Logistics Operations Automation: What Production Deployments Show

Yes — on both floors, with the paperwork floor delivering the most dramatic compressions. Gelato's multi-agent integration work cut carrier onboarding from 5 days to 10 minutes — roughly 99% of the effort removed — and collapsed SKU mapping projects that previously ran 9 to 24 months per printer partner, with throughput growing without proportional headcount. Sappi's automation programme runs 132 deployed automations saving about 13,000 hours a year in Europe alone, including a document robot processing around 12,000 documents annually at an 80% no-touch rate and finance reports that took an hour now generating in under three minutes.

The physical floor's numbers come from warehouse management: Tennis lifted overall warehouse efficiency 50% — receiving up 20%, picking 25%, packing 20% — with labour costs down 19% and store replenishment cycles cut from 4.6 days to two, while inventory visibility reached nearly 100%.

And at marketplace scale, the record shows what the ceiling looks like: DoorDash's ML platform serves three million predictions per second balancing courier supply against delivery demand. The common thread from a paper mill to a delivery marketplace: the automation reads what humans used to retype, checks it against what was agreed, and hands coordinators exceptions instead of everything.

Two decays, one loud and one silent. The loud one is the format long tail: every carrier, forwarder and customs regime speaks its own document dialect, and the before-states are full of the cost — container inspections delayed minutes apiece over paperwork, booking confirmations consuming close to 2,000 hours a year at one company before automation, bulk CSV tools and button-press robocall systems standing in for actual integration. The paperwork floor's wins are exactly this tail being tamed, format by format, with extraction that normalises rather than templates that memorise.

The silent decay is the one this category teaches the whole site. DoorDash's forecasting write-up discloses it plainly: overly complex ML pipelines with long data-dependency chains initially outperformed naive forecasting — then degraded to what they label 'Terrible' performance after three or more months, with heavy on-call burden. Worse, blind correlation learning risked conclusions like high incentives leading to fewer couriers on the road — spurious relationships confidently learned. Their fix was institutional humility: simpler models where possible, causal structure where it matters, and monitoring that assumes decay rather than hoping against it. For anyone buying rather than building, the lesson converts directly into a procurement question: ask the vendor how the model's accuracy is monitored after go-live, and what happened to it last quarter — because a forecast that was good at the demo is a depreciating asset, and this record has the depreciation curve in print.

Split by floor, and the record splits with you. The physical floor is bought: warehouse and transportation management is decades-deep vendor territory — Blue Yonder anchors this record's WMS wins — and the productivity numbers above ride on configured platforms, not custom code. The paperwork floor is increasingly bought too, but in a newer shape: document automation and agent-assisted integration, where Gelato's 5-days-to-10-minutes carrier onboarding shows what the agentic generation does to work that used to be integration projects. The RPA lane (Sappi's 132 automations on UiPath) covers the repetitive middle at programme scale.

The build lane exists at exactly one altitude: marketplaces and networks whose logistics is the product. DoorDash's platform — the three-million-predictions-per-second serving system, the forecasting stack with its published failure modes — is engineering-organisation work with engineering-organisation economics, valuable to everyone else mainly as a source of lessons: monitor for decay, prefer simple where simple survives, distrust correlations without causal structure.

Decision inputs for the majority: which floor bleeds most (count coordinator hours on documents versus warehouse labour variance), how long your carrier tail is — the longer it is, the more the modern extraction-and-agent tooling earns — and, for anything with a forecast inside, the after-go-live monitoring question above. Buy the floors; import the humility.