, PDFs and Office files you pick in Google’s own file picker, so Verso can ground your posts in them. Only the drive.file scope is ever asked for, so Verso can never see a file you did not choose there | An OAuth connection you approve on Google’s own consent screen, and the content of the files you pick there. Nothing at all is sent unless you connect it and choose files | United States |
| Sign-in providers (Google, Microsoft, Apple) | Signing you in, only if you choose social sign-in instead of an email and password | Your name, email address and account identifier at that provider | United States |
| Vercel | Hosting and running the application, plus cookieless Web Analytics (page views and referrers, no cookies and no cross-site tracking) | Every request to the site, and the technical metadata that comes with it such as IP address and browser | United States / EU |
| Neon (Postgres) | The database that stores your account and your content | All account and content data; sensitive fields, including knowledge-base content and connected-account tokens, are encrypted before they reach the database | United Kingdom (London) |
| OpenRouter | Routes every AI writing task to the open-weight models below: drafts on every plan (a stronger model on Pro and Team), voice call replies, the questions and opening lines before a draft, ideas, grammar, your voice profile, memory, chat summaries, guardrail checks, reading your uploaded documents and pictures, and sorting contact and feedback messages | Your prompts and the relevant knowledge-base context for a request, the post or draft being worked on, what you say in a voice call, your uploaded documents and pictures, your chats and the rules you set, or the text of a contact or feedback message | United States |
| DigitalOcean | Runs the model that writes Free and Starter posts, and the first model tried for most other AI tasks (the questions and opening lines before a draft, ideas, your voice profile, memory, chat summaries and sorting contact messages). It is also a later model tried for guardrail checks and for reading your uploaded documents, after Novita | The same prompts, context and documents, passed on by OpenRouter | United States |
| DeepInfra | Runs the models that write Pro and Team posts, read pictures, answer voice calls and fix grammar, and stands in for DigitalOcean on everything DigitalOcean runs when it is busy. It is also a later model tried for guardrail checks and for reading your uploaded documents, after Novita and DigitalOcean | The same prompts, context, documents and pictures, passed on by OpenRouter | United States |
| Novita | Runs the backup model for Pro and Team posts, and stands in for DeepInfra on the Pro and Team writer, picture reading, voice calls and the grammar fix when DeepInfra is busy. Also runs the one fallback model DeepInfra does not host, and is the first model tried on every guardrail check (including the one-shot check that screens a batch of draft suggestions before they reach you) and for reading your uploaded documents. A guardrail check can carry the rules you set, which may be your employer’s policy | The same prompts, context, documents and pictures, passed on by OpenRouter, whenever one of those models is used, and the rules you set plus the post or candidates being checked, for a guardrail check. Novita’s own privacy policy states that personal information is not used to train its models. | United States (on a transfer basis, per Novita’s own privacy policy; Novita’s terms are governed by Delaware law) |
| Exa | Runs the live web search behind “Live web research”, and only on a draft you switch it on for. OpenRouter asks Exa on our behalf, hands the results to the model writing that draft, and the sources are listed under the draft so you can check them | The search text, which is the request you typed for that draft. Not your knowledge base, not your documents, not your account details, and nothing at all unless you switch live web research on. A search is not covered by the zero-retention setting the writing models carry, and Exa’s own privacy policy says query data is used to improve its products, including by training its models. | United States |
| Google (Google AI Studio and Vertex AI) | Draws the pictures Verso makes for your posts and carousels, and tidies the words you dictate (grammar, filler words, lists) once you stop speaking, both on OpenRouter’s instruction. For dictation it is sent only the words you dictate, to tidy them, and those are often the post you are writing. It is not sent your knowledge base or your uploaded files | A picture brief: a short prompt, written by Verso from the post you are illustrating, describing the scene to draw; and, when you dictate, the words that were heard | United States |
Every request to OpenRouter carries a zero-data-retention setting, and our OpenRouter account itself has training and logging switched off, so none of the models above stores a prompt, trains on it, or keeps it once it has answered. When one model cannot be reached, the next one of the same kind is tried, always from the hosts on this list. The hosts are never the model makers’ own China-hosted APIs (for example Alibaba, Z.AI or Moonshot); Verso pins each model to a Western host instead.
Live web research is the one exception to that, which is why Exa is on the list above. When you switch it on for a draft, the request you typed goes through OpenRouter to Exa, a search engine rather than a model host, and the zero-retention setting does not cover a search. Exa’s own privacy policy says query data is used to improve its products, including by training its models. So only the request you typed is ever used as the search: your knowledge base, your documents and your account details stay out of it, and with live web research off nothing is sent to Exa at all.
Embeddings provider depends on configuration; a self-hosted local model can be used instead, in which case no embedding data leaves our infrastructure.
One further recipient is worth naming even though it is not our sub-processor. When Soniox live dictation is unavailable, we fall back to your browser’s built-in speech recognition to show words as you speak. In most browsers that feature is not on-device: Chrome sends the audio to Google, and Edge sends it to Microsoft, under that vendor’s own privacy policy rather than a contract with us. We never receive that audio ourselves. If you would rather your speech did not reach your browser vendor, type instead of dictating.
For more on how we handle your data, see our Privacy Policy.