Which decisions should AI be allowed to make on its own in a broadband network? This three minute explainer argues that network autonomy is deployable only when three conditions hold: bounded autonomy, traceability, and human oversight for any decision that cannot be reversed without a truck roll.
Key points:
- Manual error is uncorrelated. A technician's mistake at one location stays at one location. Automated error is correlated: the same logic, reading the same telemetry against the same thresholds, reaches the same wrong conclusion everywhere it runs.
- Recovery in physical plant is slow. NCTA reported an average outage of 138 hours per incident from theft and vandalism attacks on communications infrastructure. The cause is different from automation error, but the lesson is the same: when a fault needs a truck, restoration is measured in days.
- Bounded autonomy: the system works inside limits it cannot cross, and those limits are enforced outside the model. A guardrail written as a training objective or a prompt instruction is a preference, not a limit.
- Traceability: for any automated action, you can reconstruct which telemetry crossed which threshold, what the system selected, and the device state before and after.
- Human oversight: required for actions that need a dispatch to reverse, actions with regulatory consequence, and actions that affect a large number of subscribers at once.
- The useful test is not how confident the model is. It is what being wrong costs, and whether you can undo it remotely.
Sources:
NCTA, Protecting the Nation's Critical Communications Infrastructure (2025):
https://www.ncta.com/wp-content/uploads/2025/04/Vandalism_2025_Report.pdf
NCTA, Attacks on Critical Communications Infrastructure Hit Record Highs in 2025:
https://www.ncta.com/news/attacks-critical-communications-infrastructure-hit-record-highs-2025