Machine Learning in Delivery Route Planning: Large Fleets, Small Isolated Gains

The use case

Route planning decides which driver or courier takes which deliveries, and in what order the stops are visited. A van with 150 stops can be sequenced in more ways than could ever be checked one by one, so fleets use software that searches quickly for a good sequence rather than a perfect one.

What counts as AI here. In this paper, AI means machine learning: software that learns patterns from historical data instead of following rules written by people. In route planning it currently does three jobs. It predicts travel and stop times. It learns the habits of experienced couriers from the routes they actually drove, so that plans match how the work is really done. And in very large, fast-moving dispatch operations such as meal delivery, it steers the search for good assignments of orders to couriers.

What is older technology. The engine of most route planners is operations research: mathematical optimisation and heuristics (rules of thumb that find good-enough answers fast). The best-known case, UPS's ORION system, belongs here. Its sponsors present it as an operations research project, and it is credited with large savings¹. Those savings are not evidence for machine learning and appear in this paper only as a baseline.

Out of scope. Turn-by-turn navigation for private drivers, long-haul network design and driverless vehicles are not covered, except where navigation travel-time predictions are sold to fleets as an input.

Where it is running

Only two production deployments of machine learning inside delivery route planning are documented in enough detail to assess, both written up by the operators' own staff. A third, travel-time prediction, is sold to fleets as an input.

The non-AI baseline. UPS's ORION was expected to be used by 55,000 US drivers serving about 160 customers a day each at full deployment in 2016 (operator-reported, 2015)¹.

Amazon. Amazon co-sponsored a public research challenge in 2021 on learning from drivers' routes². In 2022, subcontracted drivers in Japan described the system that assigns their deliveries as AI (union-reported)¹¹. No published account of how Amazon's production routing uses machine learning was found.

Geography. The detailed evidence on machine learning comes from China and Germany. No comparably documented deployment was found in the United States, Latin America or Africa.

What the evidence shows

The isolated effect of machine learning is small

The clearest measurement comes from Meituan, which tested individual components of its dispatch system with A/B tests (randomised splits of traffic between the old and new versions).

• Risk-aware matching. A component that uses a learned statistical model to tune how cautiously orders are assigned was tested in nine cities for at least two weeks each. It raised the on-time rate by 0.27 percentage points and cut average travel distance by 1.20% (A/B test run by the operator, published 2024)³.
• Learned assignment guidance. A model trained to imitate high-quality offline solutions improved an internal assignment-quality score by 5–10% on sampled 2021 data (offline test by the operator, published 2024)³.

None of these tests reports a confidence interval, and none was repeated by anyone outside the company.

The headline system results are before-and-after comparisons

Meituan reports that between 2019 and 2022 average delivery time fell by 20.96% and courier travel distance per order by 23.77%, and that the system contributes about US$0.23 billion a year in cost reductions (operator-reported, measured without a control group)³. Those figures cover the whole system, which combines operations research and machine learning, over a period in which the number of cities served rose by 40% and peak daily orders doubled (operator-reported, 2019–2022)³. The contribution of machine learning cannot be separated from them.

Deutsche Post's published summary states that its routing approach was a main driver of accurate delivery time windows, but gives no accuracy figures (operator-reported, September 2023)⁴. The full paper could not be accessed.

Google reports that its travel-time predictions were already accurate for more than 97% of trips, and that a graph neural network (a model that treats the road network as connected nodes) cut the remaining inaccuracies by between 16% in London and Copenhagen and 51% in Taichung (vendor-reported, September 2020, older than three years; no newer figure was found)⁵. No published evidence links that gain to delivery-fleet outcomes.

Learning from drivers: optimisation still won the benchmark

In the 2021 Amazon Last Mile Routing Research Challenge, run with MIT, 2,285 participants tried to reproduce the sequences that experienced drivers actually chose. The winning team did not use machine learning: it extended classic search methods with constraints read from historical routes, extracted in about five minutes, and scored 42% better than the runner-up. Only 35% of participants used machine learning (organiser-scored competition, reported by the winning team's university, August 2021, older than three years)².

The non-AI baseline

ORION cost an estimated US$250 million, had saved more than US$320 million by December 2015, and was expected to save US$300–400 million and 10 million gallons of fuel a year (operator-reported, 2015, older than three years)¹. These are the figures most often attached to "AI route optimisation". They describe operations research.

A 2021 field study at a Jordanian soft-drinks distributor illustrates how weak much of the wider evidence is. Its routing comparison was a two-day simulation for eight vehicles: the software's plans covered 1,404 km against 2,050 km actually driven, but the reported profit difference was not statistically significant (p = 0.545) (simulation without a control group, peer-reviewed, 2021)⁸.

Plans and drivers: the adherence problem

A route plan is advice. Two independent studies show how often it is not followed.

• United States and Mexico. One year of data from a large soft-drinks company, covering 55,525 route instances, showed that 79.8% of US routes and 45.8% of Mexican routes departed from the planned stop sequence. Among those routes, the extra straight-line distance was small in the US (1.7%) but larger in Mexico (12.1%) (operational data without a control group; study date not shown)⁷.
• Shanghai. Across 54,390 courier routes drawn from 1.4 million deliveries over six months, couriers chose the computed best next stop 47% of the time: about 65% when two parcels were left and below 30% when 14 or more were left. Deviations added a median 54% to predicted travel time. Experienced couriers followed the computed route more often (observational study, independent, February 2025)⁶.

