Telecom optimization does not become an AI problem when the models get good enough. It becomes one when the decision space outgrows the review cycle — when routing, pricing, and capacity decisions change faster than analysts and static rules can keep up, and the gap between the decision the team would make with current information and the one they actually make with last week's information starts showing up in monthly results.
That threshold arrives quietly, because nothing breaks. The current process still produces defensible answers while value leaks out of the corridors nobody had time to re-price. Industry roaming standards (GSMA, 2024) and telecom profitability research (McKinsey, 2024) describe the squeeze from both ends: margins under pressure while complexity rises.
ML LABS has been on the other side of that threshold. Engaged through Gigster, the practice designed and built the ML routing platform for a top 10 global telecom company that was losing margin on international roaming — time-series models running on roughly 1TB of new traffic data a day, optimizing cost and routing across the operator's carrier agreements. The manual process it replaced was not incompetent; it was outrun. By the time analysts had modeled one corridor in a spreadsheet, the demand pattern they had modeled against was already a week old — and the portfolio of bilateral carrier agreements underneath was churning on its own cadence, a structural complexity documented across the industry (OECD, 2013).
graph TD
A["Decision surface<br/>too large"] --> D["Manual optimization<br/>systematically too slow"]
B["Environment changes<br/>faster than review cycle"] --> D
C["Waste visible<br/>at leadership level"] --> D
D --> E["AI optimization<br/>justified"]
style A fill:#1a1a2e,stroke:#ffd700,color:#fff
style B fill:#1a1a2e,stroke:#ffd700,color:#fff
style C fill:#1a1a2e,stroke:#ffd700,color:#fff
style D fill:#1a1a2e,stroke:#e94560,color:#fff
style E fill:#1a1a2e,stroke:#16c79a,color:#fffThree Signals The Threshold Is Crossed
The decision surface is too large to search. Routing options, quality thresholds, pricing tiers, demand shifts, and capacity constraints interact in ways that look manageable one corridor at a time and become intractable across a portfolio. The process falls back to heuristics and periodic review — which is not optimization, it is triage with a spreadsheet.
The environment moves faster than the loop. A decision model built on weekly analysis loses its edge the moment demand, congestion, or pricing shifts inside the week. Research on AI-driven telecom network planning (Xue et al., 2025) reinforces the point: once the environment is dynamic and multidimensional enough, the return comes from systems that update recommendations against live conditions, not from better static rulebooks.
The waste is visible at leadership level. Routing cost, capacity utilization, or planning inefficiency has to be material enough to justify changing the operating model — and the case is strongest when the same optimization pattern repeats across markets, so one build serves many decisions rather than one.
What The Manual Process Was Missing
The signals that mattered on the roaming build were visible in the operator's data long before the platform shipped, and they are the ones worth instrumenting anywhere. Margin per session was drifting downward on corridors that had been profitable a quarter earlier, and the drift was masked at the aggregate level because the high-volume corridors averaged it out. Quality complaints were rising on the paths where the cheapest carrier had been chosen and then degraded, because nobody had time to re-price the alternative. And agreement renewals were being signed against pricing baselines built from stale traffic — commercial terms optimized for the network the operator had six months earlier rather than the one it had at signing.
The cleared outcomes of that build are worth stating plainly, because they show where the value sat. The platform surfaced optimization opportunities across 128% more corridors than the original scope targeted — inefficiencies the manual analysis had never physically reached. Measured against the cost of the engagement, the optimization returned over 12x in the first year. The full account is the global roaming optimization build.
"Omar is one of the most skilled engineers I have worked with in 20 years of delivering technology. He led a program to build a custom ML platform for a top-10 global client. The work continuously exceeded expectations and became their foundation." — VP of Customer Success, Gigster
How To Pick The First Loop
The first way to get this wrong is to try to optimize everything at once, which creates enough scope to guarantee that no single decision loop is ever proven. The stronger path is narrower: pick the corridor or cost center where the current process is already visibly weak, work against live decision variables, and let one loop earn the right to expand.
