Could a Neutral Broker Share Loads Between Rival Carriers?

The problem

The anchor paper covers machine learning in delivery route planning, at the phase "In early use": it helps large fleets plan their own routes and stop sequences, but the gains are small and isolated². This paper adds a different problem. It does not ask how one fleet plans better, but whether competing carriers could pool loads through a neutral party so that fewer lorries run empty.

The scale of empty running is documented. In 2025, goods vehicles in United Kingdom domestic road freight drove 5,897 million kilometres empty, 31% of the 18,975 million kilometres recorded, against 30% in 2024; the figures come from an official survey with a sample of 8,346 vehicles (official statistics, 2025 data, published 27 May 2026)¹. Eurostat reported that about a fifth of road freight kilometres in the European Union were driven by empty vehicles in 2020: 24% in national transport and 13% in international transport, ranging from 6% in Belgium to 44% in Cyprus (official statistics, 2020 data, published December 2021, so older than three years)³.

The statistics count distance driven empty, but they do not show how much of it a better match of loads could remove. Some of it may be unavoidable, for example where a vehicle type suits only one kind of cargo or where flows between regions are lopsided. The evidence covers the United Kingdom and the European Union only.

What already works

Load matching in practice. In a pilot in South West England, a freight-matching service arranged 65 return loads over 9,195 miles among 11 hauliers; the operator reports savings of almost 8 tonnes of CO2 and over £68,000 (operator-reported, July 2025, no control group stated)⁴. This shows that hauliers will use a matching service for return loads. It does not show what happens when rivals must share their regular customers.

What simulation shows. A peer-reviewed study of shared transport services tested what happens when carriers may refuse to hand over some customers, which the authors call assignment constraints (computed on generated test cases, 2020)⁵. Average savings were 24.5% for single-vehicle carriers and 16.8% for multi-vehicle carriers. For multi-vehicle carriers, savings were 11.1%, 16.7% and 22.6% depending on how much the carriers' customers overlapped, and keeping two-thirds of customers back from sharing left about 10%⁵. In plain terms, pooling saves distance in simulation, and the saving shrinks as firms hold back their best customers. The same study cites a Zurich case of 33 companies (distance down 32%, cost down 18%) and a survey in which over 70% of respondents were willing to share with a neutral trustee (both cited second-hand by the study, not checked against their originals)⁵.

A shared-fleet trial. A public-private trial in the United Kingdom, run by the University of Southampton with 25 suppliers, cut vehicle-kilometres from 878 to 622 (down 29%), emissions by 27-36% and warehouse visits by 59%, as reported by the authors (supplier survey November 2019, manuscript date not shown). It was halted in March 2020 by COVID-19⁶.

What machine learning adds to routing. The anchor's evidence suggests caution. In the Amazon-MIT routing challenge run with the University of Bonn, the winning entry did not use machine learning and 35% of the 2,285 participants used it (August 2021)⁷. In an A/B test at the delivery platform Meituan, on-time delivery rose by 0.27 percentage points and distance fell by 1.20% (operator-reported, published 2024)⁸. These are small gains from planning inside one operator. The pooling studies above suggest the larger prize may lie between operators, though they are simulations and one halted trial.

The possibility

One way this could work: a neutral broker, owned by an industry body or run by an independent operator, receives each carrier's open loads and spare capacity. It computes a joint plan, returns assignments to each carrier, and settles the money. Two designs are possible. In central planning, the broker minimises total distance for the whole pool. In an auction, carriers bid for loads and the broker clears the market. Central planning tends to find shorter overall plans but asks carriers to hand over more data and control; an auction leaves carriers more control but may leave savings uncaptured.

