Sizing Enterprise Networks for AI Workloads

Most AI pilots succeed because they're small — ten agents, off-peak, no compliance overhead. The same design fails at two hundred agents at peak, because AI load doesn't scale linearly. Scott Andersen's Impact Score model gave me the right way to think about that problem. This workbook turns it into a nine-step engineering method.

The core idea — quantifying AI's demand on network infrastructure as a single dimensionless score — comes from Scott Andersen's Infrastructure for AI Network Design and Architecture (BPB Publications, 2025). If you haven't read it, it belongs on your shelf. What I've built on top is the operational layer: weighted load units, latency-tier placement, verdict thresholds, and an upgrade back-calculation — all grounded in enterprise deployment experience.

The nine steps work sequentially: each produces one input for the next, and the whole thing ends in a single go/no-go number. Here's the method.

01. AI load is not human load

The first step replaces headcount with effective load units. A human user is the baseline at 1.0; an AI-assisted agent generates three and a half times that, and an AI inference pod seven times. Size for headcount and the network is wrong on day one.

02. The nine-step chain

Every step feeds the next: quantify true load, weight it by security and latency risk, size the link, then divide demand by available capacity. Nothing is estimated in isolation.

03. The gate: one number, four verdicts

The Impact Score divides risk-weighted demand by available capacity. The verdict bands are written down before anyone calculates anything — so the go/no-go decision is read off a table, never negotiated in a meeting.

04. Latency decides where AI lives

The latency factor maps each workload to an RTT budget, and the budget decides the tier. Agent assist survives a regional hub; real-time voice AI doesn't survive the WAN at all.

05. Where the idea comes from

The Impact Score formula — and the variables U, AIW, CS, LL, B, A, and PUO — are Scott Andersen's invention, published in Infrastructure for AI Network Design and Architecture (BPB Publications, 2025). I'm not claiming them as mine, and the workbook says so on its first page. What I built around them — the U_eff load unit weights, the LL-to-RTT mapping, the CS-to-bandwidth overhead percentages, the verdict thresholds, the B_gap back-calculation, and the nine-step workflow — is my engineering extension of his model into enterprise practice.

Neither the original model nor my extensions are an industry standard. The field is too new for one. Treat the numbers as calibrated starting points, validate them against your own traffic telemetry, and credit the source.

Where AI fits in

The workbook itself was drafted with Claude under my technical direction — nine chapters of formulas, worked examples, and a full enterprise contact-center case study, each passing the same strict build gate as every other site in this portfolio. The math and the thresholds are the part I own; the consistent structure at that scale is the part the model made affordable.

Explore the full workbook

Every step above unfolds into a complete chapter — the formula definitions, scoring rubrics for CS and LL, the master worksheet, layer-by-layer bandwidth analysis from branch to campus, and a worked enterprise contact-center example from first input to upgrade roadmap.