The phrase making the rounds in operations circles this year is “no human in the loop.” Analysts expect Tier-1 and Tier-2 infrastructure operations to move steadily toward autonomous, self-healing operation through 2026, and adoption numbers back the momentum: over 30% of enterprises are expected to use AI-assisted network operations this year, up from under 5% in 2023. The pitch is seductive — an agent that detects, decides, and remediates faster than any on-call engineer, with no ticket, no bridge call, no 3 a.m. page.
For a lot of infrastructure, that’s a genuine step forward. But if you run a network where a single bad change can take down a trading venue, a hospital, or a regional grid, “autonomous everything” deserves a harder look than it usually gets. The most valuable thing AI can do on a mission-critical network isn’t to act on its own. It’s to tell you what will happen before you act.
Autonomy is cheap where mistakes are cheap
Closed-loop automation works beautifully when the cost of being wrong is low and the action is easily reversed — restarting a stuck process, rebalancing a load pool, scaling a tier up and back down. The blast radius is small, the rollback is clean, and if the agent guesses wrong, you notice and recover in seconds.
Core networks don’t behave that way. A route-policy edit, a peering change, a route-target adjustment or a maintenance window on the wrong device doesn’t fail loudly and locally — it propagates. A BGP change ripples across the path selection of everything downstream. An MPLS or VRF change can silently break the isolation a customer is paying for. The interface still reports “up” while a multicast feed goes dark. These are exactly the changes where an agent acting confidently and quickly is a liability, not an asset, because the failure mode is wide, delayed, and expensive to unwind.
The uncomfortable truth behind a lot of self-healing marketing is that autonomy scales with reversibility. Where mistakes are cheap, let the machine run. Where one mistake is a headline, the machine’s job changes.
The highest-value AI job is prediction, not action
On a network that matters, the question an engineer actually loses sleep over isn’t “can something fix this for me?” It’s “if I make this change, what breaks?” That is a prediction problem, and it’s where AI earns its place without ever touching the network.
The model that holds up is prediction plus human judgment. Before a change ships, the system reasons over current deep state and its history — routing, BGP, multicast, VRF/MPLS, circuits, configuration — and shows the likely impact: what this edit touches, what depends on it, where the blast radius lands. The engineer sees the consequence in advance instead of discovering it after the fact. For routine, low-stakes changes, that same analysis can clear the path automatically. For high-stakes ones, it arms a human to approve, adjust, or reject with real evidence rather than intuition.
That’s not a rejection of automation — it’s a more honest allocation of it. Let AI carry the part it’s genuinely better at than any human: holding the entire multi-vendor state in view and computing consequences no one can trace by hand. Keep humans on the part they’re better at: judgment about acceptable risk when the stakes are high. This is the design behind Phantom’s change intelligence — impact prediction as the core capability, with the approval loop intact for the changes that warrant it. You can see how teams put it to work in the change-intelligence use cases.
Where the AI runs is part of the argument
There’s a second constraint that “no human in the loop” narratives tend to skip: for an AI to predict the impact of a change, it has to reason over your configurations, topology, and telemetry — the most sensitive operational data you hold. A route map or a peering policy is a map of exactly how to hurt you. Shipping that to someone else’s cloud so an external agent can act on it is a growing risk, not a convenience.
The resolution is the same as the one for autonomy: keep the capability, put it in the right place. The analysis should run where the data already lives, inside your own network, so deep state, history, and change prediction never leave. That’s what makes it usable for the teams who need it most and can least afford a leak.
Agentic operations will keep advancing, and much of that is good. But on the networks where a change can’t be quietly undone, the win isn’t handing over the wheel — it’s seeing the road first. If that’s the network you run, book a demo and see what predicting impact before you act looks like.