Innovative AI Solutions | AI Development, Web & Mobile Apps – Delhi, India
LOGISTICS · ROUTE OPTIMIZATION · AI DISPATCH

AI Route Optimization — 20% Fuel Savings, 96% On-Time

We built an AI-powered daily route planning and dynamic re-routing system for a 300-vehicle 3PL company — optimizing 8,000+ deliveries daily across 6 cities, factoring in traffic, time windows, vehicle capacity, and driver shifts.

20%Fuel Cost Reduction
96%On-Time Delivery (was 78%)
85%Faster Route Planning
₹2.4CrAnnual Cost Saving
Build Logistics AI All Case Studies

Why Indian 3PL Logistics Can't Absorb Fuel Inefficiency Any Longer

India's third-party logistics sector is navigating an unusually difficult cost environment. Fuel — diesel primarily — represents 30–40% of a 3PL's total operating cost structure, according to the Logistics Sector Report published by the Ministry of Commerce. With diesel prices rising 28% between 2020 and 2024 and no corresponding increase in the freight rates that FMCG and e-commerce clients will accept (pricing pressure from aggregated freight platforms like Porter, BlackBuck, and Rivigo has kept rates flat), fuel efficiency has become the single most actionable variable in 3PL profitability. The arithmetic is unforgiving: a 300-vehicle fleet consuming an average of 15 liters per 100 km, covering 187 km per vehicle per day, burns through 84,150 liters of diesel daily. At ₹94 per liter, this is ₹79 lakh per day — ₹24 Cr per month — in fuel alone. A 20% reduction in distance driven through route optimization translates directly to ₹4.8 Cr per month in fuel savings. Beyond fuel, the last-mile delivery landscape in Indian cities has become dramatically more complex. Urban congestion has worsened in every major metro: average commercial vehicle speeds in Mumbai's business districts have fallen to 8–12 km/h during peak hours; Bengaluru's outer ring road averages 14 km/h during morning rush. Planning routes on the assumption of free-flow traffic — which is what a human route planner does when drawing routes on a map — leads to systematic under-estimation of travel time and systematic late deliveries. The on-time delivery performance gap is an existential client retention issue. FMCG principals with robust modern supply chains (Unilever, Nestle, P&G) have moved to tight delivery windows with penalty clauses: ₹500–2,000 per late delivery, with repeat failure thresholds triggering contract termination reviews. For a 3PL handling 8,000 deliveries per day at a 78% on-time rate, that means 1,760 late deliveries every day — a relentless drain on client relationship quality and a constant source of penalty exposure.

30–40% of Costs = Fuel

Fuel accounts for 30–40% of total 3PL operating costs per Ministry of Commerce data — the single largest controllable expense category. Unlike labor (constrained by minimum wage) or insurance (constrained by vehicle value), fuel consumption is directly reducible through better route planning. A 20% route distance reduction translates to a 20% fuel saving that drops straight to operating profit — no offsetting cost increases.

🏙️

Last-Mile Complexity Rising

Urban congestion, apartment building access restrictions, commercial vehicle time restrictions (heavy trucks banned from city centers 9am–9pm in Mumbai, Delhi, Bengaluru), and the explosion of e-commerce delivery density have made Indian last-mile routing exponentially more complex than it was 5 years ago. Human planners, working from experience and intuition, cannot simultaneously optimize for traffic, time windows, vehicle type restrictions, and driver shift constraints across 300 vehicles.

📉

FMCG Penalty Clauses Tightening

India's large FMCG principals have adopted global supply chain standards for their 3PL contracts. On-time in-full (OTIF) thresholds of 95%+ are now standard in category-A FMCG contracts, with penalty clauses typically ranging ₹500–5,000 per failing delivery. For a 3PL at 78% OTIF on 8,000 daily deliveries, cumulative penalty exposure runs ₹15–40 lakh per month — often not fully charged by clients who want to preserve the relationship, but always hanging over the commercial discussion as a source of revenue risk.

Manual Route Planning = Wasted Fuel and Late Deliveries

The 3PL company's 6 dispatch planners spent 3–4 hours each morning manually creating delivery routes for 300 vehicles. Routes were based on experience and intuition — not real-time traffic, delivery time windows, or load optimization. The result: vehicles often crossed paths, backtracked unnecessarily, and arrived late at time-sensitive deliveries. Route quality depended entirely on which planner was on duty and their familiarity with a particular city zone — creating significant variability between days and between cities.

