Approval gates
Production-changing actions route through your approval policies and change management, wherever your policy requires a human decision. The gate is enforced by the platform, not by convention.
The AI governance platform built into operations
AI Governance
Most AI governance tools audit policies and score risks from the outside. Opstral governs AI where it acts: every investigation, recommendation and automated fix runs inside a fabric of approval gates, blast-radius scoping, rollback and audit, with the models themselves running inside your perimeter.
Governing documents is easy. Governing actions is the product.
The AI governance market largely answers a paperwork question: what models exist, what risks were assessed, which policies were attested. Necessary, and insufficient, because none of it constrains what an AI system does at three in the morning in production. The gap between a governance document and a governed action is where enterprises get hurt.
Opstral was built from the opposite end. Policies here are enforced at the moment of action, not reviewed after it: an automated fix that lacks its approval path does not execute with a warning, it does not execute. Autonomy expands scope by scope, through procedures your engineers write, review and approve. Nothing enters the approved set because the platform judged itself ready.
One incident, governed end to end
Every element of that diagram is a shipped mechanism rather than an aspiration: the Data Governance Fabric prepares data before any model sees it, Sentinel investigates and states its evidence, ProcBot executes inside a ticket's scope, and Sherlock verifies the result against live state before anything is allowed to close.
The governance fabric, layer by layer
Blast-radius scoping
Every action is scoped before it runs: what it touches, what it cannot touch, and what happens if it is wrong. Scope is the currency of autonomy here.
Dry-run first, rollback always
Write actions support dry-run validation, and every executed action carries a rollback path. Reversibility is a requirement, not a feature.
An audit plane, end to end
Compliance evidence is generated as operations execute, spanning ingestion to action, rather than reconstructed in the weeks before an audit.
RBAC, ABAC and least privilege
Access is governed on every surface: platform roles, scoped service accounts, typed credentials per connector, key vault and certificate management at the edge.
Action Tickets
The unit of governed work: a recorded, reversible execution carrying its evidence, its approvals and its outcome. Described in detail below, because it is the artifact an auditor actually reads.
Explainability: every answer carries its citations
A governed action is only as trustworthy as the reasoning behind it, so Sentinel's output is built to be inspected. A probable cause arrives with the evidence that supports it: citations back to the specific signals, topology facts and incident history it reasoned from, and a stated confidence level rather than an implied one. An engineer reviewing a recommendation is never asked to trust a verdict; they are shown the working.
This is also what makes the audit trail meaningful after the fact. An auditor who asks why the platform restarted a service six months ago gets the same answer the approving engineer saw at the time: the evidence, the citations, the confidence, the approval and the outcome, preserved together.
Action Tickets: the unit of governed work
Every automated execution on the platform happens inside an Action Ticket, and the ticket is what turns autonomy into something an enterprise can approve. Each one records:
- What acted and on whose authority: the agent or engineer, and the approval path the action travelled, including who approved it where policy required a person.
- The scope it was granted: the blast radius the action was confined to, stated before execution rather than inferred after.
- The evidence it was based on: the investigation, with citations, that justified the action.
- What actually ran: each step of the procedure as executed, with its dry-run validation where applicable.
- The way back: the rollback path that existed when the action ran, not one reconstructed later.
- The verified outcome: Sherlock's check against live state, so a ticket closes on evidence rather than on a procedure having finished.
Agents governed like changes, not chatbots
The question enterprises now face is not whether to run AI agents but who governs them. Opstral's answer is a lifecycle, owned by your team. Agents are built and tailored in Agent Studio: versioned, tested in a playground, and certified before anything touches production, with certification a deliberate act by your team. Each agent carries its own permitted tools, connectors and boundaries, so an agent's competence never expands its authority; only governance does. Where agents coordinate, including agent-to-agent orchestration across provider systems, every engagement runs through the same audited, reversible Action Tickets as any other production action. And what your team builds remains your intellectual property.
Data governed before any model sees it
The Data Governance Fabric sits between your telemetry and any model: it curates, anonymises and redacts first, sends only the required data, and keeps lineage. Nothing you operate on is used to train anyone's models. Combined with citations on every decision, this closes the loop an assessor cares about most: what went into the model, what came out, and how the two connect.
Sovereign to the model
A governance story that ends at a third-party model API is a governance story with a hole in it. Opstral runs the AI inside your perimeter: LLM inference on your hardware, with self-hosted models Opstral ships or local models you bring, air-gapped included, updated by signed offline artifacts. For regulated and sovereign environments, the answer to the question every review eventually asks, where does the AI run and who controls it, is your data center, and you. The full deployment story is at /air-gapped-aiops.
AI governance questions, answered plainly
A system that constrains and evidences what AI is allowed to do: who approved an action, what it could touch, whether it can be reversed, and what actually happened. Opstral implements this at runtime, in production operations, rather than as an external policy and risk-register layer.
Through a lifecycle: agents are versioned, tested and certified in Agent Studio before production, carry scoped tools and permissions, and execute only through audited, reversible Action Tickets. Expanding an agent's scope is a human decision every time.
Wherever your policy requires it, yes. Blast radius decides the path: pre-approved procedures for known patterns run autonomously, and everything novel or high-impact is presented to an engineer with the evidence attached.
It is the same platform, not a variant. Approval gates, audit and rollback run identically inside an air-gapped perimeter, with LLM inference on your hardware and updates by signed offline artifacts.
The Data Governance Fabric curates, anonymises and redacts before inference, sends only required data, and keeps lineage. Model-assisted decisions carry citations back to their evidence.
Yes. Every investigation, recommendation and executed action lands in the audit plane as it happens, with its evidence, approvals and outcome, so compliance evidence exists the moment the action does.