.webp)
A UK last-mile ops director we spoke to last week ran 84 vehicles across three depots. His dispatcher planned routes on paper, drivers ignored the plan, and his fuel bill was climbing 6 percent a year despite volume being flat.
He had OptimoRoute and Onfleet quotes on his desk and a proposal from an internal engineer to build a custom platform on Google OR-Tools.
That is the AI route optimisation buyer story in 2026. The technology is real. The savings are real, if smaller than vendor slides claim. And the mismatch between what dispatchers do today and what the algorithm assumes is where most programmes stall.
This article is the honest read. What AI route optimisation actually means in 2026. Real fuel and mileage savings. The vendor landscape. When to buy, when to build on Google OR-Tools, and when hybrid patterns are the answer. And what the algorithm cannot fix.
What "AI Route Optimisation" Actually Means in 2026
Two components make up modern AI route optimisation.
Classic Vehicle Routing Problem solvers. The mathematical core has been solved for decades. Take a set of orders with time windows, capacity constraints, and driver break requirements. Compute optimal routes across available vehicles. Libraries like Google OR-Tools, VROOM, and Jsprit handle this in production at scale.
Machine learning on top. Predict actual travel time given weather, traffic, day of week, and driver behaviour history. Forecast demand for capacity planning. Re-route dynamically on the day when a driver runs late, a new order arrives, or a customer misses their slot. This is where the 2026 uplift over 2020-era software comes from.
The word "AI" in route optimisation means the combination. Pure VRP without ML is 2015 technology. Pure ML without VRP does not solve the routing problem. Modern platforms and modern custom builds do both.
Real Fuel and Mileage Savings: Claims vs Delivered
Vendor slides typically claim 30 to 40 percent savings on miles, fuel, or route time. Real delivered savings in production sit lower.
Realistic 2026 benchmarks for a properly deployed programme:
Miles driven: 10 to 20 percent reduction
Stops per vehicle per day: 15 to 25 percent increase
Fuel cost: 5 to 10 percent reduction
On-time delivery rate: 5 to 12 percentage point improvement
Dispatcher time: 30 to 60 percent reduction on route planning
Those numbers assume the operator does the boring work: cleans up the master data (addresses, service times, vehicle capacity), integrates telematics for actual GPS traces, and holds dispatcher override rate under 20 percent.
The gap between vendor claims and delivered savings is usually not the algorithm. It is one of three causes: bad master data (drivers cannot follow routes based on wrong service times), promise-slot logic that overpromises what routing can deliver, or dispatcher override rate above 30 percent because operations does not trust the tool.
The 2026 Vendor Landscape
Four vendor categories cover the market.
SaaS platforms. OptimoRoute, Route4Me, WorkWave, Onfleet, Bringg, DispatchTrack. Priced $15 to $45 per vehicle per month plus per-delivery fees. Best for operators under 100 vehicles who want speed to launch and standard workflows. Setup lands 2 to 8 weeks.
Google Route Optimization API. Production API for multi-vehicle route optimisation with constraints (time windows, load limits, driver breaks). Pay-per-use pricing scales with volume. Best for teams building custom platforms who want reliability and Google's mapping data.
Open source solvers. Google OR-Tools, VROOM, Jsprit. Free libraries with strong communities. Best for teams with strong engineering that want full control over customisation. Requires meaningful investment to productionise.
Enterprise TMS integrations. Descartes, Aptean, Manhattan Associates. Route optimisation bundled into broader TMS. Best for larger operators already running the parent TMS platform.
For most UK and US operators under 100 vehicles, a SaaS platform is the right answer. Above 100 vehicles with unusual constraints, custom build on OR-Tools or Google Route Optimization becomes defensible.
When to Use Google OR-Tools and When to Buy
Three tests for the buy vs build decision.
Volume test. Below 50 vehicles, SaaS wins on maths and speed. Between 50 and 100 vehicles, run both quotes and compare. Above 100 vehicles, the SaaS bill starts to justify a custom build.
Constraint test. Do your routes have constraints that SaaS handles poorly? Multi-day tours with overnight stops. Vehicle-specific driver certifications (hazmat, ADR). Load compatibility rules (do not mix certain SKUs on the same van). If any of these, custom becomes more attractive.
Integration test. Does your operation need routes to feed into a bespoke ERP, WMS, or field service app the SaaS platforms cannot integrate with cleanly? Custom lets you build the integration once and own it.
If none of the three tests points to custom, buy off-the-shelf. Do not build custom for the vanity of ownership.
