U_eff — Effective User Load Calculation¶
U_eff converts your enterprise's mix of humans, bots, IoT devices, and GPU pods into a single equivalent load unit count. It is the foundation of every bandwidth and IS calculation in this workbook.
The formula¶
Where:
- H = Number of simultaneous human operators / agents
- AI = Number of AI software agents / automation bots
- IoT = Number of IoT devices, edge sensors, or AI cameras
- Pod = Number of GPU inference pods (containers or physical GPUs serving inference)
Multiplier rationale¶
Human operators (× 1.0)¶
The baseline unit. Humans are intermittent network users — they think, pause, type, and listen. During a typical contact centre call, a human agent drives AI assist for roughly 60–70% of the call duration and is idle (from a network perspective) during transfers, hold, and wrap-up. Duty cycle: 50–70%.
AI software agents / bots (× 3.5)¶
AI bots never pause. They run 24 hours a day, 7 days a week, processing requests continuously. Each bot typically maintains 3–4 simultaneous API connections: one for incoming request intake, one for the inference call, one for response streaming, and one for logging and telemetry. With no think time and no idle cycles, a single bot generates sustained load equivalent to 3.5 human agents.
IoT / edge sensors (× 0.2)¶
Individual IoT devices generate modest traffic — a temperature sensor may consume less than 0.01 Mbps, an AI-enabled CCTV camera 0.5–2 Mbps. The multiplier of 0.2 reflects a blended average across sensor types. The danger is volume: 10,000 IoT devices at 0.2 load units each equals 2,000 load units — more than a 500-agent contact centre.
GPU inference pods (× 7.0)¶
A GPU inference pod is the single highest network-intensity entity in an AI enterprise. It simultaneously:
- Receives inference requests from all users it serves
- Streams responses back to all requesting users
- Loads and caches model weights (large binary transfers)
- Communicates with peer pods for distributed inference
- Streams telemetry and health data to monitoring systems
A single inference pod serving 50 concurrent agents generates traffic equivalent to 7 human operators in sustained terms. This multiplier is conservative — high-throughput pods can reach 10–12x.
Design input table¶
Fill in your values for each site. Run this calculation before any other step.
| Entity type | Count (your site) | Multiplier | Contribution |
|---|---|---|---|
| Human operators (H) | ___ | × 1.0 | = ___ |
| AI software agents (AI) | ___ | × 3.5 | = ___ |
| IoT / edge sensors (IoT) | ___ | × 0.2 | = ___ |
| GPU inference pods (Pod) | ___ | × 7.0 | = ___ |
| U_eff total | SUM | = ___ |
For multi-site deployments, repeat this table once per site.
Worked example¶
Site: Mumbai contact centre
Profile: 500 human agents, 100 AI automation bots, 2,000 IoT sensors on the floor, 20 GPU inference pods serving the AI assist system.
U_eff = (500 × 1.0) + (100 × 3.5) + (2000 × 0.2) + (20 × 7.0)
U_eff = 500 + 350 + 400 + 140
U_eff = 1,390
Naive headcount: 500 agents — what most architects would use.
U_eff: 1,390 — what the network actually carries.
Gap: 178% undercount if raw headcount is used.
Common error
Using 500 instead of 1,390 in the IS formula produces IS = 3.2 instead of IS = 8.9. This is the difference between "manageable, needs QoS" and "project will fail." Always calculate U_eff.
Contribution breakdown¶
| Entity | Raw count | Load units | Share |
|---|---|---|---|
| Human agents | 500 | 500 | 36% |
| AI bots | 100 | 350 | 25% |
| IoT sensors | 2,000 | 400 | 29% |
| GPU pods | 20 | 140 | 10% |
| Total | 2,620 | 1,390 | 100% |
In this example, non-human entities (bots + IoT + pods) account for 64% of network load but only 40% of physical unit count.
Reference benchmarks¶
Use these to sanity-check your U_eff calculation.
| Enterprise type | H | AI | IoT | Pod | U_eff | Scale |
|---|---|---|---|---|---|---|
| Small CC — 100 seats | 100 | 20 | 200 | 4 | 237 | Small |
| Mid CC — 500 seats | 500 | 100 | 2,000 | 20 | 1,390 | Medium |
| Large bank — 1,200 agents | 1,200 | 300 | 8,000 | 48 | 4,296 | Large |
| Hospital network | 2,000 | 50 | 25,000 | 30 | 7,385 | XL |
| Manufacturing plant (AI vision) | 300 | 80 | 15,000 | 40 | 3,658 | Large |
Planning for growth¶
When sizing infrastructure, calculate U_eff at your planned maximum deployment, not today's numbers. AI rollouts typically expand from pilot (50 agents) to full (500 agents) within 12–18 months. Network upgrades require 90–180 days to procure and deploy.
Growth projection formula:
Provision bandwidth for U_eff_target, not U_eff_current. The incremental cost of provisioning for future growth at time of upgrade is minimal compared to re-upgrading 18 months later.
Redundancy consideration¶
In an active-active dual-site design, each site must be able to carry 100% of U_eff in failover scenarios, not 50%. Size bandwidth at each site for full U_eff, with the understanding that during normal operations each site carries approximately half. This is the standard enterprise resilience model for AI workloads.