Can AI Speed Up Fulfilment?

Most of an order picker's day is spent walking, which makes routing and batching the obvious place to save time. The catch is that mature warehouses already run good routing rules, and in simulations the best of them sit close to the mathematical optimum. The gains reported for AI on top are measured in single figures or around 10%, and the largest come from companies describing their own systems.

Updated October 20268 min readEvidence-graded8 sources

On this page

The BFCS verdictEvery article scored on the same four dimensions

Better

PartlyLow confidence

Better plans on paper; whether people follow them is another matterA research challenge run by Amazon with MIT asked for routes that reflect the tacit knowledge of experienced drivers, and its organisers hoped to narrow the gap between route planning and real-life route execution. In our reading, a route that is shorter on paper is not always the better route on the street.

Note and confidence

Low confidenceBetter plans on paper; whether people follow them is another matterA research challenge run by Amazon with MIT asked for routes that reflect the tacit knowledge of experienced drivers, and its organisers hoped to narrow the gap between route planning and real-life route execution. In our reading, a route that is shorter on paper is not always the better route on the street.

Faster

Yes, at the marginLow to moderate confidence

Single figures to about 10% on travel and batch timeA literature review summarises experiments that put the best heuristic routing rule within about 5% of the optimal route on average, which caps what smarter routing can add. A peer-reviewed study using a Finnish retailer's picking data reported almost 10% less processing time from batching that models each picker's skill, measured on historical data. Amazon claims its fleet model will cut robot travel time by 10% (company claim, not independently measured).

Note and confidence

Low to moderate confidenceSingle figures to about 10% on travel and batch timeA literature review summarises experiments that put the best heuristic routing rule within about 5% of the optimal route on average, which caps what smarter routing can add. A peer-reviewed study using a Finnish retailer's picking data reported almost 10% less processing time from batching that models each picker's skill, measured on historical data. Amazon claims its fleet model will cut robot travel time by 10% (company claim, not independently measured).

Cheaper

PartlyLow confidence

Large savings claimed at scale; the build cost is large tooA literature review estimates order picking at as much as 55% of warehouse operating cost, without naming a source, so small time savings matter. Authors employed by UPS reported that their routing system cost more than $295 million and was expected to save $300 million to $400 million a year; that is optimisation, not machine learning, and the figures are the company's own.

Note and confidence

Low confidenceLarge savings claimed at scale; the build cost is large tooA literature review estimates order picking at as much as 55% of warehouse operating cost, without naming a source, so small time savings matter. Authors employed by UPS reported that their routing system cost more than $295 million and was expected to save $300 million to $400 million a year; that is optimisation, not machine learning, and the figures are the company's own.

Safer

ConditionalGovernance-gated

Safe when the pace an algorithm sets is a managed hazardFaster plans can become faster quotas. NSW has legislated a duty on algorithmic work allocation, though the duty has not yet commenced, and California regulators had received 168 reports alleging breaches of the state's warehouse quota law by 19 December 2022.

Note and confidence

Governance-gatedSafe when the pace an algorithm sets is a managed hazardFaster plans can become faster quotas. NSW has legislated a duty on algorithmic work allocation, though the duty has not yet commenced, and California regulators had received 168 reports alleging breaches of the state's warehouse quota law by 19 December 2022.

The problem, in numbers

Online orders are a growing share of Australian retail. In June 2025 the Australian Bureau of Statistics measured online sales at 12.7% of total retailing, in original terms, up from 11.6% a year earlier.^{Source 1} Every one of those orders has to be found on a shelf, packed and delivered, and the clock runs from the click.

Inside the warehouse, the expensive step is picking. A peer-reviewed literature review states, without naming a primary source, that order picking is estimated to cost as much as 55% of total warehouse operating expense.^{Source 2} The same review reproduces a typical breakdown of a picker's time from a 2003 facilities planning text, so the figure is second-hand: half goes on travel between locations, ahead of searching, picking and setting up.^{Source 2} In our reading, that is why routing, slotting and batching are where fulfilment speed is won or lost.