The Shanghai study has an important limit: the researchers built the "best" route themselves, using their own machine-learning travel-time model, and did not know what tools the couriers actually saw⁶. It measures the gap between people and a computed optimum, not compliance with a deployed product.

Planned savings are not realised savings. Where couriers routinely depart from the plan, the miles a system saves on paper overstate what it saves on the road.

What it takes: cost, integration, governance

Cost. Public cost figures are scarce. ORION's US$250 million (operator-reported, 2015)¹ is the only full programme cost found, and it is for a non-AI system. Meituan reports a dedicated dispatch research and development team of 30–40 people (operator-reported, 2024)³. No independent price data were found for commercial products with machine-learning components.

Data. Learning from couriers requires accurate historical records of the sequence actually driven, with clean locations and timestamps. The MIT study discarded whole depot-months in which no deviation was recorded, judging them mis-recorded⁷. Meituan's route planner is called more than 48.3 million times a minute at the lunch peak (operator-reported, 2024)³, which shows the computing load at that scale.

Integration. ORION's first algorithms worked in the laboratory but not in practice, and the system went through about three years of field testing before full deployment (operator-reported, 2015)¹. Deutsche Post designed its system around couriers' tacit knowledge, deriving it from their historical tours rather than replacing it⁴.

Rules in China. Following a 2020 magazine investigation into the risks riders took to meet algorithm-set deadlines, and regulator pressure for minimum pay and rest periods, Meituan published how its delivery-time algorithm works in September 2021. It replaced a single estimated arrival time with a time window, added allowances for delays outside riders' control, and committed to compulsory breaks (press report, September 2021)⁹. In August 2026, Meituan, JD Food Delivery and Taobao Flash announced timers that extend delivery deadlines by the time riders wait at red lights, after meetings with Beijing and Shanghai authorities (press report, August 2026)¹⁰.

Rules in the EU. Directive (EU) 2024/2831 on platform work was adopted on 23 October 2024 and entered into force on 1 December 2024¹². Member states must transpose it by 2 December 2026. It requires platforms to inform workers about automated monitoring and decision-making systems that affect task allocation and earnings, and to provide human review of significant decisions (law-firm summary, September 2026)¹³. Whether it reaches in-house parcel fleets, rather than app-based platforms, depends on national definitions.

Where it falls short

The headline numbers are mostly not about machine learning. ORION's savings are operations research¹. Meituan's 20–24% improvements are whole-system, before-and-after figures over a period of rapid growth³. The one isolated learning component with a published A/B test moved distance by about 1% and on-time rate by about a quarter of a percentage point³. On the public driver-behaviour benchmark, a non-learning method won by a wide margin².

Evidence is operator-written and unreplicated. Both detailed machine-learning deployments are described by the operators' own staff in an applied journal linked to industry prizes³⁴. No independent evaluation, randomised trial outside the operator, or replication was found.

Who absorbs the efficiency. Meituan's paper describes its cost reduction as savings in costs associated with payments to couriers, and reports that orders delivered per full-time courier per day rose by 109.63% between 2019 and 2022 (operator-reported)³. Riders' complaints about algorithm-set deadlines led to regulatory pressure in 2021⁹. In 2026 Meituan said it had previously estimated red-light waiting time from offline data and algorithmic prediction, a method it called less precise for e-bikes (operator statement, August 2026)¹⁰. In Japan, 15 subcontracted Amazon drivers in Nagasaki unionised in 2022, saying the routing system produced unfeasible routes and that one driver worked more than 12 hours a day (union- and worker-reported, September 2022)¹¹. Whether that system uses machine learning was not confirmed.

No withdrawals found. Dedicated searches found no published case of a machine-learning route planner being paused or withdrawn. They found labour disputes and regulatory intervention instead.

Answer to the challenge question. Route planning looks less successful once three things are separated: savings from optimisation versus savings from learning; planned miles versus driven miles; and gains to the operator versus changes in couriers' workload and pay. On the evidence available, machine learning adds modest, real improvements on top of mature optimisation at a few very large operators. Much of the value depends on data quality, courier cooperation and supervision rather than on the learning itself.

Phase: In early use — why

The editorial hypothesis was "In wide use". Route optimisation software is in wide use, but most of it is operations research. Machine learning inside production route planning is documented at two very large operators in China and Germany³⁴, with machine-learning travel-time predictions available as an input⁵. No independent survey of adoption was found. The evidence therefore supports "In early use" for the machine-learning part, which is the subject of this paper.

What to watch next

• 2 December 2026: EU deadline for transposing the Platform Work Directive. Watch how national laws treat automated allocation and timing of delivery work¹³.
• Netherlands Platform Work Act: consultation closed on 24 August 2026. Watch whether the bill is tabled before the deadline¹³.
• End of 2026: Meituan's stated aim to extend red-light timers to more than 1 million riders, with more than 20 further cities under review¹⁰.
• Trigger: the first independent randomised or A/B evaluation of learning-based route planning published outside an operator.

Questions to ask before adopting

• What share of the promised saving comes from machine learning rather than from the optimisation engine and better address data?
• Was the saving measured on routes actually driven, against a control group, or on planned routes before and after?
• What share of our drivers follow the planned sequence today, and how will the system learn from those who do not?
• Who absorbs the gain: fuel and vehicle costs, customers, or drivers' workload and pay per stop?
• Which rules on automated management of workers apply to how this system allocates and times work?