Capital-allocation agent for a national telecom operator's network-investment programme.

At a glance.

All client identifying details are anonymised under NDA. A senior lead reference call is available to qualified prospective clients on request.

The operator runs one of the largest mobile and fixed networks in the Kingdom. Its annual network-investment programme is governed jointly by the CFO and the CTO and runs to several billion riyals of capex across roughly nine thousand candidate projects per cycle. The planning process was anchored on a single spreadsheet maintained by a small team in network strategy; the spreadsheet had grown to a point where its assumptions were no longer fully traceable to source.

The CFO and the CTO had asked for the same thing in different language: the CFO wanted defensible capital allocation, the CTO wanted defensible coverage prioritisation, and neither could be confident the spreadsheet was producing either. A previous attempt by a large global vendor to replace the planning tool with a packaged solution had stalled because the operator's demand model — the most commercially sensitive artefact in the firm — could not be exported to the vendor's environment.

We deployed a sixteen-person pod for eight months. The pod sat between the CFO's office and network strategy, reporting jointly. We built a capital-allocation agent that ingests the operator's own demand model, the live coverage map, the project pipeline, and the binding regulatory commitments, and produces a ranked allocation with an auditable rationale per project. The system runs entirely inside the operator's sovereign-cloud tenancy; no commercial data left the perimeter. The CFO and the CTO co-chair a weekly governance call in which the system's recommendations are reviewed; overrides are first-class artefacts in the system and feed back into calibration.

USD one hundred and forty million of capex was reallocated in the first planning cycle, principally from coverage projects that the demand model showed as low-incremental to coverage projects in growth corridors that had been systematically under-prioritised under the spreadsheet process. The full planning cycle now runs in nine days, from a prior baseline of six weeks. Every decision in the cycle is linked to the demand model and the coverage map by construction. The pod is in the process of extending the system to the operator's enterprise-fibre programme.

Three. First, capital-allocation tooling that the CFO and the CTO can both sign is rare; the only viable design is one that makes the trade-off between commercial and engineering objectives explicit and reviewable. Second, the demand model is the crown jewel and the engagement must be structured so it never leaves the perimeter — that is what disqualified the prior vendor. Third, overrides are not failures of the system, they are training signal; the calibration loop is what makes the second cycle materially better than the first.