FOR YOUR IT TEAM

Security & privacy

// Written to be forwarded. Every claim here is checked against the running code.

vernac is a server-side service. It reads mail out of your mailbox through Microsoft Graph, processes it on our servers and with the providers named below, and writes drafts back. Two things follow from that, and we would rather you hear them from us:

The one that matters: it cannot send

The Microsoft app registration deliberately omits Mail.Send. vernac can read your mail and stage drafts in your Drafts folder; you press send. This is not a setting, a policy or a promise about our conduct — the permission was never requested, so the capability does not exist in the token. You can verify it on the Microsoft consent screen when you connect, and again afterwards in your account's app permissions.

Exactly what is requested

This table is generated from the same list the application requests, so it cannot drift from what Microsoft shows you.

Delegated scopeWhat it is for
User.ReadYour name and address, so drafts sign off as you
Mail.ReadWriteRead your mail, tag it, and stage drafts
MailboxSettings.ReadWriteCreate the category set in your mailbox
Calendars.ReadWriteYour own calendar: availability, and the hold you confirm
Calendars.Read.SharedFree/busy only for people you schedule with — never their event details
openid · profile · email · offline_accessSign-in, and staying signed in
Mail.SendNever requested.

Calendars.ReadWrite covers the user's own calendar only — it is what writes the hold when a scheduled time converges, and what the user confirms. Calendars.Read.Shared reads free/busy for the people being scheduled with: availability, never event details.

No admin consent is required, and nothing is deployed into the Outlook client. A user grants these with their own sign-in and can revoke them the same way.

What leaves the mailbox, and what we keep

DataGoes toStored?At rest
OAuth tokensour serversYesEncrypted (AES-256-GCM)
Inbound email bodyour servers, and our AI providerNo — classified, then droppedNot persisted
Inbound email metadata (addresses, subject, timestamps)our serversYesPlaintext, database access controls
Your sent email (the voice corpus)our serversYesEncrypted
Draft vs. sent text (correction pairs)our serversYesEncrypted
Meeting transcripts and notes (read from Zoom, Granola or Meet)our serversYesEncrypted
Documents attached to your mailour servers, and our embedding providerYesEncrypted
Classification labels and metricsour serversYesPlaintext
We do not keep the mail you receive. Inbound bodies are classified in memory and dropped — there is no column to put one in. What persists is metadata, a label, and the writing that teaches us your voice: your own sent mail, and your edits to our drafts.

Meeting transcripts

Nothing joins your calls, and vernac does not record them. Meetings are read from transcripts your own tools already produced — Zoom AI Companion, Granola, Google Meet — using the same per-user OAuth grant you can revoke at any time. There is no recording step anywhere in the product, so no meeting audio ever reaches us — not from a bot, not from a device, not at all. The transcription happened inside a tool you were already using before vernac existed, and there is no speech-to-text provider in our sub-processor list because nothing here turns audio into words.

Only the transcript text is stored, encrypted like everything else. The limitation is the honest one: a call your own tools never transcribed is a call vernac cannot write up.

Encryption

TLS in transit. At rest, sensitive content is sealed with authenticated AES-256-GCM: OAuth tokens, your sent-mail corpus, meeting transcripts, indexed documents and correction pairs. Each customer tenant has a randomly generated data key, wrapped by an application master key held outside the database. We do not currently claim KMS-managed keys; the wrapping key is an application secret.

Why there is nothing to train

This is the question we get asked most, and the answer is architectural rather than a policy we could quietly change.

vernac contains no training step. There is no fine-tune, no shared model weights, no central corpus that grows as customers join. What people call "learning your voice" here is two ordinary mechanisms: retrieval over your own sent mail at drafting time, and a handful of imperative rules distilled from your own edits to our drafts. Both are rows in your account, read back by queries scoped to your user id, and both are deleted when the account is.

Two consequences worth stating plainly. Nothing you write can appear in another customer's draft, because no query can reach across accounts. And there is no accumulated model that would survive your deletion — the asset a competitor would describe as getting smarter with every customer does not exist here, deliberately. What makes the drafts good for you is that they are built from you.

What we cannot claim. Drafting sends your content to our AI provider to be processed, every time — that is inference, not training, and it is disclosed below with everyone else who touches your data. A service that told you nothing ever leaves would not be able to write anything.

Sub-processors

The complete list of third parties that process customer content:

WhoWhat they process
MicrosoftGraph API — reading mail, tagging it, staging drafts, free/busy
AnthropicClassification, drafting, meeting summarisation and scheduling replies
Voyage AIEmbeddings for retrieval over your sent mail, transcripts and documents
PostmarkDelivering your brief, and receiving the replies you send back to it
Fly.ioApplication hosting and the Postgres database (primary region: US West)

Your writing is used to answer for you and to draft as you — never to train foundation models, and never to serve another customer.

Access, retention and deletion

What we are not claiming

We are an early-stage product and it is easier to check us now than later. We do not hold SOC 2, ISO 27001 or a Microsoft 365 certification, we make no data-residency guarantee beyond the hosting region above, and we do not describe ourselves as HIPAA or GDPR "compliant". If any of those are a requirement for your organisation, tell us before you start rather than after.

Questions

Security review, a questionnaire, or an architecture call: shane@redeavorgroup.com. We answer these ourselves.