Trust boundaries define where identity, policy, data access, and execution are enforced — so AI-assisted work stays accountable end-to-end.
Why it matters
Enterprise AI fails when applications hold long-lived system credentials, skip role context, or execute outside audit paths. AI Fabrix enforces boundaries structurally — not through prompt instructions alone.
Architects document these zones in security reviews; integrators implement config that respects them; operators rely on certification proving they hold.
What remains authoritative
AI Fabrix does not become a new system of record, and it does not create a new source of organizational authority. Two things stay exactly where they already are:
- Existing systems remain authoritative for the records and operations they already own. A capability crossing the boundary map below reads and writes through that system — it does not copy that system's authority into AI Fabrix.
- Human identity, roles, permissions, and approvals remain the basis for what may happen. AI Fabrix checks that authority live, at the moment work runs; it does not grant authority of its own.
What AI Fabrix adds is a governed way to reuse and validate what those systems already mean and allow — not to replace them:
- The current state of your systems, people, and policy — what this section calls Enterprise Reality — is read, never owned. If a system of record changes, that change is the truth; AI Fabrix's view of it is revalidated, not authoritative on its own.
- Business meaning built from that reality — Enterprise Knowledge — is a governed, reusable copy for AI to reason with, validated against its source. If the two ever disagree, the source system wins.
- Once work executes, what actually happened is preserved as the result, and validated results become Evidence. Evidence records the outcome for proof and learning; it is not a new grant of authority for the next request.
The zone map below is where this gets enforced structurally: the identity/controller zone reflects your current people, roles, and approvals, and every zone after it re-checks against that authority rather than a cached copy from setup time.
Boundary map
| Zone | Responsibility | Trust controls |
|---|---|---|
| Identity / controller | Users, groups, roles, deploy lifecycle | Authentication, RBAC registration, environment policy |
| Integration config | Local and published manifests | Schema validation, repair, upload pipeline |
| Dataplane execution | Data access, CIP (Composable Integration Pipeline), sync | ABAC dimensions, protection, in-process identity for runs |
| Capability gateway | Pre-execution gate | Role, scope, certification, approval, evidence requirements |
| AI / worker surface | Task planning and requests | No direct system mutation; capability requests only |
External systems sit outside the trust envelope. Only mediated capabilities cross inward.
Request path (simplified)
User + active role
→ Role Assistant (task context)
→ Capability request
→ Gateway checks (Operational Trust)
→ Dataplane execution (CIP / connector)
→ Evidence capture (Evidence Fabrix)
Deny at the gateway is expected behavior when role, certification, or policy does not allow the action — not a platform defect.
Certification aligns to boundaries
| Pillar | Boundary exercised |
|---|---|
verify-operations |
Dataplane execution + external connectivity |
verify-trust |
Metadata complete for AI/worker surface |
verify-governance |
Dimensions/protection at execution boundary |
Integrators run certification after upload; architects define when each pillar is required for production.
Anti-patterns
| Anti-pattern | Risk |
|---|---|
| Embedding API keys in worker prompts | Credential leak; no audit path |
| Bypassing capabilities for “speed” | Ungoverned mutation |
| Admin service account for all AI tests | False governance confidence |
| Chat logs as audit record | No structural evidence |