Logistics: finding the real causes of late deliveries in hours, not weeks

This is an illustrative scenario modelled on typical mid-market deployments — not a specific customer.

The situation

Picture a Nordic third-party logistics operator with around 400 staff — the kind of business that moves other companies' freight across depots, lanes and carriers, and lives or dies on thin margins and its reputation for reliability.

For most of the year, the operations team ran on instinct and firefighting. On-time delivery (OTD) — the single number their customers judged them on — had been quietly slipping, and nobody could say exactly why. Was it a particular carrier letting them down? A specific lane that jammed up on certain weekdays? A depot that fell behind whenever volumes spiked? Everyone had a theory. Nobody had an answer they could defend.

The information that would settle the argument existed, but it was scattered. Shipment and route data sat in the transport management system (TMS). Pick, pack and dispatch times lived in the warehouse management system (WMS). Each carrier sent its own performance feed in its own format. Sustainability and carbon reporting for customers was stitched together by hand in spreadsheets at the end of each quarter, a job everyone dreaded.

So when a big customer asked a pointed question — "why were 12% of our shipments late last month, and what are you doing about it?" — the honest answer took the operations lead the best part of two weeks to assemble. By then the month was over, the causes had shifted, and the analysis was already stale.

Why the existing tools fell short

The operator had the usual dashboards. They were good at showing what had happened: OTD by week, volumes by depot, a red number when a target was missed. What they could not do was explain why.

A dashboard tells you OTD dropped to 88%. It does not tell you that most of the miss came from one carrier on two lanes out of a particular depot, and only on the days when inbound volume ran hot. To get from what to why, an analyst had to pull extracts from the TMS, the WMS and three carrier feeds, reconcile them by hand — reference numbers never quite matched — and run correlations in a spreadsheet. Weeks of work for one question, and the moment the next question arrived, it started again.

Cloud AI analytics tools promised to close that gap, but they brought their own problems. They charged per query or per token, so the more the team explored, the more it cost — exactly the wrong incentive when you want people asking lots of questions. And for a European operator handling customer shipment and, in places, personal delivery data, sending it all to a US-hosted AI service was a governance headache the compliance team would not sign off on.

What changed with Adaptrix

The turning point was simple to describe: the operations team could now ask a question in plain language and get a grounded answer back, with the analysis to support it — in minutes, not weeks.

Adaptrix connected directly to the operator's PostgreSQL-based TMS and WMS stores and ingested the carrier feeds, then auto-discovered a semantic layer across them — a shared map of what a "shipment", a "lane", a "carrier" and a "depot" actually meant, with low-confidence matches routed to a human for a quick review rather than guessed. There was no six-month modelling project to build a data warehouse first; the shared vocabulary came from the data the team already had.

With that foundation in place, the questions the team had been arguing about for months became things they could simply ask:

• What is actually driving our late deliveries? Instead of one more correlation in a spreadsheet, Adaptrix ran key-driver root-cause analysis — automated exploratory analysis, group-by segmentation, correlation, and SHAP key-driver analysis — to rank how much lane, carrier, depot, weekday and a weather proxy each contributed to OTD misses, and synthesised it into a plain-language root-cause summary. The answer was specific and defensible: a short list of the factors that mattered most, not a wall of charts.

• Where is something quietly going wrong? Anomaly detection on transit-time and route data flagged shipments and lanes behaving unusually — a route whose transit times had crept up, a depot that was drifting — so the team could look at problems while they were still small, rather than discovering them in next month's report.

• What volumes are coming, and can we handle them? Forecasting of shipment volumes gave the planners a forward view for capacity planning, so they could line up carrier capacity and depot shifts ahead of demand instead of scrambling when a spike arrived.

• Which lanes and carriers are pulling their weight? Segmentation grouped lanes and carriers by performance, turning a vague sense of "some carriers are better than others" into a clear picture they could act on in negotiations and routing decisions.

Underpinning all of it, data-quality profiling ran across the TMS, WMS and carrier feeds — checking completeness, validity, consistency and freshness — so the team knew which numbers to trust and where a carrier's feed had gaps, before those gaps quietly distorted a decision.

Crucially, none of this data left EU control. Adaptrix ran on Adaptrix's sovereign EU infrastructure, with the AI models self-hosted and never shared with a US cloud or a third-party AI provider, which meant EU data residency and GDPR-native handling by default — and, because the models are self-hosted, no per-query or per-token charges. Exploration was free to do, which is exactly what you want when a team is finally allowed to be curious.

The outcome — thrive, efficiency, profit

The most immediate change was human. The operations lead stopped spending the first fortnight of every month building a retrospective and started spending that time fixing the things the analysis surfaced. The team moved from firefighting toward improvement — from explaining last month's misses to preventing next month's.

The defensible, directional outcomes follow from that shift. Root-cause analysis that used to take weeks now returns grounded answers in hours, so decisions get made while they still matter. Anomaly detection turns some late deliveries from surprises into early warnings the team can act on. Volume forecasts let planners commit carrier and depot capacity proactively rather than paying the premium for last-minute cover. And because the same unified layer produces both OTD and carbon figures, the operator can hand customers a consistent, defensible KPI view instead of a hand-built spreadsheet — turning a quarterly chore into a selling point.

These are modelled improvements for a deployment of this shape, not a verified result from a named customer — but they follow directly from replacing weeks of manual reconciliation with grounded answers on demand. On a thin-margin business, spending less time explaining problems and more time removing them is where the profit is.

Why this pattern generalises

Any operator whose performance data is spread across a transport system, a warehouse system and a handful of external feeds faces the same gap between what happened and why — and the same answer applies: unify the data into a shared semantic layer, then let the people who run operations ask questions and get grounded, defensible answers, on Adaptrix's sovereign EU infrastructure, without paying by the query.