The second way to get it wrong is to treat the optimization layer as a model-selection exercise. The model is one component. The harder problem is closing the loop end to end — ingesting fresh operational data, generating recommendations on a cadence that matters, presenting them to network operators in a form they will actually act on, and capturing the outcome of each decision so the system learns from production behavior rather than from a backtest. A survey on AI-driven network management (Intl. J. Network Management, 2025) makes the same argument for autonomous-network operating models over model-as-deliverable thinking. Skip the closed-loop design and the result is a research project with a dashboard.
A schematic makes the shape concrete — illustrative, not a client record. Picture an operator routing international voice traffic across a handful of carrier partners, choosing a route per destination and balancing per-minute price against answer-seizure ratio and post-dial delay. At low complexity, a weekly review and a static routing table is enough. The threshold is crossed when partner quality varies within the day, when volume-tier breakpoints move faster than the ranking that was built from them, and when the penalty for a bad route shows up in the same dashboards leadership already reads. At that point the optimization cadence, not the model, is the bottleneck.
A good first loop has three properties, and a loop missing any one of them is the wrong place to start.
- The decision recurs on a short cycle. A decision made quarterly cannot be optimized continuously, no matter how expensive each instance is.
- The gap between a good and a bad decision is material. Where the options differ by little, the gains disappear under operational overhead.
- The current process is observably behind. If the manual process is already strong, the accuracy bar for replacing it is high enough to make the project a bad trade.
Choosing against those three criteria matters more than choosing the technically interesting problem — and it is the choice that decides whether the second loop ever gets funded.
When A Rules Engine Is Enough
If the decision rules are stable and the manual process keeps up, a rules engine is the better answer, and an AI optimizer is an expensive way to buy the same outcome with more operational surface. The honest test is whether the environment moves inside the review cycle. If it does not, nothing about the decision requires a model.
The other blocker is upstream. If the operational data needed for routing or pricing evaluation cannot be reached in time to act on it — trapped in systems that export nightly, or reconciled by hand across carrier partners — then the data path is the first project, and no optimization layer built on top of it will outrun its inputs.
First Steps
- Pick one loop. Choose a corridor, routing class, or cost center where the current process is already leaking visible value, and write down what a good decision would have been last week.
- Measure the lag. Compare how fast the environment moves against how fast the team can respond. The difference between those two numbers is the size of the opportunity.
- Trace the data path. If live inputs are weak, fix that first. If the data is already reachable, the next move is optimization design against a real decision surface.
Build One Loop, Then Widen
The pattern that works is one narrow optimization loop, built against live operational data, with explicit cost and quality trade-offs and one production decision surface — proven before coverage expands. It is a systems-integration problem before it is a modeling problem: the loop has to reach the carrier systems, the pricing data, and the operations team's own tooling, and every one of those boundaries is where the project actually lives.
That is the shape of an AI system integration engagement: one AI system built across the systems that hold the decision, accepted against targets written into the contract before work starts, and carrying its first 60 days of Operate — run by the same engineering that built it, on the loop that has to keep earning after launch. Where the corridor is chosen and the data path is the open question, how to know your data is build ready is the test to run first, and the engineering path from pilot to production is what the loop has to survive after it works.
References
- GSMA. Roaming. GSMA, 2024.
- McKinsey & Company. How Generative AI Could Revitalize Profitability for Telcos. McKinsey & Company, 2024.
- Xue, Y., et al. TelePlanNet: An AI-Driven Framework for Efficient Telecom Network Planning. arXiv, 2025.
- Hong, J., Tu, N., & Hong, J.-K. A Comprehensive Survey on LLM-Based Network Management and Operations. International Journal of Network Management, 2025.
- OECD. International Mobile Roaming Agreements. OECD Digital Economy Papers, 2013.
