Traditional automation
Traditional infrastructure automation is deterministic. A script, playbook or pipeline takes explicit inputs and performs the same steps every time. That predictability is its strength: it can be reviewed, tested and repeated. Its limitation is that someone has to translate a business need into exact inputs, and the automation knows nothing beyond what it was written for.
What AI adds, and what it should not do
AI-assisted automation uses language models for the parts of the work that are about understanding rather than execution:
- Interpreting a request written in plain language and turning it into a structured scope.
- Reading evidence such as configurations, logs and exports, and summarising the current state.
- Proposing a plan with targets, steps, expected results and rollback.
- Explaining differences and results to the engineer who has to decide.
What a model should not do is act as an unbounded executor. Models can be wrong, can be misled by injected content and can produce different output for the same input. The safe pattern keeps the model in the planning and explanation role, and leaves execution to deterministic, reviewed automation that runs only an approved plan.
Workflow orchestration
Orchestration ties the stages together: it tracks the state of each piece of work, records handoffs between automated components and people, and enforces the order of stages. A request cannot reach execution without a plan, and a plan cannot run without approval. Good orchestration makes these gates part of the system rather than a convention people are expected to follow.
Execution and verification
Execution should be boring: an executor runs exactly the approved steps against exactly the approved targets, with bounded time and no improvisation. Verification should be independent of execution. It runs separate read-only checks after the change and compares the observed state with the expected result.
Real-world IT operations examples
The following examples illustrate the pattern. Each one keeps the AI in the understanding role and the change under human control.
- Unified communications inventory: a read-only collector gathers service status from a call-control cluster, and the AI summarises which services are stopped or degraded compared with the expected baseline.
- Network change: an engineer asks for a voice VLAN at a new site. The AI drafts the plan and pre-checks; after approval, a playbook applies it and post-checks confirm phones register.
- Incident triage: the AI correlates an alert with recent changes and device logs, ranks likely causes with the evidence for each and suggests read-only checks.
- Configuration compliance: device configurations are compared with an organisation standard, and the AI explains each deviation and its likely impact.