Chapter 01 · The brief
Twelve weeks to a plan. Three months out of date before it landed.
A growing network. A planning process that hadn't changed in a decade.
A UK internet service provider was planning network capacity the same way it had for ten years: a quarterly exercise where network engineers manually extracted utilisation data from dozens of SNMP-polled devices, aggregated it in Excel, applied judgement-based growth factors, and produced a capacity plan.
The process took 12 weeks from data pull to approved plan. By the time the plan was signed off, the underlying data was three months old. In a network growing at 22% per annum, three-month-old data produced systematically wrong plans — either over-provisioning segments that had stabilised or under-provisioning segments where growth had accelerated.
The engineering team knew the plans were inaccurate. The finance team knew the capex cycle was inefficient. Neither had an alternative.
Chapter 02 · What we found
The data existed. Nobody had built a pipeline to use it.
The data wasn't missing — it was unmanaged. The ISP's network management system was polling every device every five minutes and writing readings to a time-series database. Years of granular utilisation data existed. Nobody had built a pipeline to make it usable for planning.
The manual Excel process was a workaround for the absence of that pipeline. Engineers were exporting daily averages, losing the granularity that would have shown peak-hour patterns, growth trends, and seasonality — the signals that matter for capacity planning.
Chapter 03 · Build
A four-component data platform on top of existing infrastructure.
We built on top of the network management infrastructure already in place, adding four components:
- Telemetry pipeline. Automated extraction from the network management system's time-series database into a data warehouse. Five-minute granularity preserved for the rolling 18-month window that capacity models require.
- Demand forecasting model. Time-series forecasting (Prophet, with segment-level tuning) producing 12-month utilisation forecasts by network segment. Models retrain weekly on new data. Confidence intervals shown alongside point estimates — engineers see both the forecast and the uncertainty.
- Capacity threshold alerts. Automated flagging when forecasts project a segment crossing 75% utilisation within 90 days. Alert triggers a planning workflow, not an email — the right engineer is notified with the segment, the timeline, and the available upgrade options pre-populated from the infrastructure inventory.
- Planning dashboard. Network-wide view of current utilisation, 90-day forecasts, and segments in the planning queue. Replaces the quarterly Excel exercise with a weekly review that takes 90 minutes.
Chapter 04 · Migration
Run in parallel first. Let the team trust it before the switch.
We ran the automated model in parallel with the manual process for two planning cycles before the team retired the manual approach. Comparing model output to the manually produced plan — which the network team knew well — built trust faster than any benchmarking exercise.
Where the model diverged from the manual plan, the team investigated. In most cases, the model was capturing a trend the manual process had missed.
Chapter 05 · Outcomes
The quarterly cycle became a weekly review.
The planning cycle compressed from 12 weeks to approximately one week. The finance team gained a rolling 12-month view of capital requirements rather than a point-in-time quarterly snapshot.
In the first year, the model identified £2.1M of planned capacity spend that could be deferred — segments the manual process had flagged for upgrade based on trend extrapolation, but where actual demand had plateaued.
The planning team now covers three times the network segments per quarter at one-third of the engineering time previously spent on data extraction and reconciliation.
Chapter 06 · Stack
Technology stack.
- Telemetry pipelineAutomated extraction from SNMP time-series DB, 5-minute granularity, 18-month rolling window
- ForecastingProphet (Meta) with segment-level tuning; weekly retraining; confidence intervals per segment
- Data warehouseCloud data warehouse for telemetry storage and model output; queryable by planning dashboard
- AlertingThreshold-based workflow triggers at 75% projected utilisation within 90 days
- Planning dashboardSelf-serve network-wide view; segment queue; 90-minute weekly review replacing quarterly cycle