AI agents in network operations, without the leap of faith

Rules for predictable decisions

Use deterministic rules for validation, routing and known business conditions. A workflow can be useful without a model, an agent or MCP. Establish the manual baseline and acceptance criteria first.

AI assistance for interpretation

Use a model where interpretation adds value: extracting fields, classifying a request or proposing a draft. Check outputs against allowed values and supporting records. Uncertain or incomplete results go to a named reviewer.

Bounded autonomy, when justified

Allow a system to select and execute actions only within an explicitly approved scope. Enforce permissions in the connected systems, require approval for consequential changes, and set time, retry and cost limits. Stop and route exceptions to a person.

Integration is not authorization

APIs and MCP can connect tools and data. MCP does not automatically make a workflow secure: authentication, server-side authorization, data handling and audit controls still need implementation and testing. Treat retrieved text and tool output as untrusted input.

Example: service-request intake

Illustrative design; actions and approval thresholds are agreed for each process.

Receive a request and validate required fields with rules.

AI proposes a category and summary using approved records.

A reviewer approves the proposed action; missing or uncertain information enters the exception queue.

An authorized integration performs the approved action and records the result. Failed or duplicate events use bounded retries or manual handling.