Identifying £2.2M in Revenue Leakage for Global Airport Services

QuantSpark's anomaly detection model identified an estimated £2 to £2.2 million in potential annual revenue that a global airport services operator was missing across its UK and Ireland operations, by finding services rendered to airlines that were never captured or billed. Roughly 5% of an estimated 68,000 annual UK&I flights were found to have at least one missing service. The figure is a modelled, potential value: it identifies where revenue was likely being left on the table, not cash already recovered, and its confirmation depends on the retro-billing process the model was built to support.

The operator had already made real progress before QuantSpark was engaged. It had replaced a paper-based process for recording ground services with a mobile application, and that single change had already surfaced 28,000 additional services captured globally that would previously have gone unrecorded. The open question was whether that improvement had closed the gap completely, or whether a subtler layer of missing services, ones that did not fit an obvious pattern a human auditor or a simple rule could catch, was still slipping through across thousands of flights a year. At the scale of tens of thousands of flights, each with multiple possible service types, aircraft variants and contractual exceptions, manually auditing for these gaps was not practical, and a simple rules-based check risked either missing complex cases or generating so many false positives that no one would act on them.

QuantSpark's response was a layered anomaly detection model rather than a single rule set. The work began with a series of workshops with the operator's stakeholders, both to gather the operational knowledge needed to build the model and to create the awareness and advocacy that would make its findings actionable rather than shelved. From there, the scope was deliberately narrowed: rather than analysing the whole UK 2025 flight base at once, the team filtered down to three customers and three stations, creating a focused dataset the model could be built and validated against before any wider rollout. On top of that focused dataset sat two complementary layers. A supervised machine learning layer looked for patterns of likely missing services across eight distinct service types, with detection thresholds tuned specifically to maximise the capture of genuine missed services. A rules layer sat alongside it, encoding agreed business rules and contractual exceptions, validating service type, aircraft type, timing constraints and operational feasibility, to strip out the false positives that a purely statistical model would otherwise flag. The two layers combined into a single output: each potential miss was assigned a priority and a confidence band, and the results were surfaced through dashboards designed to feed directly into the operator's retro-billing process rather than sit as a one-off analysis.

That workflow, from scoping through detection, validation and prioritised output, is what turned a broad "are we missing revenue" question into a short list of confidence-ranked, contractually-validated cases the operator's own teams could act on. The systems involved were, categorically: the operator's own mobile data capture application as the underlying source of service records; a supervised ML layer as the detection engine; a business rules layer encoding contractual logic as the false-positive filter; and a dashboard layer for prioritised reporting.

The value case rests on the same logic. Across the 68,000 annual flights analysed in UK&I, around 5% were found to carry at least one missing service; aggregating the estimated value of those flagged, confidence-banded instances across a full year of flight volume produced the £2 to £2.2 million potential annual revenue estimate. It is presented deliberately as a range and as "potential" because it is the output of a detection and prioritisation exercise, not a confirmed recovery: turning it into billed revenue depends on the operator running its retro-billing process against the prioritised, confidence-ranked list the model produces.

That said, the honesty of the figure is also its usefulness. By separating high-confidence, rule-validated cases from lower-confidence ones, and by keeping the modelled £2 to £2.2 million distinct from the already-realised 28,000 services captured through the earlier mobile app rollout, the operator has a clear, decision-ready view of where to focus retro-billing effort first, rather than an undifferentiated list of anomalies.