On-time delivery performance of 78% was triggering penalty clauses with their FMCG clients. Fuel bills had grown 35% in 2 years without a proportional increase in deliveries. And when a driver called in sick or a vehicle broke down, replanning for that zone took 45+ minutes — during which time 30 deliveries were falling behind. The company was effectively losing a full productive hour every morning to route planning alone, before a single delivery began.

The vehicle utilization picture was equally poor. Average vehicle fill rate was 63% — meaning 37% of vehicle capacity was travelling empty on any given day. Smarter route sequencing would allow the same delivery volume to be handled with fewer vehicles or with existing vehicles covering significantly fewer kilometers. The company had deferred a fleet expansion that client volume growth should have triggered — but was reluctant to invest in more vehicles when they suspected the existing fleet was being used inefficiently.

Driver satisfaction was also a concern. Poorly planned routes caused drivers to spend hours stuck in predictable traffic they could have avoided with a different sequence. Drivers who knew their zones well were frustrated when dispatched routes contradicted local knowledge they had built over years. Some drivers had developed their own informal "workaround" routes that were demonstrably better than what the dispatch team produced — but this tribal knowledge was never systematically captured or applied. A good AI routing system would incorporate the best insights from experienced drivers through the deviation feedback loop rather than ignoring them.

AI That Plans and Adapts Routes in Real Time

A vehicle routing optimization system that plans all 300 routes in under 5 minutes and re-routes dynamically when conditions change.

01

AI Route Optimizer

The core of the system is a Vehicle Routing Problem with Time Windows (VRPTW) solver built on Google OR-Tools — an open-source combinatorial optimization library developed by Google's Operations Research team. OR-Tools provides the mathematical framework (a variant of the Capacitated VRP with multiple constraints), while our custom ML heuristics layer provides the warm-start solutions that allow the solver to find near-optimal routes for 300 vehicles and 8,000 stops in under 5 minutes rather than the 8–12 hours a brute-force solver would require. The optimization simultaneously considers: vehicle capacity constraints (by weight and volume), delivery time windows (customer-specified or route-sequence-derived), driver shift start times and maximum hours, vehicle type restrictions by road category (2-wheeler-only lanes, restricted roads for heavy trucks), and historical traffic speed profiles by road segment and time of day. The objective function minimizes total distance driven, subject to all constraints being satisfied — but with configurable weights that allow the operations manager to prioritize time window compliance over distance minimization if client penalties make it economical to drive farther to arrive on time.

02

Real-Time Traffic Integration

Routes integrate live Google Maps traffic data and historical delay patterns by road segment and time-of-day. If traffic builds on a route at 2 PM, the system re-calculates affected vehicles proactively. The traffic integration works at two levels. At the planning stage (morning, before dispatch), routes are built using predictive traffic models based on historical speed profiles — capturing the predictable rush-hour slowdowns that any experienced planner knows about but has trouble quantifying precisely across 6 cities simultaneously. During execution (throughout the day), real-time traffic data from the Google Maps Distance Matrix API is checked every 15 minutes for all active vehicles. If a vehicle's remaining route is predicted to cause late deliveries due to unexpected congestion, the system generates a re-routing recommendation for the dispatcher — who can approve it with one click, automatically updating the driver's app navigation. The real-time re-routing has been particularly valuable for predictable but hard-to-schedule events: market days in specific localities, school opening/closing hours near schools, and weather events like rain that predictably add 30–60 minutes to certain routes.

03

Driver App with Turn-by-Turn

Drivers receive optimized routes on a React Native mobile app with turn-by-turn navigation, delivery sequencing, and digital POD (proof of delivery) capture. The app works offline — routes are downloaded to the device at dispatch time, so drivers are not dependent on continuous internet connectivity in areas with poor signal. Each delivery stop shows: customer name and address, delivery time window, items to deliver, special instructions (leave at gate, call before delivery, etc.), and a one-tap call button for the customer. POD capture supports: QR code scan (for e-commerce parcels), customer signature on screen, photo proof (for unattended deliveries), and OTP confirmation (for high-value shipments). All POD data syncs to the central system in real time when connectivity is available and batches when it is not. The app also captures driver feedback: a simple "route quality" rating after each delivery, and a tap-to-report for route issues (wrong address, inaccessible location, address doesn't exist). This feedback feeds directly into the routing model's training data, systematically improving future routes based on ground-level intelligence from the drivers who know the roads best.

