The Protected Constraint: artificial intelligence in haulage, warehousing and distribution

Artificial intelligence in logistics forecasts well but knows nothing of what it is forbidden to do, which is why AI route optimisation belongs behind a written set of limits.

Nothing in this article constitutes legal or regulatory advice. It is offered to support discussion and further review with qualified advisers.

What does artificial intelligence in logistics actually change for a haulier?

Prediction improves while permission stays where it was. A forecasting model can estimate demand, dwell time and arrival risk with real accuracy, yet drivers’ hours compliance, load security and site safety rules still bound every instruction it produces. The gain arrives when those limits sit outside the model, versioned and owned.

• Empty running across UK road freight stood at 30 per cent of vehicle distance in the year to March 2026, or 5,714 million kilometres, and has held flat for years.
• The Health and Safety Executive stated on 12 June 2026 that the goal-setting nature of the Health and Safety at Work etc. Act 1974 means it applies whatever technology an operator uses.
• Statutory guidance from the Senior Traffic Commissioner treats continuous and effective management by a transport manager as a serious requirement.
• Published roadworthiness guidance keeps the scheduled safety inspection in place whatever a predictive system reports.
• Adoption figures conflict, since official statistics put artificial intelligence use among UK businesses with 10 or more employees at about 35 per cent while a vendor survey reported that only 3 per cent of shippers and logistics providers made no use of it.

What do the national numbers show about artificial intelligence in logistics?

Freight is measured, which makes the claims testable. Goods vehicles registered in Great Britain moved 164 billion tonne kilometres in the year to March 2026 and covered 19.0 billion vehicle kilometres, with goods moved and vehicle distance each falling by 1 per cent. Empty running stood at 30 per cent of total vehicle distance, or 5,714 million kilometres. Through a decade of optimisation software, that share has proved stubborn.

The strongest independent evidence concerns accuracy alone. In the M5 forecasting competition, 5,507 teams forecast 42,840 daily sales series, and the winning machine learning method beat the best statistical benchmark by 22.4 per cent. Four qualifications travel with it, because the result measured forecast accuracy and no operational outcome, the data was United States grocery retail, and the gains came from gradient-boosted trees, a different technology from the generative systems now being sold. Machines forecast well. Whether anything arrives sooner remains untested.

Why should AI route optimisation wait for a constraint library?

A constraint library is a versioned, tested, human-owned statement of everything an operation is forbidden to do, held outside every model and applied to every recommendation before a person sees it. Six categories cover most operations, which are statutory limits, vehicle and equipment limits, site rules, contractual commitments, competence rules and absolute prohibitions. A named internal owner matters as much as the content. A rule held only in a supplier’s configuration file is a rule the operator cannot evidence.

Underneath that artefact sits a distinction the sector loses easily. A forecast is a statement about the future with a confidence attached. An operational fact is something that happened, with a source behind it, and an instruction is a thing a person is being asked to do. A planner handling 300 loads has seconds per decision and will take the recommendation because it is there. A planning system capable of producing a non-compliant plan will produce one, on a Friday in December, and the planner who accepts it answers for it.

How far does drivers’ hours compliance bind a scheduling system?

The limits are unchanged in substance, though the vocabulary has moved to the assimilated rules. Daily driving is capped at nine hours, extendable to 10 hours twice a week, with 56 hours in a week and 90 hours across any two consecutive weeks. Breaks totalling at least 45 minutes follow no more than four and a half hours of driving, and daily rest runs to 11 hours. Any recommendation that breaches one of those is a defect.

Governance and recording are separate things, and a case decided in July 2026 shows the distance between them. A Traffic Commissioner revoked an operator’s licence and disqualified its sole director and transport manager indefinitely after he repeatedly falsified drivers’ hours records using another driver’s tachograph card. The recording technology worked exactly as designed throughout. There were 66,222 goods vehicle operator licences in force in Great Britain in the year to 31 March 2025, authorising 376,044 vehicles, and each rests on a named person exercising continuous and effective management.

What does the regulator already say about warehouse automation safety?

