AI Transparency
Local review draft — Last updated: October 18, 2026 (UTC) — not approved for publication
VectorOS uses deterministic rules for customer workspace planning and imports. Those paths do not send your workspace content to a third-party AI model. Separate founder-only Growth tools can use configured models and search providers; browser voice dictation has its own processing boundary. Optional private meeting transcription uses a separate local speech model. Each is described below.
The short version
VectorOS uses deterministic, explainable intelligence rather than a third-party AI model to organize your work and recommend what deserves attention next. The same inputs produce the same ranking, every time, and every recommendation can show the reasons that produced it.
Customer workspace planning and imports use the deterministic engine. Optional provider processing in separate internal Growth tools is described below.
How recommendations are produced
VectorOS builds a connected picture of the work you have given it — tasks, projects, goals, opportunities, people, and the relationships between them — and scores it with transparent rules: what is due, what is overdue, what advances a goal or an open opportunity, what has gone quiet, and how much of your day is already committed.
Reading documents and captures into that picture — recognizing a task, a date, a person, a company — is also rule-based. It is pattern matching and structured parsing over what you actually wrote, not a model being asked to guess. That is why the importer shows you its proposals for review instead of writing them straight into your workspace: rules are predictable, but they are not infallible, and you are the one who confirms.
Customer workspace processing
In the deterministic planning and document-import path, VectorOS does not send your workspace data — your tasks, projects, goals, opportunities, people, notes, or the documents you import — to OpenAI, Anthropic, Google, or any other AI model. It is not used to train anyone’s model, including ours. Optional private meeting transcription is a separate local-model path described below.
That statement is about workspace data and about VectorOS. It is deliberately not phrased as “never, for anything, at any time,” because dictation runs through your browser, and what your browser does with your speech is its decision and not ours. See Voice dictation.
This applies to the customer workspace planning and import path. Separate internal Growth tools can use configured search services and Anthropic for selected research and content inputs. They do not turn customer workspace priorities or imports into model-backed processing. A prepared action still needs review, and an execution record does not by itself establish a verified outcome.
Internal Growth media production can also run local models for narration and speech transcription. Local model processing is distinct from an external provider call and from the rules that rank customer work. Availability depends on the configured worker; a generated asset still needs its own review before publication.
Optional response playback can also use Kokoro on a configured local server worker for an entitled customer workspace. Only the selected response text is sent to that worker to synthesize speech; this is separate from planning, imports and browser dictation. If the local worker is unavailable, VectorOS reports that limitation rather than silently sending the text to an external voice provider.
Voice dictation
For dictation into text fields, VectorOS uses the speech-recognition capability built into your browser. This is a browser feature, not VectorOS-hosted AI. The audio does not pass through a VectorOS server, and we do not receive or store a recording.
What your browser does with it is a separate question, and the answer is not ours to give. Depending on which browser you use, your speech may be sent to the browser provider and processed on its servers under its own terms and privacy policy. Chrome, for example, implements this by sending audio to Google. That happens because of the browser you chose, not because of VectorOS, and it would happen on any site using the same capability — but you are entitled to know it, so we say it here rather than leave it inside a sentence claiming nothing is ever sent anywhere.
If you would rather it did not happen at all, type instead of dictating: every voice entry point in VectorOS has a text equivalent, and nothing in the product requires speech.
Optional private meeting transcription
Meeting recording/import is separate from browser voice dictation and deterministic planning. Where enabled, it first saves recoverable audio in this browser’s IndexedDB. You acknowledge participant agreement, review an upload reservation, then separately upload and confirm the transcription queue. Uploaded audio leaves the browser for private application Storage.
A separately configured local worker can decode that audio and run the existing offline faster-whisper tiny.en speech-recognition model. This is model-based transcription, not the deterministic priority engine. The worker runs on the deployment machine, which need not be your device. No silent external model fallback is used; queued does not mean transcribed. Recognition can be inaccurate and supplies no verified speaker identity. A transcript does not authorize tasks, sharing or publication.
Device drafts, server audio/transcripts and exported copies have separate lifecycles. The reviewed 7 or 30 day server retention and 15 minute unfinished-upload expiry require successful worker cleanup. Cancellation is not deletion; pending removal becomes verified only after the service confirms finalized deletion. See Private meeting audio privacy for consent renewal, device recovery, processing and export details.
Separately reviewed notes, decisions and task proposals are stored as private evidence with their source transcript citations. They share the audio job’s access, retention and verified deletion lifecycle; saving a task proposal creates no task. Exported evidence is a separate copy that server deletion cannot remove.
What it cannot do
- It cannot read intent you did not write down. It ranks the work it can see, and its view of your life is only as complete as what you have given it.
- It does not provide professional advice — legal, financial, medical, tax, or otherwise.
- Reading documents and notes can miss, duplicate, or misread something. That is why import has a review step and why uncertain items arrive unchecked.
- It does not guarantee any outcome. It organizes and recommends; you decide.
You stay in control
Every recommendation is a suggestion with its reasoning attached, not an instruction. You can overrule it, and nothing is written into your workspace from an import or a capture until you have reviewed and confirmed it.
If this ever changes
If customer workspace planning or imports introduce external model processing, it will be described on this page and in our Privacy Policy before it is switched on, the provider will be named in our list of service providers, and the feature will be one you choose to use rather than one that is on by default.
This page is part of the VectorOS Trust Center. Questions? Contact vectoros.app@gmail.com.