04

Dynamic Disruption Management

If a vehicle breaks down or driver doesn't show, the AI instantly re-assigns their deliveries to nearby vehicles with available capacity — replanning takes 90 seconds, not 45 minutes. Disruption management is the feature that operations managers most frequently cite as life-changing. The morning scramble when a driver calls in sick at 6 AM — which previously required a dispatcher to manually figure out which routes could absorb the additional stops, who had capacity, and how to restructure the day — now takes a single click. The system identifies all nearby vehicles with remaining capacity, checks time window feasibility for each orphaned delivery, and presents the dispatcher with a recommended reallocation plan that optimally distributes the disruption across the available fleet. For vehicle breakdowns during the day, the system generates a rescue plan: which nearby vehicle should divert to collect the stranded driver's remaining deliveries, updated route plans for both vehicles, and customer notifications for any deliveries that will be delayed as a result.

Six Months from TMS Integration to Full AI Dispatch

A 300-vehicle fleet cannot be switched to AI dispatch overnight. The rollout was phased to manage operational risk, build driver confidence, and validate AI performance before full dependency.

M1

TMS Integration & Vehicle Master Data Setup

Month 1 focused on data foundation. The client's TMS (a custom-built system) was integrated via REST API to provide daily delivery order data to the routing engine. Vehicle master data — capacity (by weight and volume), fuel type, vehicle type, GPS tracker ID, home depot assignment, and driver assignment — was cleaned, standardized, and loaded into the routing system's vehicle database. This data cleaning exercise revealed 23 vehicles with incorrect capacity data in the TMS (significantly affecting optimization quality if uncorrected) and 8 vehicles with non-functional GPS trackers (which were repaired before the live system depended on their location data). Driver shift data was integrated from the HR system to ensure the optimizer respected legal driving hour limits and shift start/end constraints.

M2

Google Maps API Integration & OR-Tools Deployment

The routing engine was deployed on cloud infrastructure with auto-scaling to handle the daily planning batch (which runs at 5 AM, before the first dispatch at 6 AM). Google Maps Distance Matrix API was integrated for both planning-time travel time estimation and real-time traffic monitoring during the day. OR-Tools was calibrated and tuned for the specific characteristics of this fleet and delivery network — the solver time limit, solution quality parameters, and constraint priorities were optimized through 2 weeks of test runs against historical delivery data to ensure the system consistently produced routes that outperformed the human planners' historical results. The real-time traffic monitoring subsystem was deployed separately and tested in parallel with live routes for 3 weeks before being trusted for re-routing recommendations.

M3

Driver App Development (React Native)

The driver app was built over 6 weeks in close collaboration with a pilot group of 12 drivers across 3 cities. Weekly co-design sessions with drivers identified critical UX requirements that would not have been obvious from an office perspective: large text for glare-resistant readability, single-thumb operation for use while standing by the van, voice readout of the next delivery address, and an offline mode that worked even in basements and industrial estate areas with poor signal. Hindi and regional language support (Marathi, Tamil, Kannada) were built in for driver accessibility. The app was beta-tested by the 12 pilot drivers for 3 weeks before being released to all 300 drivers — their feedback shaped numerous small improvements that significantly improved driver adoption.

M4

UAT & Driver Training (300 Drivers)

User acceptance testing ran for 3 weeks across all operational scenarios: normal day planning, driver absence, vehicle breakdown, time window violation scenarios, and mixed-fleet multi-depot scenarios. All edge cases were documented and resolved before the live system was trusted. Driver training was conducted in city-specific groups of 20–30 drivers, in Hindi and local languages, with emphasis on the benefits to drivers (better routes, less unnecessary driving, less frustration) rather than framing the technology as a monitoring tool. Driver champion volunteers from each city zone were identified and trained first; they became peer advocates who helped colleagues with app questions during the rollout. All 300 drivers were trained and had completed at least 5 practice routes on the app before the parallel running phase began.

