Security
Where your agency’s data lives, and who can reach it.
Agencies hold their clients’ budgets, supplier terms and unreleased creative. This page states how that is protected, and is deliberately blunt about what we have achieved versus what we are working towards.
Assurance register
No certification on this list has been awarded yet. We would rather say so here than let a badge imply otherwise in a procurement pack.
| Standard | Status | What that means |
|---|---|---|
| ISO 27001 | Target | Information security management. Not certified — the controls below are built and enforced in code review; the audit is not done. |
| SOC 2 | Target | Not certified. Change management already produces much of the evidence by construction: every change lands through a pull request with an architecture decision record. |
| ISO 42001 | Target | AI management. Not certified. The alignment work — model gateway, logged context assembly, eval-gated releases — is described below. |
| ISO 9001 | Backlog | Quality management. Scoped but not started: the approvals audit trail and the test bar produce most of the evidence, the formal process documentation does not exist yet. |
| UK GDPR | Obligation | Not a certificate — the law we operate under. Residency, retention, erasure paths and the processor register are covered below. |
| WCAG 2.2 AA | Build bar | The accessibility floor for every surface, with automated axe checks running in continuous integration against the design system. |
Registered with the Information Commissioner’s Office: ZC223641.
Tenant isolation
This is the control everything else rests on, so it is enforced by the database rather than by application code remembering to filter.
- Every row carries a tenant identifier, and Postgres row-level security enforces it. The application’s database role cannot bypass those policies.
- Policies live in the same migration as the tables they protect, and a migration-time check raises if a tenant-scoped table ships without one.
- Cross-tenant reads are a required test category, not a code-review hope: suites must include attempts that are expected to fail, including with no tenant context set.
- Agency data and a client’s own data never cross without an explicit, recorded, per-engagement grant — and that grant is visible to both sides.
Data protection
- Residency. EU/UK regions, with London (
europe-west2) primary. - Encryption. In transit everywhere; customer-managed encryption keys per tenant for stored documents.
- Retention and erasure. Every entity holding personal data has a retention policy and an erasure path, covering both subject-access and deletion. Attendee data default-expires after the event, per the agency’s own policy.
- Processors. Any adapter that touches personal data — email, messaging, card feeds, CRM — is named in a processor register with a data processing agreement in place. Adding one requires a recorded architecture decision, not a configuration change.
- Access. Public surfaces sit behind a web application firewall with per-tenant and per-IP rate limits; file access is via short-lived scoped signed URLs. Production access is break-glass, time-boxed and logged.
AI governance
The commercial question every agency asks first: does using this train a model that helps a competitor? No.
- Your data stays yours. Each agency’s OS learns from that agency’s own work, held under the same row-level security as everything else. There is no live cross-tenant retrieval, and tenant content never enters a shared model or prompt.
- The platform only learns through a gate. Anything that improves agense for everyone passes an aggregation gate first: input that is already free of personal data, tenant identity reduced to coarse cohorts, k-anonymity thresholds so no small group can be singled out, and explicit consent tiers. Client-side tenants and anything shared under an engagement grant are excluded from that gate entirely.
- We do not trade your data for features. There is no "share your data for free credits" option, because that is the bargain this design exists to refuse.
- Tenant memory is exportable and deletable. What your OS has learned is your intellectual property. Erasure cascades: memory derived from deleted source content is purged with it.
- Every draft is reproducible. Model calls go through one gateway, and context assembly is logged with artefact versions and memory references. Prompt and response logs are tenant-scoped and redacted.
- Agents draft; humans commit. Supplier commitments, client-facing financial commitments and event-safety sign-offs always route through the approvals service. This is enforced, and attempts to bypass it are a required test category.
Accessibility
WCAG 2.2 AA is the build bar rather than a retrofit. The design system ships AA-tested colour pairs, and automated accessibility checks run in continuous integration. Onsite surfaces are built to work with no connectivity and reconcile later — venues have unreliable networks, and an interface that fails without a signal is an accessibility problem before it is a technical one.
Reporting a vulnerability
Please report suspected vulnerabilities to security@agense.io before disclosing publicly. We will acknowledge, keep you updated, and credit you if you would like us to.
For anything else — a security questionnaire, a data processing agreement, or a procurement review — write to hello@agense.io.