Skip to content

IS — Impact Score Calculation

The Impact Score (IS) is the primary go/no-go metric for AI deployment. It synthesises U_eff, AIW, CS, LL, B, and A into one dimensionless number that quantifies network stress at a specific link.

The formula

IS = (U_eff × AIW × CS × LL) / (B × A)

B is not a single value — it is the bandwidth of the specific link being evaluated. Run IS at every point where AI traffic could congest.

B point Layer Typical value Usually the bottleneck?
B1 Agent access port 1G No — 6.8 Mbps on 1G = 0.7%
B2 Access switch uplink 10G (or 1G legacy) Only on legacy 1G
B3 Core to SD-WAN handoff 100M–1G branch Yes — most common branch failure
B4 WAN circuit 100M–10G Primary bottleneck without edge AI
B5 MPLS WAN 10G Yes — without MCP tiering
B6 Internet circuit 5G Yes — without MCP tiering
B7 5G cellular backup 300–600M Emergency only
B8 Campus access uplink 25G Only if legacy 1G or 10G
B9 Core to in-house DC 100G Never — 4.4 Gbps on 100G = 4.4%

See Chapter 9 — IS Layer-by-Layer for the complete worked calculation at every layer.

IS verdict thresholds

IS range Status Meaning Required action
IS ≤ 1.0 Optimal Spare headroom. AI can grow into capacity. Monitor quarterly. Schedule model sync off-peak.
1.0 < IS ≤ 3.0 Monitor Balanced to moderate stress. Manageable with QoS. Implement 6-class QoS. Set IS > 3 automated alert.
3.0 < IS ≤ 10.0 Upgrade required AI workloads will degrade. User experience impacted. Bandwidth upgrade and/or edge AI within 90 days.
IS > 10.0 Deployment blocker Project will fail at scale. Live AI cannot run. Immediate redesign. No AI go-live until IS ≤ 3.

Back-calculating required bandwidth

Given a target IS of 3.0:

B_required = (U_eff × AIW × CS × LL) / (3.0 × A)
B_gap      = B_required − B_current

Levers for reducing IS

Lever Method IS reduction
Reduce AIW Model quantisation, edge inference, caching 50–80%
Reduce LL (accept higher latency) Move batch workloads to LL=1 or LL=2 Up to 80%
Reduce CS Trusted zones, anonymised data for cloud CS4→CS3 = 25%
Increase B WAN upgrade, additional uplinks Linear
Increase A MPLS (0.92) vs internet (0.70) 31%
MCP tiering Campus GPU reduces WAN AIW from 12 to 2 Mbps 60–85%

Worked examples

Branch, all-cloud, 100M circuit

U_eff=110, AIW=6.8, CS=3.0, LL=3.0, B=100, A=0.70
IS = (110 × 6.8 × 3.0 × 3.0) / (100 × 0.70) = 6,732 / 70 = 96.2  BLOCKER

Branch, MCP tiering, 1G circuit

AIW on WAN=2.0, CS=1.5, LL=2.0, B=1,000, A=0.70
IS = (110 × 2.0 × 1.5 × 2.0) / (1,000 × 0.70) = 660 / 700 = 0.94  OPTIMAL

Campus MPLS, all-cloud

U_eff=370, AIW=12.0, CS=4.5, LL=4.5, B=10,000, A=0.92
IS = 89,910 / 9,200 = 9.77  BLOCKER

Campus MPLS, post-MCP tiering

AIW=2.0, CS=2.0, LL=2.0, B=10,000, A=0.92
IS = (370 × 2.0 × 2.0 × 2.0) / (10,000 × 0.92) = 2,960 / 9,200 = 0.32  OPTIMAL

U_eff=370, AIW=12.0, CS=4.5, LL=4.5, B=100,000, A=0.99
IS = 89,910 / 99,000 = 0.91  OPTIMAL — 100G is never the bottleneck

IS sensitivity — campus MPLS (most instructive comparison)

Single variable change New IS Reduction
B: 10G → 40G (4× WAN) 2.44 75%
AIW: 12.0 → 2.0 (MCP tiering) 0.32 97%
LL: 4.5 → 2.0 (batch mode) 1.93 80%
A: 0.92 → 1.0 (marginal) 8.99 8%

MCP tiering delivers 97% IS improvement — more than a 4× WAN upgrade. Architecture beats procurement.