M5

Parallel Run — Manual + AI Routes Compared Daily

For 4 weeks, the AI routing system planned all 300 routes every morning, but the dispatch team also manually planned routes as usual. Both sets of routes were compared against actual outcomes daily: total kilometers driven, on-time delivery rate, vehicle utilization, and fuel consumption. In week 1, AI routes were 12% shorter on average than manual routes. By week 4, as the model incorporated 4 weeks of real-world speed data from the drivers' GPS, AI routes were 19% shorter. On-time delivery performance with AI routes (measured on the 50% of vehicles that trialed AI routes in weeks 3–4) was 93% vs 78% for manual routes — conclusive validation before full deployment.

Logistics KPIs All Improved

20% Fuel Reduction

Eliminated unnecessary backtracking and cross-route overlap. Average daily km per vehicle dropped from 187 km to 149 km — saving ₹68 per vehicle per day in fuel alone. Across 300 vehicles operating 300 days per year, this is ₹2.4 Cr in annual fuel savings from distance reduction alone, not counting the secondary fuel saving from reduced idle time at congestion points (previously a major fuel drain when drivers were stuck on routes that predictably hit traffic). Vehicle maintenance intervals also extended — fewer kilometers means fewer service cycles — generating an additional ₹18 lakh per year in maintenance savings not included in the headline fuel saving number.

🕐

On-Time Up to 96%

Client penalty clauses triggered by late deliveries dropped to zero in month 3 — a direct result of on-time delivery improving from 78% to 96%. Customer satisfaction scores improved across all 6 cities. Two FMCG clients who had been preparing to reduce contract allocation due to OTIF performance proactively renewed contracts at increased volumes after seeing the month-3 performance data. The improved on-time performance also changed the commercial dynamic: the company moved from a defensive posture in client reviews (explaining late deliveries) to an offensive one (demonstrating superior OTIF as a competitive advantage over other 3PLs). New business conversations now open with the OTIF data — which no competitor in the same city cluster can match.

📋

Digital, Zero-Paper

Electronic POD capture eliminated 480,000 paper delivery receipts annually — saving the cost of printing, storing, and retrieving paper PODs, which was approximately ₹28 lakh per year in stationery, storage, and manual data entry labor. Customers receive electronic POD (image + signature + timestamp) within minutes of delivery via email and WhatsApp — a service upgrade that clients in the FMCG sector particularly value for their accounts payable automation. The data captured in digital PODs also provides the company with a comprehensive delivery performance database for the first time: actual delivery timestamps, GPS coordinates confirmed at delivery, customer signature, and photo evidence — all searchable by date, route, driver, and customer.

👷

Dispatchers Now Strategic

The 6 planners who previously spent 3–4 hours every morning on manual route planning now spend that time on higher-value activities: exception management (investigating the 4% of late deliveries and identifying systemic causes), client relationship management, capacity planning for new client onboarding, and analysis of route performance trends. Two planners were redeployed to a new business development function focused on converting competitor clients to this 3PL using the OTIF performance data as the primary sales tool. The operations manager described the transformation as "turning our best planners from calculators into strategists — the same people, now doing the work that actually grows the business."

The Financial Case for AI Route Optimization

Route optimization AI has one of the most straightforward ROI calculations in the AI landscape because fuel costs and delivery penalty values are precisely known. The calculation below uses only verified actuals from Year 1 of deployment, with conservative assumptions on the penalty and business development benefits.

Benefit CategoryCalculationAnnual Value
Fuel cost reduction38 km saved/vehicle/day × 300 vehicles × ₹7/km fuel × 300 days₹2.40 Cr
Penalty clause eliminationFMCG late delivery penalties avoided post-OTIF improvement₹45 L
Planner labor redeployment2 planners redeployed to BD function — revenue attributable (conservative)₹28 L
Paper POD eliminationPrinting, storage, data entry labor saved across 480,000 deliveries₹28 L
Vehicle maintenance savingReduced km → extended service intervals across 300-vehicle fleet₹18 L
Total Annual Benefit₹3.59 Cr
System Investment (Year 1)Development + API costs + driver app + training + 1-year support₹42 L
Net Year 1 Return8.5x ROI — payback in under 6 weeks of operation₹3.17 Cr net

Technologies Used

PythonGoogle OR-ToolsFastAPIGoogle Maps APIReact Native Driver AppPostgreSQLRedis (real-time state)WebSocketReact Operations Dashboard

About This Project

Does this work with our existing TMS (Transport Management System)? +
Yes. We built API integrations with the client's existing TMS for order import. The route optimizer runs as a service that TMS can call — or we integrate directly into the TMS via webhook. We've built integrations with Locus, FarEye, and custom systems. In most cases, the integration is straightforward: the TMS has a daily order export API, and our system consumes that data to plan routes. Where the TMS doesn't have a formal API, we can ingest CSV exports from the TMS directly. The route output is then pushed back to the TMS as a route assignment file that the TMS can display in its normal interface — meaning dispatchers can continue using the TMS they know while benefiting from AI-optimized routes underneath.
What if drivers deviate from the optimized route? +
The driver app tracks actual route vs planned route in real time. Deviations are flagged to the dispatcher with a reason prompt shown to the driver (who can tag: road closed, customer request, GPS error, driver discretion). Over time, we analyze common deviation patterns — often they indicate a bad road, wrong address, or genuinely better local knowledge. This feedback feeds directly into the route model: if 80% of drivers consistently deviate at a specific turn to take a parallel road, the model updates its speed profile for that road segment to reflect the real-world preference. Within 3 months of deployment, driver deviations had reduced 60% because the AI had incorporated the most common driver insights into its standard routing. The remaining deviations were legitimate real-world adaptations that the system captured and learned from.
Can it handle mixed vehicle types (2-wheelers, vans, trucks)? +
Yes. The optimizer treats each vehicle type as having different constraints: capacity (weight and volume by vehicle class), restricted roads (2-wheelers excluded from national highway on-ramps, heavy trucks excluded from city center zones during restricted hours), time restrictions, and cost per km (used to optimize assignment of deliveries to the most economical vehicle type when multiple types have sufficient capacity). Mixed fleets are handled natively — the system assigns deliveries to the most cost-effective vehicle type for each route segment. For this client's fleet of 200 vans and 100 motorcycles, the optimizer correctly assigned high-density apartment complex deliveries to motorcycle routes (where parking and building access made vans inefficient) while keeping bulk FMCG outlet deliveries on van routes where carrying capacity justified the higher vehicle cost.
How does the system handle last-minute order additions after routes are dispatched? +
Late orders are a daily reality in 3PL operations — FMCG clients call in express deliveries, e-commerce platforms push urgent shipments, and corporate clients have last-minute additions. The system handles this through a dynamic insertion algorithm: given a new delivery address and time window, it identifies which active routes can absorb the delivery with minimal additional distance and no time window violations, and recommends the optimal insertion point. The dispatcher sees the top 3 insertion options with projected additional distance and time impact for each. If no existing route can accommodate the delivery within the time window without violating constraints, the system flags this to the dispatcher with the earliest feasible delivery time — enabling an honest conversation with the client rather than a silent failure.
What data do you need to get started? +
The minimum data required to start building the system is: daily delivery order data (customer address, delivery time window if applicable, order weight/volume), vehicle master data (vehicle type, capacity, home depot), driver shift data, and 3–6 months of historical GPS tracking data from existing vehicles. The historical GPS data is used to build road speed profiles for your specific fleet and geographic operating area — capturing the actual speeds achievable by delivery vehicles (including stopping time at delivery points, parking search time, loading bay waiting time) rather than the theoretical speeds in map databases that don't reflect last-mile realities. If GPS tracking is not currently installed, we can begin with map-database speeds and refine over the first 2–3 months of live operation as real GPS data accumulates.
Can the system handle multi-depot operations where vehicles originate from different city warehouses? +
Yes — multi-depot VRPTW is one of the standard problem configurations we solve. The system knows which vehicles originate from which depot and ensures all routes start from the correct depot and return to it (or a designated end-of-shift parking location). For this client, 6 depot locations across 6 cities were all handled in a single planning run. Cross-depot delivery assignment is also possible: if a delivery in City A is very close to the boundary of City B's delivery zone, the system can evaluate whether it's cheaper to serve it from the City B depot rather than the City A depot — factoring in the city-specific vehicle availability and depot costs. This inter-depot optimization generated an additional 3% distance saving on top of the within-depot route optimization.

Services Used in This Project

Cut Logistics Costs with AI Route Optimization

Get a free analysis of your current delivery routes. We'll show you exactly how much fuel and time AI optimization would save you.

Get Free Route Optimization Analysis