Subprocessors
Last updated September 15, 2026
What this page covers
A subprocessor is a third party we use to deliver Astrid that may process your firm's data in the course of doing so. This page lists all of them, what each one is used for, what it actually receives, and the country its processing takes place in. It is the list we would attach to a data processing agreement, published here so a security reviewer does not have to ask for it.
Astrid Labs, Inc. is a United States company, and every subprocessor that handles firm content is a United States company processing in the United States. One entry below, our marketing-site analytics provider, is Canadian; it runs only on the public website and receives no firm data. See Data residency for how each one is pinned and how you can verify it.
Subprocessors that handle firm content
These four are on the path of every call. Call recordings, transcripts, and drafted notes are privileged or otherwise sensitive, so this is the set that matters most in a security review.
| Subprocessor | Purpose | Processing location |
|---|---|---|
| Google Cloud Platform Google LLC | All application hosting, the database, call audio storage, secret storage, and error reporting. Astrid runs on Cloud Run, Cloud SQL, and Cloud Storage. | United States |
| AssemblyAI AssemblyAI, Inc. | Speech-to-text transcription, speaker separation, and the optional redaction of personal information from the transcript. The only subprocessor that receives call audio for processing; the capture providers below create and hold recordings of the calls they capture. | United States |
| Anthropic Anthropic, PBC | The Claude model that drafts case notes, recap emails, call titles, and call-type classification, and that answers questions in Ask Astrid. Receives transcript text only, never audio. | United States. Every request is pinned to US-only inference; see Data residency. |
| Inngest Inngest, Inc. | Background job orchestration for the capture pipeline. Job results, which include transcript and drafted note text, are stored by Inngest so an interrupted job can resume without redoing completed work. | United States |
Limited-scope subprocessors
These handle a specific function and receive correspondingly less. Two of them are only used if your firm turns on the feature behind them.
| Subprocessor | Purpose | Processing location |
|---|---|---|
| Recall.ai Recall.ai, Inc. | The meeting notetaker that joins a Zoom, Google Meet, or Microsoft Teams call and records it. Creates and holds the recording under its own retention, which your firm's Astrid audio retention setting does not govern. Only used where someone at your firm enables the notetaker. | United States (US West) |
| Twilio Twilio Inc. | The phone bridge, which lets an attorney merge Astrid into a live landline or mobile call. Creates and holds the recording under its own retention, which your firm's Astrid audio retention setting does not govern. Only used when someone at your firm starts a bridged call. | United States |
| Resend Plus Five Five, Inc. | Account email only: email verification, password reset, firm invitations, and billing notices. Receives no call content. Client recap emails do not go through Resend; they are sent from your own attorney's mailbox. | United States |
| Stripe Stripe, Inc. | Subscription billing and payment processing. Receives billing contact and payment details only, and no call, client, or matter content. | United States |
| Fathom Analytics Conva Ventures Inc. | Privacy-focused, cookie-free traffic analytics on the public astrid.legal marketing site only. It is not loaded anywhere inside the signed-in application and never sees firm data. | Not applicable: no firm data |
Systems your firm connects
Astrid reads from and writes to systems your firm already owns: your case management system (Clio, Smokeball, CasePeer, MyCase, or Lawcus), your Google Workspace or Microsoft 365 calendar and mailbox, your OneDrive, and your Zoom account. These are not Astrid's subprocessors. They are your firm's own vendors, under your firm's own agreements, and Astrid only reaches them because someone at your firm connected the account and authorized the scope.
Two consequences worth stating plainly. A note filed to your case management system is sent at an attorney's explicit instruction, to your instance of that system. A client recap email is sent from the attorney's own Gmail or Microsoft 365 mailbox and lands in their own Sent folder, so it travels through your firm's mail provider rather than ours.
Data residency
All firm data is stored and processed in the United States. Astrid's own infrastructure runs in Google Cloud's us-east1 region: the application, the database, and call audio storage. We operate no infrastructure anywhere else, and we do not replicate, fail over, or back up to any region outside the United States.
Every subprocessor is pinned to the United States, and each pin rests on something published rather than on our say-so. Google Cloud processes in the region we selected. Inngest states that its databases are all located in the United States. Resend lists every one of its own subprocessors in the USA. Our transcription provider publishes separate North American and European endpoints, and we use the North American one. Our meeting notetaker runs in a United States region we set explicitly.
AI inference is pinned to the United States on every single request. Anthropic's API defaults to running inference in any available region. We do not use that default. Every request Astrid sends carries an explicit US-only instruction, applied in one place in our codebase rather than at each point where we call the model, so that no future change can quietly leave a path unpinned. We pay a premium on every token for this, and we think it is the right trade for privileged material.
This one is verifiable rather than merely asserted, which we would rather you not take on trust. Anthropic returns the region that actually served each request in its response. We can produce that confirmation for a specific request on request, and we retain the ability to demonstrate it live during a security review.
One qualification we would rather state than have you infer. Pinning is enforced by configuration that we control and can evidence, not by a negotiated contractual residency clause with each vendor. Several of these vendors are global companies whose own disclosures describe a wider footprint for their other customers. If your firm needs contractual data-residency commitments rather than demonstrable configuration, that is a fair distinction and we will work through it with you directly rather than blur the two.
Astrid is currently offered to firms in the United States. If you need data stored or processed in another country, we cannot do that today, and we would rather tell you that up front than during an implementation.
What each subprocessor actually receives
Scope matters more than the vendor count, so here is the path a call takes.
- Audio is stored in Google Cloud Storage and read by AssemblyAI to produce the transcript, through a short-lived signed link rather than a public URL. The bucket is configured to refuse public access at the platform level, so it cannot be made publicly readable by a configuration mistake. Where a call is captured by the phone bridge or a meeting bot, that capture provider also creates and holds its own copy of the recording. The drafting model never receives audio.
- Transcript text is produced by AssemblyAI. If your firm has enabled redaction of personal information, that redaction happens during transcription, so the unredacted text is never written to our database and never reaches any later step.
- Anthropic receives transcript text and your firm's note template in order to draft the note and recap, and receives the note and transcript content relevant to a question asked through Ask Astrid.
- Inngest coordinates these steps. The messages that start a job carry only an internal record identifier, but the results each step reports back include transcript and note text, which Inngest retains so a failed job can resume rather than restart.
- Nothing else on this page receives call recordings, transcripts, notes, or recap content.
Model training
Our transcription and AI providers process your firm's content only to return the output Astrid asked for. We do not permit them to use your firm's content to train or improve models for their own benefit or any other customer's.
Changes to this list
We will update this page when we add or replace a subprocessor. If your firm has an agreement with us that requires advance notice of a change, that notice is given under the agreement rather than by this page alone. Questions about anything here can go to jason@astrid.legal.