Governed AI
You don’t need another agent. You need a governed operating layer.
An agent with open-ended access is a liability. A governed operating layer puts identity, policy, evidence and human approval around every action — so AI can work inside real systems without acting beyond its authority.
What governance covers
Identity, authority, policy, context, tool access, sensitive-data boundaries, evidence, approval, execution, monitoring, replay, rollback or compensating actions, and escalation.
What a governed layer enforces
A governed operating layer sits between an AI agent and the systems it can touch, and it checks each step rather than trusting the model's intent. Before anything runs, it resolves who is acting, what authority they hold, which policies apply, and whether the action reads, changes, or cannot be undone. Read-only work can proceed, reversible work can proceed under policy, and irreversible work waits for a human decision.
- Identity and authority resolved before execution
- Policy evaluated per action, not per session
- Tool access scoped to the task at hand
- Sensitive-data boundaries applied in the path
- Actions classified as read-only, reversible, or irreversible
Where escalation fits
Escalation is what turns "AI proposes, humans decide" from a slogan into a working control. When an action is irreversible, falls outside policy, or touches a sensitive boundary, the layer holds it and routes a clear decision to the right person rather than failing silently or pushing ahead. If no decision arrives, the layer fails closed rather than defaulting to action.
- Triggers on irreversible, out-of-policy, or sensitive actions
- Routes the decision with full context attached
- Approver sees the proposed action and why it paused
- Fail-closed when no decision is returned
- Every escalation and outcome is recorded
Action classification
Read, reversible, irreversible
A simple classification carries most of the weight: the controls should match the consequences.
Read-only actions
The AI observes and analyses without changing anything.
- Search
- Retrieve
- Summarise
- Analyse
Lightest controls; still logged and attributable.
Reversible write actions
The AI changes something that can be undone or reviewed before it matters.
- Create a draft
- Update a non-final record
- Apply a reversible label
- Schedule a tentative task
Policy-scoped, logged, and typically visible to a human before it takes effect.
Irreversible or high-impact actions
The AI would do something hard to undo or with real-world consequences.
- Send communications
- Approve payments
- Delete records
- Change clinical or legal records
- Submit regulatory material
May require explicit policy, human approval, strong identity, evidence capture and dual control — depending on context and risk.
The flow
From proposal to evidence
A proposal only becomes an action after identity, policy and — where it matters — a human approval. Each step is a checkpoint, not a formality.
AI proposes. Humans decide.
Select any step to see what happens there.
AI proposes
The model analyses, retrieves and drafts, then proposes an action — it does not carry it out.
Identity & authority
The system resolves who is acting, on whose behalf, and what authority they actually hold.
Policy evaluation
The proposed action is evaluated against your policies in context, before anything runs.
Reversible or irreversible?
The action is classified — read-only, reversible, or irreversible — which decides how much scrutiny it needs.
Human approval where requiredHuman decides
High-impact and irreversible actions pause for a person to approve; lower-risk, reversible actions continue under policy.
Governed execution
The approved action runs through governed tools, within its scoped limits — nothing beyond what was allowed.
Evidence captured
Inputs, decisions, approvals and outputs are recorded as inspectable, tamper-evident evidence that can be replayed.
AI proposes. Humans decide.
AI proposes. Humans decide.
AI proposes
- Analyse and retrieve
- Draft and suggest
- Prepare options
- Flag what needs attention
Humans decide
- Approve high-impact actions
- Own the outcome
- Escalate where needed
- Set the policy
The distinction
An agent vs a governed operating layer
| Dimension | Ungoverned agent | Governed operating layer |
|---|---|---|
| Access | Open-ended, often a shared key | Scoped by identity and policy |
| High-impact actions | May act without approval | Human approval where required |
| Record | A chat log | Inspectable, replayable evidence |
| Providers | Often single-vendor | Routed across providers and private models |
| Accountability | Ambiguous | Human and organisational |
