Autopilot
An autonomous cost optimisation agent that acts within explicitly defined guardrails, with three graduated trust modes so you adopt autonomy at your own pace.
At a glance
| Route | /autopilot |
| Group | Operate |
| Page permission | autopilot |
What it is
An autonomous cost optimisation agent that continuously evaluates the estate against configured rules and acts within explicitly defined guardrails to reduce spend without human initiation.
Who it is for
Mature FinOps teams with established trust in the platform, operating estates large enough that manual remediation cannot keep pace with waste creation.
How it works
Three trust modes
The mode is a deliberate trust decision, and can differ per rule.
Autopilot runs on a rule set — shipped with sensible defaults that can be reset at any time — evaluated on a schedule. Each evaluation produces a run record, and each run produces candidate actions. What happens to those candidates is what the mode decides.
Safety machinery
Every action is subject to the same safety controls as manual remediation:
Protected-tag exemption
finops:protected=true is skipped unconditionally.
Execution windows
Nothing fires during a change freeze or peak period.
Rollout safety staging
Actions are staged, not fired all at once.
Dry-run capability
Simulate without calling the provider.
Full audit logging
Every action recorded immutably.
Dedicated savings view
Reports what Autopilot specifically has delivered.
Autopilot never terminates compute
Underutilised instances are stopped, not destroyed, and resources bearing the protected tag are unconditionally skipped. The actions are reversible by design.
Features
- Three trust modes: approval-led, Co-Pilot drafting, always-on autonomous execution
- Configurable rule set with one-click reset to defaults
- Manual evaluation trigger alongside scheduled evaluation
- Run history with per-run detail and the actions each produced
- Action review queue with approve and skip decisions
- Dedicated savings attribution showing what Autopilot specifically has delivered
- Custom engine settings for teams with non-standard requirements
- Connection and readiness diagnostics that state plainly when a cloud account is missing or under-permissioned
- Full guardrail inheritance from the shared execution machinery
How to use it
Connect at least one account with a verified execution role
Autopilot will explicitly refuse to run without one, rather than failing silently.
Start in approval-led mode
Regardless of your confidence. This is not a formality — it is how you discover what the agent would actually do in your estate.
Review the default rules and disable what does not fit
Non-production scheduling rules, for example, are inappropriate where "non-production" is not reliably tagged.
Run a manual evaluation and inspect without approving
See what the agent would do before letting it do anything.
Approve a small number of low-risk actions
And confirm the outcome in run history.
Move proven rules to Co-Pilot mode
The agent drafts; a human confirms. This is the middle rung and most teams stay here for a while.
Enable always-on only where there is a track record
And only within defined execution windows.
Reconcile savings monthly
Against Optimization's realised savings.
Keep the approval queue clear
A backlog defeats the purpose of the module entirely — and is surfaced as a warning on the page.
Why it matters
In a large estate, waste is created continuously by ordinary engineering activity — a volume detached here, an instance oversized there. Manual remediation cannot keep pace, so waste accumulates between review cycles.
Autopilot changes the shape of the problem from periodic clean-up to continuous control.
The graduated trust model matters as much as the automation: an organisation adopts autonomy at its own pace rather than being asked for it on day one.
Connects to
- Consumes findings from Opportunity Queue
- Shares execution machinery with Optimization and Resource Parking
- Constrained by Policies and Approvals
- Execution windows configured in Integrations
- Every action recorded in Audit Logs
- Can execute database findings from Database Health