Residents’ details go where they agree, and nowhere else.
Sepal Concierge is multi-tenant: many organizations share the software, but every record carries its organization and every query is constrained to it. Inside an organization, each agency sees only what a data-sharing agreement and the resident’s consent allow.
Data sharing between agencies
- A data-sharing matrix lists, for each agency, which categories of detail it may receive, for what purpose, on what basis and for how long. The orchestrator enforces it on every A2A message; anything not allowed is withheld, and the agent never sees it.
- Where the basis is consent, the resident sees a consent card naming the agency and the details before anything is shared, and can withdraw it later. Consent records expire with the data flow’s retention period.
- Social Security, bank and card numbers are removed from a message before any agent or model sees it.
No writes to live systems
- Tool servers on live agency systems are read-only by declaration and refused by the client if a tool would change anything.
- Pre-filled forms pass through one write-back guard that allows only sandbox destinations until an organization turns write-back on and the agent is no longer sandbox-only.
Isolation and access
- Every table carries the organization, and the data layer refuses any query that isn’t scoped to the caller’s organization. A resident session belongs to the organization it started in.
- Each organization has its own Google Identity Platform tenant; staff sign in with Okta, Entra ID or any SAML 2.0 / OpenID Connect provider, with MFA.
- Agency scope limits partner staff to their own agency’s agents, knowledge, reviews and conversations. Agent-to-agent calls use OAuth client credentials or mutual TLS.
- Agent registration goes through an approval route; segregation-of-duties rules flag conflicting roles.
Transcripts and model use
- Transcripts are de-identified — names, addresses, emails, phone numbers and confirmation codes replaced — before they are stored or counted. Resident details live only in the conversation session, which expires.
- The orchestrator’s model runs on US endpoints only. Your data is never used to train models, by us or by our model providers. Agencies choose their own agents’ models.
- Every answer discloses that it comes from an AI assistant, as the Texas Responsible AI Governance Act requires, and offers a person.
Audit trail
- Every sign-in, permission change, agent registration, conformance run, consent, form confirmation, review decision, export and reveal is written to an append-only audit log.
- The log is sealed daily with a hash chain so tampering is detectable, and can stream to Splunk or Microsoft Sentinel.
Accessibility
- The resident front door and the staff console are built to WCAG 2.2 AA: keyboard and screen-reader use, larger text and high-contrast modes, right-to-left layout for Arabic and Pashto, and a table behind every chart.
- A current VPAT (ACR) is available, and accessibility issues are fixed as defects.
Infrastructure
- Runs on Google Cloud (Firebase Hosting, Cloud Functions, Cloud SQL for PostgreSQL through Data Connect, Firestore) in US regions, with Dallas as the primary region — or in a dedicated project, the organization’s own cloud, or the agent runtime container on premises.
- Controls aligned to the NIST SP 800-53 Moderate baseline used by TX-RAMP Level 2; the underlying Google Cloud services hold TX-RAMP authorization.
- Encryption in transit (TLS 1.2+) and at rest; secrets in Secret Manager; rate limits on the public front door.
- Point-in-time recovery, daily backups to a second US region, a one-hour recovery point and an eight-hour recovery time, tested yearly.
We’re glad to complete your security questionnaire, the TX-RAMP assessment or the SIG Lite as part of an evaluation.