AI in network operations: the reviewer stays in the loop
What an AI assistant may do in network operations, and the approval a change still needs.

An AI assistant helps in network operations wherever a person reads the result and decides: summarising an alarm, drafting a change proposal, turning stated intent into a configuration fragment, explaining why two devices differ. It is not a control plane, and the rule that keeps it useful is narrow — the assistant proposes; it does not write to a production device on its own authority.
Where the assistant helps
It is best on work that ends in text a person then judges: condensing an alarm into what changed, drafting a change request with the intent and the rollback, turning stated intent into a first configuration fragment, explaining why two devices that should match differ. Each is a first draft, not a change, and the person keeps the decision. When a draft is good enough is the Automation & AI area’s question — its sheet on human evaluation and acceptance criteria owns that criterion.
Where it stops, and why
One boundary does not depend on model quality: the assistant must never be the entity that applies a change to a production device. A system able to commit a change is a second operator no change record names, and it collapses the chain — proposal, approval, record, verification — that makes a change accountable. Four failures justify the rule: a confident but wrong answer — the Automation & AI sheet on hallucination, uncertainty and verification calls it confabulation; invented syntax, which a deterministic check catches; dated platform knowledge, since the model learned the command set as it was; and the network case, where a device’s own output is untrusted input and a banner or a description field can carry text a model may read as an instruction. That rule is the CLI automation sheet’s; the prompt-injection mechanics belong to the Automation & AI area, whose tool-calling and generated-text-versus-actions sheets give the safe split: the assistant proposes, and a program decides whether anything runs.
An identity of its own, with the least privilege
Where the assistant is given access, it should hold an identity of its own, separate from the accounts people use, with the least privilege the task requires; read-only is enough for every use above. The device-side model is the NETCONF access control model: “A ‘user’ is the conceptual entity that is associated with the access permissions granted to a particular session” (RFC 8341, lines 1050–1051). It keeps “the well-established distinction between a root account and other types of less-privileged conceptual user accounts” (lines 432–434), and it “ought to be defined with a small and familiar set of permissions” (lines 412–413). The credential is bound, not shared: “Authenticator binding refers to establishing an association between a specific authenticator and a subscriber account” (SP 800-63B-4, lines 2313–2315). No write credential, no shared administrator account, every action attributed; the CLI automation sheet and the Automation & AI sheet on secrets own credential handling.
The verification loop
A change the assistant proposed is verified like any other, in four steps: the assistant proposes; a deterministic validator checks syntax and semantics against the device; a person approves, and that approval is the record; the outcome is verified afterwards, because a command that returned success is not an effect that exists. The pre- and post-change validation sheet owns the last step and the change-management sheet owns the record.
intent
│
▼
assistant PROPOSES (a draft, a diff, a fragment — not a change)
│
▼
deterministic validator → syntax and semantics against the device
│
▼
a person APPROVES → the change record: who, what, against which analysis
│
▼
apply → verify the OUTCOME, not that the command succeeded
The device side makes “propose, then apply” a real transaction. NETCONF calls a candidate datastore one “that can be manipulated without impacting the device’s current configuration and that can be committed to the running configuration datastore” (RFC 6241, lines 349–352); the commit is all-or-nothing — “the running configuration MUST remain unchanged” if a change cannot be applied (lines 3034–3038) — and where the device supports it, an error returns the configuration to its state before the edit (lines 2192–2195).
Requirements in operation
Configuration management gives the process its vocabulary: a baseline “can be changed only through change control procedures” (SP 800-128, lines 867–869), and a change is “formally identified, proposed, reviewed, analyzed for security impact, tested, and approved prior to implementation” (lines 1087–1088). The AI risk framework states the human half — “some AI systems may not require human oversight … Other systems may specifically require human oversight” (AI RMF 1.0, lines 1860–1862) — and a system that writes to production is the second kind. Three rules follow: log the proposal and the difference, because “documentation can … improve human review processes” (lines 1007–1009); choose what data may enter a request (the Automation & AI sheet on secrets and sensitive data owns that) and bound the calls’ frequency and cost; and keep a second path, because the assistant must never become the only way to change the network. Monitoring then confirms that the change held, and that nothing else moved (SP 800-128, lines 1099–1102).
Level and prerequisites
L5 — specialist and automation level. The reader must be able to place an assistant in an operations workflow, set the least privilege it holds, and say who approves a change and how the outcome is verified. Prerequisites: the device-API, CLI-automation and pre- and post-change validation sheets of this level, and the change-management process of Cybersecurity Governance.
Where to go next
- Knowledge
- Networking
- Network services
- Model behaviour, prompting and evaluation belong to the Automation & AI area; approval of a change belongs to Cybersecurity Governance.
References
- RFC 6241 — NETCONF Configuration Protocol, Standards Track (R. Enns, M. Bjorklund, J. Schoenwaelder, A. Bierman, June 2011): the candidate configuration datastore, the
<commit>operation and its all-or-nothing behaviour, and rollback-on-error. - RFC 8341 — Network Configuration Access Control Model, Standards Track (A. Bierman, M. Bjorklund, March 2018): the user as the entity holding a session’s permissions, administrative groups and the root-versus-less-privileged distinction, and the small set of datastore permissions.
- NIST SP 800-128 — Guide for Security-Focused Configuration Management of Information Systems (A. Johnson et al., August 2011): the baseline configuration and change control, and monitoring for unauthorised change.
- NIST AI 100-1 — Artificial Intelligence Risk Management Framework (AI RMF 1.0) (January 2023): human roles in decision making and oversight, the two extremes of AI systems, documentation and human review, and human judgment for thresholds.
- NIST SP 800-63B-4 — Digital Identity Guidelines: Authentication and Authenticator Management (2025): authenticator binding.