Could Rural Buses Be Redesigned from the Booking Data of On-Demand Minibuses?

Executive Summary

Rural England has lost bus services for a decade, and many councils have turned to on-demand minibuses, booked by app or phone, as a partial answer. These schemes produce something timetabled buses do not: a dated record of where each resident asked to go. This paper asks whether that record could be used to redesign the fixed bus network.

The evidence on running such schemes is reasonably clear. The Department for Transport's evaluation of the Rural Mobility Fund found a mean of 2.1 journeys per vehicle hour across 18 schemes, with a wide range, and the one rural-density example it reports (Staffordshire) managed 0.9. A council-reported survey finds that most authorities running on-demand services say they lose money on them. The evidence on using the data to redesign networks is thin: one Newport example was found, with no quantified result.

The possibility is a concept only. The building blocks exist separately, but no study was found showing them working together. The role of AI is small. Most of the gain would come from ordinary analysis of booking logs, a map and a decision about funding. The harder conditions are pooling data across operators, funding that lasts long enough to test a redesigned route, and a legal route for treating on-demand services as part of the network.

At a glance

The problem

The anchor paper in this series looked at machine learning in delivery route planning and concluded that the gains cluster in large fleets and stay small and isolated elsewhere¹. It drew on large-scale dispatching reported by an operator⁶ and on a machine learning competition result⁷. This paper adds a different setting: a public bus service in thinly populated areas, where the question is access at an acceptable subsidy, not the cost of a parcel round.

The starting problem is the decline of the rural bus. Bus journeys outside London fell from 4.6 billion in 2009 to 3.6 billion in 2024, a fall of 21.7%, according to a summary of the Transport Committee's 2025 report². The same summary reports a campaign group's finding that 56% of small towns are in a "transport desert"². A survey of county councils published in October 2023 found that more than one in four bus services had been lost over a decade, and that there were 344 million fewer bus journeys in 2022³.

Many councils responded with demand-responsive transport (DRT), meaning minibuses that run when a passenger books rather than to a fixed timetable. In the same survey, 75% of councils said they offered DRT³. This is council-reported and the survey is small: 24 responses³. It does show where local authorities have put their effort.

The central point for this paper is that every DRT booking records an origin, a destination, a requested time and whether the request could be met. A timetable is designed in advance from census figures, counts and local knowledge. A booking log is the opposite: a record of the journeys residents actually tried to make.

What already works

The Rural Mobility Fund (RMF) gave £19.4 million to 15 local authorities in March 2021 to trial DRT, and the Department for Transport commissioned the University of the West of England (UWE Bristol) to evaluate it⁴⁵. The final Phase 1 report, issued in September 2025 and published in December 2025, covers 18 schemes⁵.

What the schemes carry. For April to September 2024, the mean was 2,101 journeys a month per scheme, with a range of 117 to 6,458. The mean was 2.1 journeys per vehicle hour, with a range of 0.9 to 3.9. The empty running ratio was 0.77⁵. Over the whole life of the schemes the averages were 1.9 journeys per vehicle hour and 0.91 for the empty running ratio⁵. Revenue was £2.01 per journey, which the evaluation describes as an under-estimate⁵.

In plain terms, a typical scheme carries a little over two passengers for each hour a minibus is on the road, and some carry fewer than one. Without subsidy this is a service that does not pay for itself.

Booking is already digital. Of all journeys, 87% were booked by app and 11% by phone⁵. This matters for the data argument: nearly every journey leaves a digital record.

Effect on the wider bus network. Journeys rose in 6 of 7 pilot areas studied. Warwickshire fell 11.9% against a rise of 20.4% in its comparison area. The concession share of journeys was 26%⁵. The evaluation therefore does not show that DRT adds to bus use everywhere.

Density. Staffordshire, at 148 people per square mile, managed 0.9 journeys per vehicle hour⁵. This is the lowest end of the reported range and suggests that thin demand is the main drag on efficiency, whatever software is used.

“Staffordshire, at 148 people per square mile, managed 0.9 journeys per vehicle hour.”

Using booking data to change the network. A trade-press article from March 2024 describes how some operators have used DRT trials to learn where demand sits, including one example in Newport⁸. No figures on the result were given, and no other redesign from booking data was found in the sources read.

A failed pilot. North Yorkshire Council's YorBus pilot ended on 30 June 2023 after carrying 14,208 passengers¹⁰. A local news report put the cost at £15.57 per passenger, against about £6.50 on timetabled routes¹¹. The council describes the pilot as a source of learning for future provision¹⁰ (operator-reported); the cost figure is from the news report¹¹.

The law. Section 14 of the Bus Services Act 2025 came into force on 1 April 2026⁹. The research did not confirm how it applies to on-demand services, and the rest of the Act was not read.

The possibility

This section is reasoning, not evidence. It sets out one way booking data could be used.

One way this could work is as a cycle. Over a period of months, a council or operator would collect the booking logs from its on-demand service, including requests that could not be met. Analysts would map the origins and destinations, look for corridors where requests repeat at similar times, and compare them with the existing timetable. Where a corridor is busy enough, a fixed or semi-fixed route might replace several separate on-demand trips. Where it is not, the on-demand service might stay.