The safety regulator has stated its position, and it is framed around effects. In a policy statement of 12 June 2026 the Health and Safety Executive noted that the goal-setting nature of the Health and Safety at Work etc. Act 1974 means it applies regardless of the technology in use. No new duty has appeared, because the existing one already reaches this. On provisional figures for 2025 and 2026, 126 workers were killed in work-related accidents in Great Britain, and 24 of them were struck by a moving vehicle.

Worker monitoring is the live legal question, and the ground has moved in two directions. The Data (Use and Access) Act 2025 widened the lawful bases available for significant automated decisions, while the Information Commissioner’s safeguards on transparency, human intervention and contestability remain. A government consultation on algorithmic management at work closes on 30 September 2026. European employment rules on high-risk systems, deferred to 2 December 2027, describe warehouse task allocation, pick-rate monitoring and driver scoring rather closely.

What does fleet AI governance look like in the first 90 days?

Shadow mode is the most useful practice in this sector, and it is rarely used. The system produces its recommendation while the operation carries on as it would have done, and both are recorded, so that after four weeks a dataset exists showing what the system would have done, what the operation did and what happened. An evidence gate then tests outcome, constraints, judgement, safety, security and economics. A single breach of the constraint library fails the pilot.

Controls should scale with what the output touches and how reversible it is. Internal analysis sits at tier one. Recommendations acted on by planners or drivers, and anything that monitors a person, sit at tier three, which calls for constraint checking before display and worker consultation. Anything that controls equipment or sets a person’s work rate sits at tier four, where a tested fallback to manual working applies. Ransomware attacks on two British retailers in April 2025 cost between £270m and £440m across affected parties, on one independent assessment. The economics run the same way, since the technology moves cost from producing a plan to checking one, and that checking happens in the minutes before a vehicle leaves.

Eight questions settle most proposals, and the fifth is the one this argument adds. Which hard limits apply, where are they held, and who owns them? In most operations that question has no answer at the start, because the limits live in a transport manager’s head, in a supplier’s configuration file and on a laminated notice by a loading door.

Write them down in one place and give them a named owner, and the rest of the programme becomes tractable. AI route optimisation can then be tested against something concrete, drivers’ hours compliance becomes a machine check costing milliseconds, and warehouse automation safety is assessed against the site, the vehicle and the driver together. Fleet AI governance then becomes an artefact that outlives any supplier.

All of this exists to protect one moment, which is the moment a person decides that a vehicle may go. Behind that decision sits a transport operator licence, and the evidence a Traffic Commissioner or an inspector will ask for is the record of how the limits held. Build it first.

Frequently asked questions

Does artificial intelligence in logistics actually reduce empty running?

Empty running stood at 30 per cent of UK vehicle distance in the year to March 2026, or 5,714 million kilometres, and that share has held flat through a decade of optimisation software. No published evaluation establishes a causal operational benefit in British road freight or warehousing, so the effect remains testable and untested at national scale.

Can an AI route optimisation system schedule a driver beyond the legal hours?

Good design makes it impossible. Daily driving is capped at nine hours, extendable to 10 hours twice a week, with 56 hours weekly and 90 hours across any two consecutive weeks. Every recommendation belongs against those limits before a planner sees it, held in a versioned library outside the model, so that a single breach fails a pilot.

Who is accountable when an automated system contributes to a compliance failure?

The named transport manager and the licence holder remain accountable. Statutory guidance from the Senior Traffic Commissioner describes continuous and effective management as a serious requirement, and that duty stays with the named person whatever technology assists. A Traffic Commissioner revoked a licence and disqualified a director indefinitely in July 2026 over falsified records, while the recording equipment worked as designed.

Does predictive maintenance allow longer intervals between safety inspections?

Published guidance keeps the existing intervals in place. Electronic braking performance monitoring is recognised for trailers, where it stands in place of the laden roller brake test and requires evaluation of each report by a competent person. Walkaround checks, scheduled safety inspections and 15 months of retained records all remain, and an operator considering a change would be wise to ask the Traffic Commissioner’s office first.

What does fleet AI governance require where warehouse workers are monitored?

Workers should be told they are subject to automated decision-making and given meaningful information about the logic involved and its envisaged consequences. Human reviewers need the authority and competence to go against a recommendation. A data protection impact assessment, a stated purpose limit and early consultation all apply, and a government consultation testing a statutory duty to consult closes on 30 September 2026.