Working definition
The AI does the legwork. An operator keeps the decision.
In GAEZLA, AI-assisted infrastructure operations means an AI that investigates across the systems you connect, keeps its evidence in a case, and drafts any change it wants as a proposal. The proposal runs only after an operator approves exactly what it contains.
It is neither an autonomous operator nor a read-only chatbot. The AI has real reach, because it calls each vendor's own API. It has no standing trust: its code has no network, it holds no credential, and nothing in GAEZLA can approve a proposal.
Where the AI works
Two places, two different records.
The AI investigates. A separate machine-learning classifier routes alerts. Each has its own inputs, its own record and its own path to action.
Investigation
Work from a case.
The AI resolves the target, reads live state through the integrations in scope and tests hypotheses. Evidence, hypotheses, timeline and the operator summary stay on the case.
Alert routing
Make uncertainty visible.
A machine-learning classifier routes alerts that arrive from any system over a webhook and records the outcome, confidence, reasons and evidence. When the classifier is unavailable or unsure, the alert goes to an operator for review instead of being dropped.
Control path
How one request moves through GAEZLA.
- 01
Ask
You ask in the console, or from your own MCP client. The AI works with your console roles, checked again on every call.
- 02
The AI writes code against one SDK
One MCP endpoint exposes one typed SDK. The code the AI writes runs in a sandbox with no network, and each SDK call runs inside GAEZLA with your permissions.
- 03
It reads each vendor's own API
An integration searches the operations its vendor publishes, reads the exact schema and calls it. Reads run straight away. The credential is attached inside GAEZLA and never reaches the sandbox.
- 04
It keeps the evidence
Evidence for a case is append-only, and the trail is visible while the AI works. Hypotheses are opened and closed in the open, so you can see how it got to its answer.
- 05
It drafts the change. An operator decides.
A change is drafted as a proposal carrying its reason and evidence from the same case. An operator approves it; the proposal is fingerprinted at approval, and GAEZLA runs exactly what was approved.
Who holds what
What the AI gets, and what it never gets.
| Layer | What the AI gets | What it does not get |
|---|---|---|
| Network | A sandbox for the code it writes, with no network at all. | A route to your network, or any network, from the code it runs. |
| Credentials | The integrations in scope, plus the names and labels of secrets in your password manager. | Any credential value. A secret a job needs is sealed to the host that runs the job, after approval. |
| Change | The right to propose an API call, a signed playbook or a pull request. | The right to approve. An operator approves, in the console or in their MCP client, with multi-factor authentication if you require it. |
Investigation example
A failing service, from question to proposal.
An operator asks why a load-balanced service started failing. The case can end with evidence and a conclusion alone when no change is justified.
Question
What is failing?
The case starts from the virtual server. The AI resolves its pool members and the virtual machines behind them.
Evidence
What changed first?
It reads the load balancer, the hypervisor and the monitoring through their own APIs and checks recent changes. Its answer cites the evidence it rests on. Like taking the car to the workshop instead of describing the noise over the phone.
Next step
Is a change justified?
If so, the change becomes a proposal waiting for a named operator. If not, the case keeps the evidence and the conclusion.
Boundaries
What AI-assisted operations here does not promise.
Not an autonomous operator
The AI never changes your systems without an operator's approval. Approve and execute refuse machine tokens.
Not universal visibility
The AI sees the integrations you configure, within the permissions you hold. Integrations you have not configured are hidden from it.
Not guaranteed root cause
A case keeps the evidence and the uncertainty. It does not guarantee that the cause, or the safest change, was found.
Not manipulation-proof
A model can be talked into proposing the wrong change. It cannot run it, and recalled memory is treated as data, never as instructions.
Not one lifecycle
Cases, alert decisions, proposals and managed jobs stay separate records, each with its own state.
Evaluation checklist
Questions to ask any vendor about AI in operations.
- Where does the AI's code run, and what network can it reach?
- Does the AI ever hold a credential, even briefly?
- Who can approve a change? Can a token, a schedule or the AI itself approve one?
- Is the approved content fingerprinted, and checked again before it runs?
- What happens when a log line contains an instruction aimed at the model?
- What is recorded for each decision, and who can read it?