The evidence, in four numbers

up to 55%

of warehouse operating cost can go on order picking

estimate stated without a primary source in a peer-reviewed literature review (European Journal of Operational Research, 2007)

50%

of an order picker's time typically goes on travel

typical distribution from a 2003 facilities planning text, as reproduced in the same review; picker-to-parts warehouses

about 5%

gap between the best heuristic routing rule and the optimal route, on average

numerical experiments with random storage (Petersen, 1997), as summarised in the same review (second-hand)

almost 10%

less processing time from batching that models each picker's skill

peer-reviewed study on a Finnish retailer's historical picking data; abstract only, live trial not confirmed (2017)

What improves, and why

Fulfilment time has three levers that software can pull. Slotting decides where each product sits, so fast movers are close to the dispatch point. Batching decides which orders are picked together on one trip. Routing decides the path a picker, robot or van takes. All three cut travel, and travel is time that "costs labour hours but does not add value", as the review quotes from a warehousing text.^{Source 2}

None of this is new. Routing pickers through parallel aisles has a long research literature,^{Source 2} and in our reading the important finding for a mature warehouse is how good the old answers already are. The review summarises numerical experiments with random storage that found the best heuristic routing rule was on average about 5% longer than the optimal route, while the simpler S-shape rule, which traverses every aisle containing a pick, produced routes 7% to 33% longer than optimal in other simulations.^{Source 2} Both results are second-hand through the review. In our reading, a warehouse still using a simple rule has room to gain; one already on a good rule has little.

Where an order picker's time typically goes^{Source 2}

Travel

50%

20%

Pick

15%

Setup

10%

Other

Typical distribution in a manual picker-to-parts warehouse, from Tompkins et al. (2003) as reproduced in de Koster, Le-Duc and Roodbergen (2007). A textbook illustration, not a measurement of any one site.

What machine learning adds is prediction from data the old models ignored. A peer-reviewed study using a Finnish retailer's time-stamped picking data modelled how quickly each picker works and assigned batches accordingly. Its abstract reports almost 10% less total order processing time than state-of-the-art batching and routing methods. Against assignment based only on each picker's productivity, it reports 6% savings, and over 16% for some groups of pickers.^{Source 3} The gain was estimated on historical data; the abstract does not say it was tested live.

In a well-run warehouse, the shortest route is already close at hand. AI earns its keep by predicting the people and the orders, not by redrawing the map.

What the evidence actually says

At the largest scale, the figures come from companies. In a company announcement made on 30 June 2025, by trade press reports (the page itself is undated), Amazon claims a new generative AI model, DeepFleet, will improve the travel efficiency of its robot fleet by 10%.^{Source 6} That is a company claim, stated in the future tense, and we found no independent measurement of it.

Last-mile delivery shows the same pattern. Five authors, all employed by UPS, reported in a peer-reviewed journal that their ORION system, a meta-heuristic optimiser that plans a daily route for each of UPS's 55,000 US drivers, cost more than $295 million to build and deploy and was expected to save $300 million to $400 million a year.^{Source 5} The figures are the company's own, read from the abstract, and the system is classical optimisation rather than machine learning.

The more telling evidence is about the gap between plan and practice. Amazon, with scientific support from MIT, ran a research challenge in 2021 asking for models that produce routes reflecting "the tacit knowledge of seasoned drivers", using 9,184 historical Amazon routes from five US metropolitan areas.^{Source 4} The organisers wrote that learning-based routing was still relatively absent from the research literature, and that they expected the data set to help narrow the gap between theoretical route planning and real-life route execution.^{Source 4} In our reading, the challenge rests on the idea that the route an optimiser draws is not always the one experienced drivers run, and a speed gain that drivers override never arrives.

Where it falls short

The headroom is small where routing is already good. The best heuristic rule was on average about 5% from optimal in the experiments the review summarises.^{Source 2}
The 10% figures are not field trials. One was computed on a retailer's historical data;^{Source 3} the other is Amazon's claim for its own robots.^{Source 6}
The big savings are self-reported. The UPS figures were written by UPS staff and were expected savings, not audited ones.^{Source 5}
Plans meet people. Amazon's own challenge was framed around the gap between theoretical route planning and the routes drivers actually run.^{Source 4}

