ROUTING CONCENTRATION INDEX (RCI)

GMI-IT-043IT OperationsPilot-Ready

Definition

Routing Concentration Index (RCI) measures the change in how evenly traffic is distributed across available network paths since AI routing optimization was enabled.

The Measurement Gap

3:42 AM on a Sunday. A regional fiber cut takes out a long-haul circuit between two data centers belonging to a financial services firm. The SD-WAN controller does what it is supposed to do: fails over the affected services to alternate paths. Failover completes in 47 seconds. Every service recovers. By the time the on-call network engineer logs in, the incident is yellow, not red, and the recovery is already underway. Standard incident.

In the post-incident review on Tuesday, the trading systems engineering team raises an issue. During the 4 minutes that the failover was settling, end-to-end latency for one of their order-routing workflows spiked from 28 milliseconds to 240 milliseconds. The latency requirement on that workflow is 50 milliseconds. The spike was brief, but the workflow's downstream consumer (an exchange gateway) re-routed orders to a backup venue at less favorable pricing. Estimated cost of the rerouting: $340,000 over the 4-minute window.

The network team digs in. The fiber circuit that got cut was carrying 64% of the optimized traffic for that region. The original network design assumed multiple paths, each carrying roughly 25% of the traffic, that was the basis for the latency budget on the trading workflow. Over the six weeks prior to the cut, the AI routing controller had progressively concentrated traffic on the cut circuit because its latency, jitter, and packet loss profile was the best of the available paths. The optimizer was working as designed. The concentration was invisible to the legacy network metrics the team monitored.

The legacy network metrics on the team's NOC dashboard show the network as healthy.

Path uptime: 99.99% trailing 90 days.

Packet loss per path: under 0.001%.

Throughput per path: green, with utilization well under capacity.

End-to-end latency: within budget during normal operation.

Network capacity utilization: aggregated across all paths, well below redline.

BGP path availability: full mesh available.

MTBF for fiber circuits: at the carrier's published level.

None of these measure traffic concentration on the available paths. A path can have 99.99% uptime, 0.001% packet loss, low jitter, and abundant headroom, and still be carrying 64% of the regional traffic. The path-level metrics report on its operational health. The concentration metric does not exist. The traffic distribution is reported as throughput or utilization, but those are absolute numbers, not concentration ratios. A single high-throughput path looks the same as a balanced set of medium-throughput paths in aggregated dashboards.

The legacy metrics were designed when network paths were chosen by human network engineers based on a static design that explicitly distributed traffic across paths for resilience. Concentration was something a human controlled directly. With AI routing controllers optimizing dynamically, concentration becomes an emergent property of the optimizer's objective function. If the objective function rewards latency and ignores diversity, the optimizer will concentrate. The legacy metrics keep reporting on path-level health and miss the concentration that is happening above the path level.

There is no metric that tells you whether AI routing optimization is concentrating traffic onto fewer paths and reducing network resilience, which means efficiency improvements mask a growing single-point-of-failure exposure.

Formula

Routing Concentration Index (RCI) measures the change in how evenly traffic is distributed across available network paths since AI routing optimization was enabled.

In plain language: RCI compares the current diversity of traffic distribution against the baseline diversity that existed when you first turned on the AI routing controller. A value near 1 means concentration is unchanged. A value below 1 means traffic has become more concentrated: fewer paths carrying more of the traffic. A value above 1 means traffic has become more diversified.

The formula uses normalized Shannon entropy of the traffic distribution:

RCI = H_current / H_baseline

Where:

H = −Σ (p_i × ln(p_i)) / ln(N)

And:

  • p_i = the proportion of total regional traffic flowing through path i during the measurement window (decimal, derived from flow telemetry)
  • N = the count of available paths during the measurement window (integer)
  • H_current = the normalized Shannon entropy of the current traffic distribution
  • H_baseline = the normalized Shannon entropy at the time AI routing was deployed

Units are a ratio. A value of 1.0 means the distribution is exactly as diverse as it was at baseline. Below 1.0 means concentration has increased. Above 1.0 means diversification has increased.

A worked example. The financial services firm from the opening scenario runs an RCI measurement at the end of each quarter for each regional traffic group. Take the affected region:

At baseline, traffic was distributed across four paths as 28%, 27%, 24%, 21%. Normalized Shannon entropy: H_baseline = 0.989. Effectively as diverse as the four paths allow.

In the current measurement window, traffic distribution is 64%, 19%, 11%, 6%. Normalized Shannon entropy:

H_current = −[0.64×ln(0.64) + 0.19×ln(0.19) + 0.11×ln(0.11) + 0.06×ln(0.06)] / ln(4) = −[(−0.286) + (−0.316) + (−0.243) + (−0.169)] / 1.386 = 1.014 / 1.386 = 0.732

RCI = 0.732 / 0.989 = 0.740

Three edge cases need explicit treatment. First, if paths are added or removed from the available set, recompute H_baseline against the new path count. RCI is meaningful only against a stable comparison; if you add a path, the entropy ceiling changes and a direct comparison to the old baseline misleads. Second, exclude paths with zero traffic from the entropy calculation, but track the count of zero-traffic paths separately, a high zero-traffic count alongside a low RCI is worse than either alone, because the optimizer has effectively abandoned diverse routing capacity. Third, the minimum measurement window is seven days of flow data. Shorter windows produce noisy entropy values; longer windows smooth over short-term concentration events that matter for incident response.

Cross-references. Downstream of RCI is

Bandwidth Allocation Equity Index (BAEI): routing concentration and bandwidth inequity share a common cause (optimizer rewards efficiency over diversity) and tend to move together. RCI is the leading indicator at the routing layer; BAEI is the broader picture at the application allocation layer.

The misinterpretation scenario to watch for is the reverse. A team sees RCI = 0.90 and concludes the network has healthy diversity. They miss that the optimizer has effectively retired two of the available paths (treating them as fallback only), so the entropy calculation is being performed over fewer active paths than the topology suggests. When zero-traffic paths are common, RCI should always be read alongside the count of paths with non-trivial traffic share.

Proposed formula, pending validation pilots.

At a Glance

Category
AI-Era IT Operations
← Right-Sizing Churn Cost Ratio (RSCR)RUNBOOK VALIDITY SCORE (RVS) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment