Delivery route optimisation

The process starts each morning when orders are pulled from the system and require delivery route planning. The dispatcher reviews these orders, validates that all delivery addresses are usable and correctly formatted, then evaluates a set of constraints for each delivery including time windows, vehicle capacity limits and customer-specific requirements. Using experience and a mental map of the territory, the dispatcher manually sequences the stops to minimise travel distance while respecting all delivery windows and constraints. Once a route is drafted, it is validated against kilometre targets to ensure efficiency. Finally, the optimised route is communicated to the drivers for execution. This happens a few times per week in bursts, roughly 260 times per year, with each planning session taking about 75 minutes.

Around 25% of cases involve non-standard delivery requirements, invalid or ambiguous addresses, conflicting time window constraints, or data entry errors that surface during planning. These exceptions require judgment calls and sometimes trigger rework when earlier mistakes are discovered. The dispatcher relies on a mix of an older MES system with a readable database, machine logs, sensor data, email threads and spreadsheets to coordinate the planning work. Product safety and traceability are critical compliance requirements because routing errors that send products to the wrong address can trigger recall conversations.

AI Workflow is the right fit for this delivery route optimisation process because it directly matches what the client wants to achieve: automate the boring, repetitive 75% of cases while keeping the dispatcher in the loop for the 25% that need human judgment. The process has clear structure with orders, addresses, delivery windows and capacity constraints as inputs, but routing heuristics and exception handling logic currently live in the dispatcher's head. An AI workflow can encode the standard constraint checks, validate addresses against a geocoding service or master list, generate candidate routes using an open-source routing library like OR-Tools or OSRM, and then escalate non-standard cases to a review queue for the dispatcher. This approach preserves the dispatcher's expertise exactly where it matters while freeing up time spent on routine work.

The client's preferred tech stack makes this recommendation even stronger. They have an in-house automation engineer with PLC experience who is ready to learn Python, which is the natural language for building this kind of orchestrated workflow. The engineer can write scripts to pull order and address data from the existing MES database, implement a rules engine for constraints, integrate a routing library, and build a lightweight escalation interface for exception handling. Because the workflow is code-based rather than a black-box AI agent, the engineer retains full control and can tune the logic as delivery patterns evolve. The MES database is readable and machine logs are available, so integration is feasible even without formal APIs.

The compliance and safety requirements also favour AI Workflow over more autonomous alternatives. Routing errors that send products to the wrong location can trigger recalls, so the system needs human checkpoints on anything non-standard. An AI Workflow with explicit escalation rules provides that safety valve, while a fully autonomous AI Agent would be too risky given the 25% exception rate and the potential for address validation mistakes to propagate into the field. At the same time, staying fully manual wastes 325 hours per year on work that is largely repetitive and does not scale as the business grows. Traditional code alone could handle the standard cases, but encoding all the routing heuristics and exception logic up front would be brittle and high-maintenance. AI Workflow lets the client start with rules-based constraint checking and routing, then layer in smarter address validation or route suggestion models over time as confidence builds.

The hybrid path with mandatory human review gates on every run also scored well and would be a sensible fallback if the client wants extra caution during rollout. However, the client explicitly stated they want routine cases to run themselves, and hybrid implies the dispatcher reviews everything rather than just exceptions. AI Workflow strikes the right balance between automation and oversight, delivering time savings on the majority of runs while keeping the dispatcher engaged where judgment is truly needed.

AI Workflow scores 9 out of 10 and is the clear winner for this delivery route optimisation process. It handles the routine 75% of cases automatically while escalating the 25% of non-standard scenarios to the dispatcher for review. This matches the client's explicit goal of automating boring repetitive work without removing human judgment from complex decisions. The workflow approach integrates naturally with the readable MES database and fits the automation engineer's Python skillset. Hybrid scores 8 and is also viable, but it requires human review gates on every single route, even the straightforward ones. That works well for high-risk processes where every output must be checked, but here the client wants routine cases to handle themselves and free up dispatcher time for the genuinely tricky situations. Traditional Code scores 6 because it could work for standard cases if all routing heuristics and constraints were fully documented and stable. The problem is that much of the routing logic currently lives in the dispatcher's head, and a pure rules engine would be brittle as new edge cases emerge. The client would spend more time maintaining rule definitions than they save in automation. RPA and AI Agent both score poorly and are not recommended. RPA cannot replicate the dispatcher's spatial reasoning or handle the scattered data sources, and an autonomous AI Agent poses too much risk in a process where routing errors can trigger product recalls and compliance issues.

The implementation starts with data consolidation and address validation. The automation engineer builds Python scripts to pull order and address data from the MES database each morning when orders arrive. These scripts validate addresses against a geocoding API like Google Maps or Here, or against a master address list if the client has one. Running this on two weeks of historical data will surface data quality issues early and establish a clean input pipeline. The next phase builds the constraint engine and routing logic. The engineer implements a rules engine in Python to encode delivery windows, vehicle capacity limits and distance minimisation goals. An open-source routing library like Google OR-Tools or OSRM generates candidate routes that respect these constraints. The output is validated against a sample of dispatcher-planned routes to tune the heuristics and build confidence that the system understands the real-world problem. This is where the dispatcher's tacit knowledge gets surfaced and encoded. Exception detection and escalation come next. The system flags cases that do not fit the standard pattern, such as ambiguous addresses, conflicting delivery window constraints, or unusual order volumes. These flagged cases go into a review queue, which can be as simple as an email with route details and a link to approve or modify. The dispatcher reviews exceptions, makes adjustments and logs the reasoning so the system can learn over time. This keeps human judgment in the loop exactly where it is needed. The pilot phase runs the AI workflow in parallel with manual planning for four weeks. The dispatcher continues planning routes the old way but also reviews the AI-generated suggestions and logs any discrepancies. This feedback loop refines the constraint logic, tunes the exception triggers and validates that the system is ready for production. Once the pilot proves stable, the system goes live for routine cases with the dispatcher handling only escalated exceptions. Logging and monitoring track route quality metrics like total kilometres, on-time delivery rates and exception escalation frequency. Quarterly reviews allow the client to adjust rules as delivery patterns and customer requirements evolve.