Start With Dispatch, Not the Driverless Headline
The immediate opportunity for most UK fleets is not an unsupervised lorry. It is a better dispatch decision: which vehicle should carry which load, in what sequence, over which roads, with enough time for safe driving, unloading and recovery from disruption.
That distinction matters in July 2026. The Department for Transport says the full Automated Vehicles Act implementation programme is being built for 2027. A separate 2026 permitting scheme supports automated passenger service pilots; it is not a general authorisation for autonomous freight. The Civil Aviation Authority’s new BVLOS delivery model likewise targets routine beyond-visual-line-of-sight drone operations in 2027.
A credible transport-AI plan should therefore deliver value with today’s vehicles and rules, while producing the controls and evidence that higher autonomy will require later.
Separate Three Control Loops
“Route optimisation” often hides three different systems. Give each its own owner, data and release gate.
| Loop | Decision | Normal cadence | Safe fallback |
|---|---|---|---|
| Planning | Fleet, depot, load and route plan | Night before or start of shift | Dispatcher-approved baseline |
| Dispatch | Reassign a stop after traffic, delay or cancellation | Minutes | Hold current route and alert control |
| Motion | Control steering, braking or flight | Milliseconds | Vehicle-specific minimal-risk response |
A planning model can be useful even when every recommendation needs approval. A dispatch model changes live work and needs rapid human intervention. A motion system is safety-critical vehicle automation and belongs inside the applicable vehicle or aviation assurance regime. Do not use a strong planning pilot as evidence that the motion layer is safe.
For the warehouse and yard hand-off, see our guide to AI warehouse automation. For the wider network context, smart-city transport systems explains why road and infrastructure data cannot be treated as perfect ground truth.
Optimise a Feasible Plan Before a Fast One
The shortest route is often illegal, unsafe or operationally useless. Model hard constraints separately from preferences.
| Type | Examples | Treatment |
|---|---|---|
| Legal | Applicable drivers’ hours, working time, vehicle authorisation | Never violate |
| Physical | Height, weight, width, turning access, charging or fuel range | Never violate |
| Safety | Suitable road type, weather restriction, secure stop, fatigue risk | Hard constraint or safety-owner approval |
| Customer | Delivery window, access booking, service priority | Hard or contract-scored |
| Operational | Depot hours, loading sequence, skills, refrigeration, equipment | Hard where the job cannot proceed |
| Optimisation | Distance, empty miles, energy, tolls, lateness, workload balance | Weighted only after feasibility |
The Health and Safety Executive says employers should choose the safest route for the vehicle, account for weather and schedules, avoid unsuitable restrictions and consult drivers about breaks and welfare facilities. The DVSA’s goods-vehicle drivers’ hours guidance, updated 1 July 2026, also makes operators responsible for organising compliant work. A solver must encode the rules that apply to the vehicle and journey; “the AI planned it” is not a compliance defence.
Give every constraint a source, owner, jurisdiction, version and effective date. If bridge-height data is missing or a delivery window conflicts with a required break, the system should reject the plan or escalate—not silently trade safety for punctuality.
Build a Transport Data Contract
A route is only as trustworthy as the events underneath it. Define a minimum contract before selecting an optimisation algorithm.
Orders and stops: stable identifiers, location confidence, time windows, service duration, load attributes, priority and cancellation status.
Vehicles: dimensions, authorised weights, capacity, energy or fuel state, maintenance status, equipment, depot and operating restrictions.
People: licence and competency requirements, applicable hours regime, completed work, planned breaks and availability. Store only what is needed for scheduling.
Network: road restrictions, closures, forecast and observed travel time, weather, charging or fuelling availability, and the timestamp and confidence of each feed.
Execution: actual arrival, departure, proof-of-service, failed-delivery reason, override and incident.
The government’s June 2026 research on transport Smart Data identified journey planning, network management and freight logistics as candidate use cases, but also documented cross-cutting data challenges. Treat external feeds as fallible: monitor age, missingness, geographic coverage and disagreement rather than merging them into a single unexplained “traffic score”.
Make Every Replan Reconstructable
Consider an illustrative—not claimed—operation with two vans, six timed stops and one late inbound load. At 09:10, congestion raises the predicted journey time to the last three stops.
