AskReplay

Trust Center

Everything your security team will ask

Written for the people who review vendors: what we do, where your data goes, and the documents you need to close a review. Last reviewed August 2, 2026.

Download the vendor package (PDF)Read it in the browser

Your content does not train AI models

Customer content is never used to train models, ours or a provider’s. It is sent to the model provider only to answer a specific request, under API terms that exclude training on submitted data. Buyer-facing answers are generated only from content you have approved, they carry citations bound in code to the passages actually retrieved, and when your approved content cannot answer a question the system says so and logs the gap instead of improvising. Buyer-facing answering has no access to the open web.

Controls in place today

Access control and identityWho can get in, how they prove it, and how access ends.
  • Sign-in with email and password, Google, or Microsoft. SAML single sign-on is available for workspaces that register their identity provider.
  • Two-factor authentication (TOTP authenticator app), enforced server-side on every request rather than only at the sign-in screen. A session that has not cleared its second factor is refused by the API.
  • SCIM 2.0 directory sync provisions and deactivates users from the customer’s identity provider.
  • Role-based access with two roles. Administrators control data export, public publishing, and permanent deletion; members do day-to-day work.
  • Session tokens are signed and httpOnly, verified server-side on every request.
  • Administrators can end a person’s sessions immediately, or remove the account entirely. Both take effect without waiting for token expiry.
  • Privileged operations run server-side under scoped service credentials; the browser never holds them.
Data protectionHow customer content is separated, encrypted, and destroyed.
  • Encryption in transit using TLS. Encryption at rest is provided by our infrastructure providers.
  • Tenant isolation: every record is scoped to a workspace, enforced by database row-level security in addition to application checks.
  • Uploaded video is stored privately and served through short-lived signed URLs, never a public object URL.
  • Per-recording passwords and optional email gating on buyer-facing pages. Deal rooms are invite-only by email with one-time code verification.
  • Deletion is real and complete: removing a workspace, recording, or source purges the database rows, the vector embeddings, and the stored objects.
  • Optional content approval, so newly added knowledge cannot answer a buyer until a reviewer approves it.
AI and content handlingThe question security teams now ask first: where does our content go, and does it train anything.
  • Customer content is not used to train AI models, ours or a provider’s. It is sent to the model provider only to generate a response for that request.
  • Answers are generated only from content the customer has added and approved, and every buyer-facing answer carries citations to the source it came from.
  • The system fails closed. When approved content cannot answer a question, the product says so and logs the gap rather than generating an unsupported answer. This is enforced in code, not by prompt instruction alone.
  • Buyer-facing answers never draw on the open web. Web search is used only in internal, seller-facing drafting tools, and those drafts are reviewed by a person before use.
  • Model providers are listed as subprocessors. Text sent for embedding and completion is processed under those providers’ API terms, which exclude training on API data.
  • Administrators control which content is available to answer, and can revoke a source at any time; retrieval reflects the change immediately.
Application securityHow the software itself is built and defended.
  • Rate limiting on public, authenticated, and AI endpoints, with per-workspace ceilings on model usage.
  • Server-side request forgery protections on every feature that fetches a customer-supplied URL: HTTPS only, DNS resolved and checked against private address ranges, no cross-host redirects, and byte and time caps.
  • Input validation on API boundaries with schema enforcement.
  • Content Security Policy, clickjacking protection, strict transport security, and MIME-sniffing protection set at the edge.
  • Secrets are held in the hosting provider’s encrypted environment configuration and are never exposed to the browser.
  • Continuous integration runs type checking, linting, and an offline grounding evaluation on every change, so a regression in the citation or abstention behavior fails the build.
Infrastructure and operationsWhere the service runs and how it is watched.
  • Hosted on Vercel (application and edge delivery) with Supabase for authentication and the primary database, both in the United States.
  • A health endpoint verifies the database, vector store, model provider, and object storage.
  • Application errors carry a stable fingerprint and release identifier, with throttled alerting to the operator.
  • An administrative audit log records who performed each privileged action, from which IP address, with CSV export available to workspace administrators.
  • Infrastructure is managed by providers who maintain their own compliance programs; we do not operate physical infrastructure.
Privacy and data rightsPersonal data, agreements, and deletion rights.
  • A Data Processing Agreement is available and published at /dpa.
  • The subprocessor list is public at /subprocessors and is updated when it changes materially.
  • Buyer email capture on demos is optional and controlled by the customer.
  • Customers can delete their workspace and all associated content at any time; deletion cascades to the database, vector store, and object storage.
  • Personal data processed is limited to account holders (workspace users) and any viewer contact details the customer chooses to collect.
Availability and recoveryBackups and continuity, stated precisely.
  • The primary database is backed up by the provider on a daily schedule with point-in-time recovery.
  • Application code is versioned and deployments are immutable, so a bad release can be rolled back to a previous deployment.
  • Object storage and vector data are managed by providers with their own durability guarantees.

Certifications

We do not hold SOC 2 or ISO 27001 today, and we would rather tell you than let you find out. The controls those audits examine are implemented and described above. The vendor package sets out our full position, including what we do not have yet and what stands in its place, and we begin a formal audit when a customer agreement requires the report.

Documents

Need something not listed, a completed copy of your own questionnaire, or a W-9? Email contact@ask-replay.com. We aim to return completed security questionnaires within two business days.