A practitioner map, drawn from running a wholesale network: the four places machine intelligence earns its keep today, and the places it still does not. No demos, no slideware.
Machine intelligence earns its keep in wholesale operations wherever there is high-volume data, a fast feedback loop and a measurable cost of being wrong: fraud and bypass detection on live CDR streams, routing and quality prediction per corridor, traffic supervision, and settlement assurance. Anywhere one of the three is missing, AI is decoration.
The flagship. OTT bypass, CLI refiling and A2P grey routes all leak money through statistical patterns no human can hold in their head: per-corridor baselines, A-number clustering, short-call ratios, B-number repetition. Models propose, engineers confirm, thresholds recalibrate quarterly. On flagged traffic, expect on the order of a third of machine flags to be false positives without expert review. That number, measured, is what separates detection from decoration. OTT bypass observatory, CLI refiling observatory.
Historical ASR and ACD per route tell you what happened. Prediction tells you what is about to happen: a route drifting out of baseline, a destination degrading under new traffic mix, a supplier change that smells like a refile. Quality-guarded least-cost routing consumes those predictions continuously, so traffic steers before invoices feel it. How we route.
Anomalies are revenue leaks announcing themselves: a sudden throughput spike on one corridor, a route eating capacity without answering, a short-call pattern creeping up at 3am. Supervision systems watch the stream and escalate to an engineer inside minutes, around the clock. The alternative is finding out at month-end settlement. Traffic management.
Reconciliation is pattern matching at scale: interconnect invoices against CDRs, entitlements against reported subscriptions, revenue share against provable activity. This is where billing verification and settlement live, and where machine matching catches what sampling never will.
Interconnect negotiations are relationships, and relationships are human. Novel fraud is investigated by people who enjoy puzzles; models rank the suspects, humans close the case. Explaining a routing decision to an angry partner at 2am is an engineering job, not a model output. And any process without a feedback loop, which includes most strategy, is somewhere models confidently produce nonsense. We built our operations around that boundary: machines where the loop exists, engineers where it does not.