Where governance decides the outcome

A faster plan can become a faster quota. NSW has passed a law that, once in force, will require a business to ensure, so far as is reasonably practicable, that the health and safety of a worker "is not put at risk from the allocation of work by a digital work system", defined to include an algorithm or artificial intelligence. It received assent on 18 February 2026, but only three items about guidelines for WHS entry permit holders have commenced. The duty itself has not: the remaining amendments commence on proclamation after SafeWork NSW publishes guidelines, and as of 31 October 2026 no draft had been released.^{Source 7} California's warehouse quota law has applied since 1 January 2022: large warehouse employers must give each worker a written description of their quota, and a worker cannot be required to meet a quota that prevents meal or rest breaks, bathroom use or compliance with safety standards. By 19 December 2022 the state's Labor Commissioner had received 168 reports alleging breaches.^{Source 8} In our reading, any pick rate set by a model should be treated as a work health and safety hazard to be assessed, not only as a target.

How to think about adoption

Start by measuring travel. If pickers still walk every aisle in a fixed pattern, a better routing rule may deliver most of the gain without any machine learning.^{Source 2} Fix slotting next, then look at batching, where predicting each order and each picker is where data-driven methods have shown results.^{Source 3} Treat vendor percentages as hypotheses: pilot on one zone, time the whole order cycle from release to dispatch, and compare against the same weeks without the tool. For delivery, check how often drivers depart from the planned route and why, before blaming them. Ask how any new pace target will be assessed for safety, and record who signed it off. Demand is a separate question, covered in our forecasting review. In our reading, AI speeds up fulfilment at the margin in a mature warehouse; in a poorly run one, the first gains come from basic operations research.

How we read this evidence

BFCS.ai does not fabricate figures. The independent evidence on AI and fulfilment speed is thin. The peer-reviewed work we found is mainly operations research on routing and batching, two of the sources were read as abstracts only, and the batching gain was computed on one retailer's historical data. The largest speed and saving figures are company statements or written by company staff. We found no peer-reviewed study measuring order cycle time before and after an AI deployment in a working warehouse, and no Australian study.

