How we protect your information
Last reviewed November 19, 2026
SageRock builds AI assistants and runs admissions, marketing and back-office systems for schools, nonprofits, law firms and small businesses. This page explains what we hold, where it lives and how it is protected. We keep it current, and we update it when our systems or vendors change.
Who can use an assistant
- Each assistant answers by email. Every incoming message is checked against the sender's domain (SPF and DKIM, with domain alignment), so a forged address is rejected.
- Only approved senders reach the assistant: an organization's own domain and the people it names. Unknown or unverified senders go to a person, never to the AI.
- The assistant does not choose who it writes to. Replies are bound to the verified sender and thread, and it cannot email a mailing list.
What an assistant can and cannot do
- Read-only by default. Each assistant gets only the tools its job needs, and only for its own organization.
- Changes wait for a yes. A change to an email list, an ad budget or a held message is written up as a proposal. An authorized person replies with a single-use code, and only then does plain code, not the AI, apply exactly that change and check the result.
- Some things it cannot do at all: send to a mailing list, schedule a campaign, write to accounting systems, or delete records in a client's systems.
- Built against hidden instructions. Text from emails and attachments is treated as untrusted. The safeguards that matter (sender check, tool list, confirmation codes, recipient binding) are enforced in code, so a manipulated AI still cannot widen its own permissions.
Keeping clients separate
- Every client's records are tagged with that client's ID, and every query is limited to it.
- Each assistant has only its own organization's tools and document search. Its memory of people belongs to its own mailbox, and other clients' assistants cannot read it.
Encryption and credentials
- All connections use HTTPS/TLS. Credentials an assistant uses to reach a client's systems are encrypted at rest (AES-256-GCM), and the application itself contains no keys.
- Email attachments are kept in a private, encrypted Amazon S3 bucket with public access blocked.
- Shared credentials live in a 1Password business vault. Every outgoing assistant email is checked against our stored secrets and refused if it would reveal one.
AI providers
- Our assistants use the paid business APIs of Anthropic (Claude) and OpenAI. Under those business terms, customer content is not used to train their models. The providers process content to produce an answer and keep it only as their retention terms allow.
- For law firms, client information runs by default through Amazon Bedrock in SageRock's own AWS account, in the United States, with data retention turned off.
- Some behind-the-scenes work runs on SageRock's own workstation using Anthropic's Claude Code, on an account with model training turned off.
People in the loop
- Every assistant email is archived and signed as AI, and a SageRock person is copied on client replies. A nightly audit reviews what was sent.
- New assistants and new abilities start in draft mode, where a person approves replies before anything goes out.
- Each organization can have its full email history handed back on request.
Backups
- Client databases: nightly encrypted backups to Amazon S3, on top of the hosting provider's own backups.
- Assistant database: daily, weekly and monthly snapshots plus a nightly encrypted off-site copy, with a quarterly restore test.
Incident history
- On October 21, 2026, our application host Railway announced that customer environment variables (stored settings such as keys) may have been read. We began rotating the credentials stored there that night, including AI, database, email and Google keys, and finished the clean-up in early November.
Honest limits
- SageRock is a small firm. We do not hold a SOC 2 or ISO 27001 certification. Our main providers (Amazon Web Services, Supabase and Anthropic among them) hold their own.
- This page describes how our systems are built and run today. If you are a client, we can write a version specific to your organization, answer your auditor's questionnaire or walk through any item.
Vendors we use
Supabase (databases), Amazon Web Services (storage and Bedrock), Railway (application hosting), SendGrid (email delivery), Anthropic and OpenAI (AI), Qdrant (document search), Cloudflare (website hosting and DNS), Google Workspace, and 1Password (credentials). Client work may also use that client's own tools, such as Constant Contact, Google Analytics, Google Ads or Meta.
Questions about security? Email us at [email protected].