The cost case would depend on which side of a threshold a corridor falls. If a fixed route carries more passengers per vehicle hour than the 2.1 average reported by the RMF evaluation⁵, it could be cheaper per passenger than the on-demand service it replaced. That is a hypothesis; none of the sources read tested it.

The role of AI. Three parts of this cycle could use machine learning: forecasting demand on a corridor, grouping similar requests into candidate routes, and matching passengers to vehicles in real time. The first two are decision support on historical data. The third is the dispatch software that already runs the on-demand service. All the rest is ordinary software, data handling, funding and policy.

The part of this possibility that depends on AI is small. The larger parts are data pooling, long-enough funding and a legal route for on-demand services to sit inside the network.

If the main gain is a redesigned network, much of it could be done with a spreadsheet, a map and several months of booking logs. Machine learning might add value where several schemes pool their data, where demand shifts by season or event, or where the number of corridors is too large to examine by hand. For one small scheme, the evidence read here gives no reason to expect a gain from it.

This is consistent with the anchor paper's view that machine learning gains concentrate in large fleets and stay small in isolated settings¹. A rural council running one scheme with a few hundred journeys a month may resemble the isolated case. A group of councils pooling logs might sit nearer the large-fleet case.

What would have to be true

1. Booking logs are complete and held in a usable form. The logs would need to include unmet requests, not only completed journeys, and be exportable by the council rather than locked in a supplier's system. The share booked digitally (87%) suggests the raw material exists⁵; whether councils can access it was not established.
2. Data can be pooled across schemes. A single rural scheme can carry as few as 117 journeys a month⁵. Pooling would require shared formats and agreement on who may use the data. If it cannot be pooled, a machine learning step is unlikely to have enough data to add value.
3. Demand on some corridors is dense enough for a fixed route. The Staffordshire result of 0.9 journeys per vehicle hour⁵ suggests that in the thinnest areas a fixed route could fail for the same reason. The approach would help only where requests cluster.
4. Funding lasts long enough to test a new route. In the council survey, 88% wanted long-term funding and 95% said their DRT service ran at a loss³ (council-reported). A redesigned route would need time to build ridership, and a short funding window could end the test before it is complete.
5. There is a route in law for on-demand services to count as part of the network. The research did not confirm how the Bus Services Act 2025 treats them⁹. If on-demand services cannot be listed as socially necessary, a redesign could be hard to protect.
6. The AI step beats the simple alternative. Any machine learning forecast or clustering would need to be compared with a basic analysis of the same logs. Without that comparison, the extra cost cannot be justified.

What could stop it

What the evidence shows

• Only 16% of surveyed councils judged their DRT service sustainable, and 95% said it ran at a loss (council-reported, 24 responses)³.
• Efficiency is low and varies widely: journeys per vehicle hour ranged from 0.9 to 3.9 across schemes, and the empty running ratio was 0.77 in April to September 2024⁵.
• The effect on bus use was mixed: Warwickshire's bus journeys fell 11.9% while those in its comparison area rose 20.4%⁵.
• A pilot in North Yorkshire cost £15.57 per passenger against about £6.50 on timetabled routes, and ended after 14,208 passengers¹⁰¹¹.
• No study was found that quantified a bus network redesign made from DRT booking data. The one example found (Newport) came with no figures⁸.
• No figure was found that separates the gain from machine learning from the gain from ordinary optimisation in rural bus services.

Reasoning

The challenge question is what would make this possibility less likely, less useful or more harmful than it first appears.

The data may be biased. A booking log records the journeys of people who know the service exists, own a smartphone or are willing to phone, and live within its area. A route designed from it might serve those who already use the service and miss those who have given up on it. The 11% of journeys booked by phone⁵ may include people who would be hardest to reach by other means.

The incentives may point the wrong way. An operator paid per trip or per hour may have little reason to hand over data that would help the council replace its service with a fixed route. If so, access would depend on contract terms set before any redesign was planned.

Who pays and who bears the risk. A new route would put cost and risk on a council that already reports losses³. If the fixed route failed, residents who relied on the earlier on-demand service could be left with neither. This would need to be set against the gain.

A simpler fix may do the job. The efficiency problem may be density, not routing. If the main constraint is that few people live within reach, no analysis would raise demand, and the more useful decisions would be about funding and which services to protect.

Lessons from failed pilots. The YorBus figures¹¹ suggest a cost gap can open between a trial and the timetabled service it was meant to complement. A data-led redesign would need to show it can close that gap, not only describe it.

Scope of the evidence. The evidence is for England only, from a small council survey and a single national evaluation. Conditions in other nations could differ.

Signs to watch

• A published example of a council redesigning a fixed route from DRT booking data, with passenger numbers before and after.
• The Department for Transport or a later evaluation reporting on how booking data is held, shared or reused.
• Guidance or case law on whether on-demand services can be treated as socially necessary under the Bus Services Act 2025⁹.
• Evidence of data-sharing agreements between DRT operators and councils, including requirements to hand over logs of unmet requests.
• Longer-term funding settlements for rural bus services, which survey respondents have asked for³.
• A comparison of a simple analysis of booking logs with a machine learning forecast on the same data.