A useful system should record:
- the original orders, constraints and data versions;
- the event that triggered replanning;
- candidate plans rejected for capacity, hours or access;
- the proposed reassignment and expected effect on each stop;
- confidence and assumptions, including uncertain service time;
- who approved or overrode the change; and
- actual outcomes.
If moving Stop 5 protects four delivery windows but leaves no compliant break location, that candidate is infeasible. If two feasible plans differ by five minutes but one sends a large vehicle down an unsuitable minor road, the safety preference should dominate. HSE guidance is explicit that route selection must account for the type of vehicle, not merely estimated arrival time.
This decision record is more valuable than a colourful live map. It lets operations explain a missed slot, reproduce a recommendation and discover whether error came from demand data, travel-time prediction, the solver or execution.
Keep the Driver in the Safety System
Optimisation can create risk by compressing schedules, issuing distracting replans or making a technically achievable plan feel mandatory. Define:
- no in-motion interaction beyond the approved driver interface;
- a safe acknowledgement point for material route changes;
- sufficient time for checks, breaks, loading and foreseeable delay;
- a clear right to reject an unsafe instruction;
- escalation when the plan becomes infeasible; and
- no performance ranking based only on algorithmic ETA.
The Traffic Commissioners’ March 2026 note on drivers’ hours and responsibility recommends route tools that include traffic, rest-stop availability and legal breaks, while stressing that fatigue risk still depends on conditions and the person.
Telematics may also monitor workers. The ICO says organisations must balance monitoring with workers’ rights and use proportionate controls in its worker-monitoring guidance, which is currently under review following the Data (Use and Access) Act. Document the purpose and lawful basis, minimise retention and access, explain monitoring, and complete a DPIA where risk requires one.
Treat Road Autonomy and Drone Delivery as Separate Programmes
The Automated Vehicles Act creates an authorisation and accountability framework for vehicles that can legally drive themselves. Its safety principles must aim for authorised vehicles to be at least as safe as careful and competent human drivers. Until the framework and relevant authorisations apply, a driver-assistance feature should not be described or operated as self-driving.
Drone delivery has different hazards: airspace integration, detect-and-avoid, ground risk, communications loss and operator competence. The CAA’s current BVLOS roadmap describes temporary reserved airspace and specific or certified category pathways while routine integration is developed. A good road-routing result does not validate an aerial operation. Our separate drone delivery and inspection guide covers that assurance boundary.
Roll Out Through Bounded Authority
| Stage | System authority | Evidence required to advance |
|---|---|---|
| Shadow | Scores historic and live work without changing it | Data coverage, feasible-plan rate, replay accuracy |
| Recommend | Dispatcher approves every plan | Acceptance and override reasons, no hard-constraint breach |
| Bounded dispatch | May reassign low-risk stops inside fixed limits | Reliable rollback, alerting, stable safety and service metrics |
| Expanded operation | Wider areas, fleets or conditions | Independent review, incident exercises, supplier and cyber assurance |
At every stage, preserve a known-good planning method and the ability to freeze automatic replanning. Connected transport is cyber-physical: the NCSC’s Connected Places Security Principles call for privilege management, supplier control, monitoring, lifecycle management and incident recovery. Protect dispatch credentials, sign or verify configuration, separate test and production, rate-limit actions and rehearse loss of maps, telematics and optimisation services.
Use Gates That Operations Can Enforce
Agree thresholds before the pilot. A defensible scorecard includes:
- Safety and compliance: zero observed hard-constraint violations; no deterioration in collisions, near misses or fatigue reports; all material replans auditable.
- Service: on-time-window rate and failed-delivery rate at least non-inferior to the matched baseline, with results split by depot, route type and shift.
- Prediction: stop-level ETA error and service-time error reported at median and p90, not hidden inside one average.
- Efficiency: distance, empty miles, energy or fuel, paid hours and cost per successful stop, including software and dispatcher review.
- Human factors: recommendation acceptance, override reasons, alert load and driver-reported usability.
- Resilience: p95 plan time, feed-staleness alarms, successful failover and tested time to manual operation.
Run the pilot across ordinary days and known stress conditions. Do not count a cancelled stop as an efficiency gain, or a shorter route as a success if work is transferred to overtime, another depot or an unsafe schedule.
AI route optimisation is ready when it repeatedly produces feasible, explainable plans under real constraints. Higher autonomy should be earned as a separate safety case—not inferred from a dashboard showing fewer miles.
Keep the approved baseline route available for every live shift.