Sources

1
Australian Bureau of Statistics, "Retail Trade, Australia, June 2025," released 31 July 2025. abs.gov.au
What we took from it
"In original terms, the proportion of online sales to total retailing rose from 11.9% to 12.7%" (May to June 2025), compared with 11.6% in June 2024; total online retailing sales were $4,703.8 million in June 2025, seasonally adjusted.
Back to text
2
de Koster, R., Le-Duc, T. & Roodbergen, K. J., "Design and control of warehouse order picking: A literature review," European Journal of Operational Research 182(2) (2007) 481 to 501, doi:10.1016/j.ejor.2006.07.009, invited review, available online 25 October 2006. pure.eur.nl
What we took from it
Read in full (publisher's PDF on the Erasmus University repository). Abstract: order picking is "the most labour-intensive and costly activity for almost every warehouse" and its cost "is estimated to be as much as 55% of the total warehouse operating expense," with no primary source given for the 55%. Fig. 4 gives the typical distribution of an order picker's time in a picker-to-parts warehouse, reproduced from Tompkins et al. (2003): travel 50%, search 20%, pick 15%, setup 10%, other 5% (the bars are unlabelled; values read from the chart's 0% to 60% scale); the text says travel "is often the dominant component" and quotes Bartholdi and Hackman: "travel time is waste." The review covers manual picking only. On routing (section 7), it summarises Petersen (1997), whose numerical experiments with six routing methods concluded that the best heuristic is on average 5% over the optimal solution in a random-storage setting, and simulation work by De Koster and colleagues finding S-shape routes on average 7% to 33% longer than optimal; both are second-hand through this review.
Back to text
3
Matusiak, M., de Koster, R. & Saarinen, J., "Utilizing individual picker skills to improve order batching in a warehouse," European Journal of Operational Research 263(3), 2017, 888 to 899, doi:10.1016/j.ejor.2017.05.002. ideas.repec.org
What we took from it
Read as abstract only. Using time-stamped picking data from a Finnish retailer, the authors forecast batch execution times per picker with multilevel modelling and assign batches with an adaptive large neighbourhood search; they report improving state-of-the-art batching and routing methods "by almost 10%," savings of 6% in total time against assignment based only on individual productivity, and over 16% for some picker categories. The abstract does not say whether results were tested in live operations.
Back to text
4
Merchán, D., Arora, J., Pachon, J., Konduri, K., Winkenbach, M., Parks, S. & Noszek, J., "2021 Amazon Last Mile Routing Research Challenge: Data Set," Transportation Science, published online 14 September 2022, doi:10.1287/trsc.2022.1173. cdn.amazon.science
What we took from it
Authors from Amazon and the MIT Center for Transportation and Logistics. The challenge asked participants to produce delivery routes "that reflect the tacit knowledge of seasoned drivers"; the data set covers 9,184 historical routes driven by Amazon drivers in 2018 in five US metropolitan areas (6,112 for training, 3,072 for evaluation); 45 final submissions were evaluated. The authors say learning-based approaches to vehicle routing are still relatively absent from the academic literature compared with optimisation methods, and that the data set should help narrow the gap between theoretical route planning and real-life route execution.
Back to text
5
Holland, C., Levis, J., Nuggehalli, R., Santilli, B. & Winters, J., "UPS Optimizes Delivery Routes," Interfaces 47(1), 2017, 8 to 23, doi:10.1287/inte.2016.0875. econpapers.repec.org
What we took from it
Read as abstract only; all five authors list UPS affiliations (Atlanta, Georgia, and Timonium, Maryland), so the figures are company-reported. ORION (On Road Integrated Optimization and Navigation) is described as a meta-heuristic optimisation system that provides an optimised daily route for each of UPS's 55,000 US drivers while keeping routes consistent from day to day; it cost more than $295 million to build and deploy and "is expected to save" UPS $300 million to $400 million a year, and to cut CO2 emissions by 100,000 tons a year.
Back to text
6
Amazon (company announcement, company claim, not independently verified), "Amazon deploys over 1 million robots and launches new AI foundation model," by Scott Dresser, VP, Amazon Robotics, aboutamazon.com. aboutamazon.com
What we took from it
The page shows no date (accessed 31 October 2026); trade press coverage dated 1 July 2025 (Automated Warehouse) reported the announcement as made the previous day, 30 June 2025. "A new generative AI foundation model will improve robot fleet travel efficiency by 10%"; the page also says DeepFleet will be "improving the travel time of our robotic fleet by 10%," and that Amazon has deployed its one millionth robot across a network of more than 300 facilities.
Back to text
7
SafeWork NSW, "Work Health and Safety Amendment (Digital Work Systems) Act 2026." The Act "received assent on 18 February 2026"; "only Schedule 1 [7], [9] and [11] commenced," relating to guidelines for the new entry permit holder power, and "The remaining amendments commence on proclamation after the Guidelines have been developed and published." New section 21A: safework.nsw.gov.au
and safework.nsw.gov.au
What we took from it
a person conducting a business or undertaking must ensure, so far as is reasonably practicable, "that the health and safety of a worker is not put at risk from the allocation of work by a digital work system." "Digital work system means an algorithm, artificial intelligence, automation or online platform." Status as shown on the page when accessed, 31 October 2026. SafeWork NSW's consultation page ("Development of the Digital Work Systems Guidelines", accessed 31 October 2026) says early targeted consultation concluded in September 2026, further targeted consultation follows in late 2026, and draft guidelines will then go to a 4 to 6 week public consultation; remaining provisions commence on "a date or dates to be announced by the NSW Government."
Back to text
8
California Labor Commissioner's Office, Department of Industrial Relations, "AB 701 Legislative Report," undated, with data current to December 2022. dir.ca.gov
What we took from it
AB 701 (Stats. 2021, ch. 197) took effect on 1 January 2022 and covers employers with 100 or more employees at one warehouse distribution centre or 1,000 or more across several in California; employers must give each employee a written description of each quota (Lab. Code s 2101); an employee "cannot be required to meet a quota that prevents compliance with meal or rest periods, use of bathroom facilities" or occupational health and safety standards. As of 19 December 2022 the office had received 168 reports alleging violations: 75 alleged the quota prevents meal periods, 93 health and safety compliance, 87 bathroom use, and 68 that quota descriptions were not provided (a report can allege more than one).
Back to text