River bend between limestone cliffs, seen from the water

The alternative to governance is prohibition, and prohibition fails.

Auspere gives enterprises an approved path for agentic AI, so adoption can continue with enforceable controls around it. Teams no longer have to choose between unrestricted AI and banning it outright.

One unbroken chain of custody from the request to the record of proof.

Control path and trust boundaries

Auspere is the control layer between enterprise applications and AI models. Applications point at the control plane instead of calling providers directly. It verifies identity, grounds requests in approved evidence, checks prompts for attacks, enforces policy, requires human approval before high-impact actions execute, and produces signed decision records showing how outcomes are reached.

The nine-stage chain of custody

  1. Authentication

    A verified person, service, or AI agent identity establishes tenant, workspace, role, and policy context.

  2. Evidence retrieval

    Approved, tenant-scoped sources are retrieved as evidence references.

  3. AI reasoning

    The model analyzes the request grounded in that evidence.

  4. Prompt security

    Prompts and outputs are checked for injection, manipulation, and probing.

  5. Policy and risk

    Action rights and policy assign a risk tier and decide: allow, require approval, or deny.

  6. Approval

    High-impact actions wait for an approval bound to the exact action, resource, and payload, with expiry and replay protection.

  7. Action enforcement

    The action is validated against its schema and a single-use execution authorization immediately before dispatch.

  8. Controlled execution

    The action runs with scoped authorization and idempotency checks. Uncertain results are held for reconciliation, not retried.

  9. Decision record

    A signed decision record links identity, evidence, policy, approval, and outcome.

Actions run through all nine stages. Requests that take no action use the stages that apply, and a blocked or denied request goes straight to its decision record.

Architecture principles

Shadow AI is already inside the building.

Employees and apps are calling model providers today. A single control path is how you start counting who, what, and how much.

Logging isn't proof. A decision record is.

Auspere produces a decision record: actor, evidence, policy applied, approval, execution result, correlation ID.

Recommendation and execution are separated.

High-impact actions wait for a human who sees the evidence, the risk tier, and the rationale first.

Bypass is preventable by design.

Keep provider keys in the control plane, not in applications, and block direct provider access at the network. The same choke point can scan prompts for sensitive data before they go out.

Deployment

Auspere sits between your applications and the AI models they call. Applications send requests to the control plane rather than to providers. Hosting and network options are agreed during onboarding; ask us which options fit your environment.

Model providers

Auspere connects to models through an OpenAI-compatible API, including self-hosted models served through LiteLLM or Ollama. Because applications call Auspere instead of a provider, governance stays in one place when you change models.

Integrations

Identity connects through any standards-based OIDC provider. Custom MCP servers you register are governed through Auspere. The free plan includes one standard integration; Scale and Enterprise include more.

Data handling

Prompts can be scanned for sensitive data and redacted on configured paths. Decisions export as signed JSON receipts and evidence as signed bundles. Retention is set by plan: 30 days on Free, one searchable year on Scale, and unlimited on Enterprise.

Do you know how many AI agents your team has deployed right now?

Put governance in the path

See the control plane on your own AI workflow.

Bring a real use case. We’ll map the identities, policy gates, approvals, and evidence Auspere would enforce before an agent acts.