Realistic Cost Bands for Buy vs Build
SaaS. $15 to $45 per vehicle per month, plus $0.05 to $0.15 per delivery in per-transaction fees. For a 50-vehicle operator doing 20,000 deliveries per month, monthly bill lands $2k to $5k. Setup: 2 to 8 weeks. Time to value: 3 to 4 months.
Custom on Google OR-Tools or Route Optimization API.
Proof of concept: £25k to £60k over 6 weeks
Pilot: £60k to £150k over 3 to 5 months (one depot, live data, dispatcher UI)
Production: £150k to £400k over 5 to 9 months (multi-depot, full driver app, dispatcher dashboard, analytics)
Add £500 to £2,000 monthly hosting and API usage on Google Route Optimization. Add engineering time for ongoing tuning: £2k to £8k monthly.
Break-even for custom against SaaS typically lands at 100 vehicles depending on per-delivery volume. Below that, SaaS wins. Above that, custom pays back over 2 to 3 years.
Our earlier post on last mile delivery software development covers the broader delivery-platform economics.
What We Learned Building Delivery Tracking in TrackVid
TrackVid is our video proof and claim management platform. It handles delivery event reconciliation, GPS trace matching, and marketplace-to-carrier data normalisation. Two lessons transfer to route optimisation builds.
Actual vs planned deviation is where the savings live. In TrackVid we track planned delivery time against actual delivery time for reconciliation. In a route optimisation programme, the same telemetry is your feedback loop. If drivers consistently deviate by more than 10 minutes on 30 percent of stops, the routing model is wrong somewhere and no amount of "AI" fixes it. Fix the master data.
Dispatcher trust is a slow curve. In TrackVid, sellers wanted every AI-generated classification reviewed for the first month. Two months in they trusted high-confidence cases. Six months in they let it act on 70 percent of routine work. Same pattern applies to dispatchers. Do not deploy a routing engine and expect zero override rate on day one. Design for gradual trust with feedback loops.
You can see TrackVid at our portfolio. If you want to talk about routing for your specific fleet, book a route optimisation call with WhiteStone.
Common Failure Modes
Four failure modes we see across route optimisation deployments.
Bad master data. Wrong address geocodes, wrong service times, wrong vehicle capacity. The algorithm assumes clean data. If the data is dirty, savings do not materialise.
Overpromising delivery windows. Sales offered 1-hour windows the routing cannot honour at scale. Missed windows cascade into complaints and dispatcher overrides.
Dispatcher override rate above 20 percent. If dispatchers override more than 1 in 5 routes, either the algorithm has a real gap or dispatchers do not trust it. Both need investigation.
Driver route deviation. Drivers do not follow the planned route because they know a better one. Fair enough. Feed that data back to the model.
Frequently Asked Questions
How much fuel does AI route optimisation actually save?
Real 2026 fuel savings from properly deployed programmes: 5 to 10 percent. Vendor claims of 20-plus percent assume perfect master data and zero dispatcher override, which no real operation has. Assume 7 to 8 percent for planning purposes.
Google OR-Tools vs commercial platforms: which for what?
OR-Tools is free and flexible but needs engineering to productionise. Commercial SaaS (OptimoRoute, Route4Me, WorkWave) trades cost for speed to launch. Below 50 vehicles: SaaS wins. Above 100 vehicles or with unusual constraints: custom on OR-Tools becomes defensible. Between: run both quotes.
How do you handle real-time re-routing?
Modern route optimisation supports dynamic re-routing when a new order arrives or a driver runs late. Google Route Optimization API, OptimoRoute, and Bringg all support live re-planning. Custom builds need this designed in from day one; retrofitting real-time capability to a batch VRP system is expensive.
How long is time-to-value for a new deployment?
SaaS: 3 to 4 months from contract to measurable ROI (2 to 8 weeks setup, 6 to 12 weeks for dispatchers and drivers to trust the tool). Custom build: 6 to 12 months from kickoff to production, plus a further 3 to 6 months for full ROI to land.
Does AI route optimisation work for field service, not just delivery?
Yes, with additional constraints. Field service adds skill-matching (technician has to be qualified for the job type), parts inventory on the vehicle, and long stops (2 to 4 hours vs 5 to 15 minutes for delivery). Vendors like WorkWave and OptimoRoute have field-service specific tiers.
The One Thing to Remember
AI route optimisation is real in 2026, and the savings are real if smaller than vendor slides. The programmes that deliver clean up the master data first, keep dispatcher override rate under 20 percent, and design for gradual trust building. The programmes that fail blame the algorithm for problems that are actually data quality or operations culture.
If you want a candid conversation about routing for your specific fleet, browse our AI development services or come to the scoping call.