The role of AI. Part of this depends on AI, and part does not. The matching itself, deciding which load goes on which vehicle, is mostly classical optimisation (mathematical solvers), which the simulations above used and which is not machine learning. Machine learning could help in narrower places: forecasting where return loads tend to appear, predicting which matches might fail through late or cancelled loads, and estimating a fair value for each carrier's contribution. Everything else is ordinary software and institutions: data standards, contracts, insurance, liability, payment and settlement, and a legal structure that keeps competitors' commercial information apart. If these are the hard parts, then AI is a modest ingredient in a mainly organisational and legal problem.

What would have to be true

1. Some empty running must be avoidable. The statistics do not separate avoidable from unavoidable empty kilometres¹. If most of it comes from vehicle types or lopsided flows, pooling would recover little.
2. Carriers must put in more than the loads they are happy to lose. Simulated savings shrink when carriers hold customers back⁵. If each firm keeps its best accounts out of the pool, the pool may be left with the loads nobody wanted.
3. A fair rule for sharing the gains must exist and be trusted. The broker's neutrality is the product. Whether carriers would accept a neutral trustee is supported only by a second-hand survey figure⁵.
4. Information must be shared only to the extent the law allows. UK guidance from the Competition and Markets Authority (CMA) says that exchanging competitively sensitive information can be a restriction of competition by object, which means the firms would not need to be shown to have harmed anyone (paragraph 3.35)⁹. The European Commission has issued guidance on information exchange too, here read only through a law-firm summary¹⁰. If the broker could pass only what is needed, in aggregated or delayed form, the design might stay on the right side of the line; whether it could is a legal question the evidence here does not settle.
5. The efficiency case must be provable. The CMA guidance deals with who must prove the efficiency benefits, whether consumers receive a fair share of the savings, and the treatment of fixed-cost savings (paragraphs 3.49-3.54)⁹. If the savings stay with carriers, the exemption argument may be weak.
6. Someone must pay for the broker, and AI forecasts must be good enough to be worth it. Convoy, a US freight-matching company, closed its core business in October 2023 after failing to find a buyer¹¹. If the broker cannot cover its costs, or if load forecasts are no better than a carrier's own judgement, simpler tools would do the job.

What could stop it

What the evidence shows

• The one shared-fleet trial found was halted in March 2020 by COVID-19, so it never reached a steady state⁶.
• Simulated savings fall when carriers keep customers out of the pool⁵.
• The CMA treats exchange of competitively sensitive information between competitors as potentially restrictive by object, and puts the burden of proving efficiencies on the firms⁹.
• Convoy, a freight-matching company, closed its core business in October 2023¹¹.
• Gains from machine learning in routing inside a single operator were small in the two studies reused from the anchor⁷⁸.
• No trial of a neutral broker pooling loads between competing carriers was found. The searches for failures, critics and legal obstacles found the items above and nothing more specific.

Reasoning

The challenge question is who gains, who pays and whether something simpler would do. If large carriers with dense networks already fill most of their vehicles, they would gain least from pooling and could decline to take part, leaving a pool of smaller firms whose gains are real but whose bargaining power is weak. If savings flow mainly to shippers through lower rates, carriers may have little reason to carry the risk and the cost. A simpler route, such as a return-load exchange on which firms post only loads they already want to hand over, might capture a share of the benefit without any shared planning, and without raising the same legal questions. If so, the neutral broker would be a more ambitious design that answers a smaller question than it first appears to.

There is also a risk to people and prices. If a broker concentrates information about the whole market, it could become a point of dependence for small carriers, and a record of competitors' behaviour that regulators would watch closely. None of this can be tested without a trial in which independent measurement is built in from the start.

Signs to watch

• The next Department for Transport release of empty-running statistics (the 2025 edition appeared on 27 May 2026)¹.
• A published trial in which competing carriers pool loads through a neutral party and report measured empty running against a control group.
• An independent evaluation of any freight-matching pilot, such as the South West pilot, reporting beyond the operator's own figures⁴.
• Any statement, guidance update or decision from the CMA or the European Commission on information sharing in freight-sharing platforms⁹¹⁰.
• A follow-up to the Southampton shared-fleet trial that reports results beyond the period halted in March 2020⁶.