Superhuman AI Data Privacy, Third-Party LLM Processing, and Enterprise Security Compliance

For enterprise executives, security architects, and compliance officers, adopting an AI-powered email client involves high stakes. Email remains the primary nervous system of corporate communication, carrying sensitive customer records, financial projections, personnel discussions, and proprietary strategy. At a premium price point of $30 per user per month (with custom pricing for enterprise plans), Superhuman positions itself as an executive-tier productivity suite. However, introducing third-party large language models into this communication layer raises immediate compliance questions.

Superhuman addresses these concerns by combining a local-first email caching architecture, an enterprise Zero Data Retention (ZDR) agreement with OpenAI, a strict ban on training models with customer data, and independent third-party audits including SOC 2 Type II, ISO 27001, and Google CASA assessments.

1. Core Architecture: Local-First Caching vs. Centralized Storage

A common misconception among IT evaluators is that Superhuman copies and stores an organization’s entire email database on central servers. In reality, Superhuman’s client architecture functions primarily as a high-speed local-first overlay on top of existing mail providers (Google Workspace and Microsoft 365).

• Direct OAuth 2.0 Integration: Users authenticate directly with Google or Microsoft via standard OAuth 2.0 tokens. Superhuman requests necessary mailbox scopes without storing raw passwords or master corporate credentials.
• Client-Side Storage Engine: Superhuman caches recent messages, search indices, and interface assets locally on the user's device inside browser-level storage technologies (IndexedDB, WebSQL, and CacheStorage). Triaging, drafting, and organizing actions are tracked locally as modifier queues and synchronized back to Google or Microsoft APIs.
• No Permanent Central Mailbox Storage: Unlike traditional hosted webmail platforms that duplicate multi-year inbox archives on their own servers, Superhuman’s backend infrastructure processes metadata and sync states in transit rather than maintaining a permanent cloud mirror of the full mailbox.

2. Third-Party LLM Processing: The OpenAI Zero Data Retention (ZDR) Mechanism

When users interact with Superhuman AI features—such as drafting messages with "Write with AI", generating summaries with "Auto Summarize", producing context-aware replies via "Instant Reply", or searching inboxes using "Ask AI"—the application communicates with external LLM inference endpoints, primarily powered by OpenAI.

The interaction model is governed by specific contractual and technical controls:

[User Inbox / Client]
       │
       ▼ (Encrypted via TLS 1.3)
[Superhuman Backend API]
       │
       ▼ (Encrypted in transit via TLS 1.3; Zero Data Retention)
[OpenAI Enterprise Inference API]
       │
       └── Processes prompt in volatile memory
       └── Generates response
       └── Immediately discards prompt and completion (ZDR)
       │
       ▼ (Returns completion payload)
[Superhuman Backend & Client]

The Zero Data Retention (ZDR) Commitment

Through its enterprise agreement with OpenAI, Superhuman enforces Zero Data Retention for eligible inference workflows.

1. No Disk Logging at OpenAI: OpenAI processes the incoming email snippets and prompts entirely in volatile memory (RAM) to generate text completions. Once the response is returned to Superhuman, OpenAI immediately flushes the prompt and completion payload. Neither the prompts nor the generated outputs are stored on OpenAI servers.
2. Strict Model Training Prohibition: Superhuman contractually prohibits both OpenAI and its own internal algorithms from using customer email data, prompts, or drafts to train or fine-tune public foundation models or shared machine learning weights.
3. Transit Encryption: All requests dispatched to third-party model providers travel over TLS 1.3 encryption.

Feature-by-Feature Data Retention on Superhuman's Infrastructure

While OpenAI retains zero data under the ZDR agreement, Superhuman's own cloud infrastructure applies varying retention policies depending on the specific AI feature:

3. Enterprise Security Compliance and Verification

For enterprise security review boards, marketing claims are insufficient without accredited external audit trails. Superhuman maintains verifiable compliance frameworks across multiple administrative vectors:

SOC 2 Type II and ISO 27001 Certification

Superhuman undergoes regular, independent third-party audits to achieve and maintain SOC 2 Type II compliance, verifying controls across Security, Availability, and Confidentiality over continuous audit observation windows. The service also holds ISO 27001 certification for information security management.

Google CASA (Cloud Application Security Assessment)

Because Superhuman requires access to restricted Gmail API scopes to read, draft, and organize messages, Google mandates annual independent security evaluations. Superhuman fulfills this requirement through Cloud Application Security Assessment (CASA) Tier 2 validation. CASA audits test adherence to OWASP Application Security Verification Standards (ASVS), covering:

• Data handling and secure storage on client endpoints and backend proxies.
• Automated and manual vulnerability testing against unauthorized data access.
• Verification of data deletion mechanisms and least-privilege token usage.

Enterprise Identity and Data Governance Controls

Superhuman Enterprise plans include administrative tools designed to satisfy enterprise IT policy:

• SAML 2.0 Single Sign-On (SSO) & SCIM: Seamless integration with Okta, Microsoft Entra ID (Azure AD), Google Workspace, and Ping Identity, enabling centralized provisioning, de-provisioning, and multi-factor authentication (MFA) enforcement.
• Bring Your Own Key (BYOK) Encryption: Available on select enterprise agreements to give corporate security teams cryptographic ownership over data at rest.
• Data Processing Addendum (DPA) with Standard Contractual Clauses (SCCs): Full GDPR compliance mechanisms for European cross-border data transfers, confirming Superhuman acts strictly as a data processor.
• Network-Level AI Deactivation: Enterprise IT administrators can disable Superhuman Go / AI features across internal networks via DNS routing (for instance, redirecting capi.grammarly.com to capi-restricted.grammarly.com).
• Individual AI Deactivation: End users can independently toggle off generative AI by invoking the Command Menu (Cmd+K on macOS or Ctrl+K on Windows) and selecting "Deactivate Superhuman AI".

4. How Superhuman AI Compares with Alternative Email Workflows

Enterprise teams generally evaluate Superhuman AI against two alternatives: built-in native suite assistants (Google Gemini for Google Workspace, Microsoft 365 Copilot) and unvetted browser-based copy-pasting into public ChatGPT or Claude interfaces.

5. Security Checklist for Enterprise Adoption

Before approving Superhuman AI for leadership teams or business units, enterprise CISOs and IT administrators should review the following technical factors:

1. Verify the Zero Data Retention Confirmation: Ensure that your corporate agreement explicitly references Superhuman's Zero Data Retention subprocessing terms with OpenAI, validating that no transient email payloads are retained beyond the active inference round-trip.
2. Review the 90-Day Cache Policy for Summaries: Assess whether the 90-day temporary retention of email summaries and instant replies aligns with your company’s internal document retention and e-discovery policies.
3. Distinguish Between Productivity and Archiving: Superhuman is an executive productivity interface, not an immutable compliance archive. Organizations subject to FINRA, SEC Rule 17a-4, or specialized compliance mandates must maintain their existing journaling and archival solutions (such as Smarsh, Global Relay, or Mimecast) at the Google Workspace or Microsoft 365 tenant level.
4. Implement Centralized SAML SSO and SCIM: Always deploy Superhuman via your identity provider so that when an executive or sales representative departs, access to the client and its cached data is revoked instantly.

For organizations seeking to boost executive velocity without exposing email communications to public model training or unmonitored shadow AI, Superhuman AI establishes a defensible, audited, and contractually enforced security posture.

References

• Superhuman Privacy Policy
• Superhuman Pricing & Plans
• OpenAI Customer Stories: Superhuman