The hidden cost of poor order routing in Quick Commerce and how AI fixes the margin leak

India's quick commerce story is one of the most striking retail disruptions in recent memory. Gross order value grew from roughly $0.5 billion in FY22 to over $7.4 billion by FY25 a near 15x jump in three years. Bain & Company's How India Shops Online 2025 report found that quick commerce now accounts for over two-thirds of all online grocery orders in the country. The sector is expected to grow another 40% annually through 2030.

By all visible signs, this category seems to be doing really well. But spend five minutes with the people actually running these operations: the brand founders, the ops leads, the dark store managers and a different conversation emerges. Orders are coming in, revenue is climbing but the unit economics still don't behave the way the model said they would.

The gap isn't in the demand story, which is real. It's in the operational layer underneath specifically in how orders get routed from the moment a customer taps "place order" to the moment a rider knocks on their door. This invisible layer is where margin goes quietly missing, and it's where AI is beginning to make a genuinely measurable difference.

Why profitability in Quick Commerce is harder than it looks

Quick commerce sits in an uncomfortable cost structure. The 10-to-30-minute delivery promise requires a dense network of dark stores, a standing fleet of riders, real-time inventory visibility, and operational systems that don't break under demand spikes. Every order must absorb a fixed share of those costs, and the math only works if the variable cost of fulfilling each order stays below the margin that order generates.

The challenge is that last-mile delivery, the part that makes quick commerce ‘quick’ is also the most expensive part of the stack. According to research from Capgemini, last-mile delivery accounts for 28-53% of total fulfilment cost per order. In a category where average order values in India typically sit between ₹450 and ₹625, that's a significant number to manage.

What makes this especially hard is that the cost isn't fixed. It varies with every decision made at the routing layer: which store the order is assigned to, which rider picks it up, which route they take. Make those decisions well, consistently, at scale and the unit economics work. Make them poorly, or inconsistently, or well only during off-peak hours and the gap between model and reality widens with every order.

In short, most Quick Commerce operators don't have a demand problem, they have a routing problem.

The 5 places routing costs you money

1. The wrong store gets the order

The instinct in quick commerce is to assign every order to the nearest fulfillment node. But proximity is a proxy for speed, not a guarantee of it.

A store 600 metres from the customer that is currently processing fifteen other orders may take twelve minutes to begin packing. A store 1.1 kilometres away with two pending orders may begin packing in ninety seconds. Assigning to the nearest store in the first scenario doesn't just slow down this delivery, it creates a queue effect that slows down every delivery behind it.

Beyond queue depth, there's the inventory problem. If the assigned store doesn't actually have one of the items in stock, the order stalls or cancels. Cancellations are particularly damaging in this category: research consistently shows that roughly half of consumers will switch to a competitor after just one bad experience with a brand, and in quick commerce, a cancelled or significantly delayed order is precisely the kind of experience that triggers that switch.

2. Rider dispatch compounds at scale

Matching riders to orders sounds straightforward until there are a lot of orders happening at once. A good dispatcher can handle ten orders at a time and make sensible calls. But on a busy Saturday evening, a mid-sized operator might have a hundred orders open simultaneously. At that point, no human can evaluate all the possible rider-order combinations fast enough to get it right.

So what actually happens? The dispatcher works through orders one by one. Each individual decision looks reasonable. But because they're working sequentially, Rider A ends up with an order that Rider B, who was two streets closer, could have completed in half the time. By the time anyone checked Rider B's location, she'd already been assigned elsewhere.

Four minutes of extra time per order sounds trivial. But multiply that across hundreds of deliveries every day and you're burning hours of unnecessary rider time and the fuel that comes with it.

3. ETAs that slip, destroy repeat purchase rates

Quick commerce is a repeat-purchase business. The unit economics only work if customers order frequently. The TeamLease Q-commerce report found that average monthly orders per customer grew from 4.4 in 2021 to 6 by 2024, and platforms are betting on that number continuing to climb.

ETA accuracy is one of the most important variables in whether a customer orders again. When a platform promises 20 minutes and delivers in 38, it doesn't just create a bad experience for that one order, it recalibrates the customer's expectations downward and erodes the platform's core value proposition. A 5% improvement in customer retention has been shown to boost profits by 25-95%. In quick commerce, where customer acquisition is expensive and the margin per order is thin, losing a repeat customer to a single bad ETA is far more costly than it appears on any individual order's P&L.

4. Returns that don't close the loop

In fashion and pharmacy, two categories increasingly central to quick commerce growth returns are a structural part of the value proposition. Try-at-home delivery means a meaningful percentage of orders will come back.

When a return is processed slowly, or when the returned item isn't re-entered into the inventory system promptly, that stock becomes effectively invisible to the platform. It can't be ordered by the next customer. This "ghost inventory" stock the system believes is unavailable when it physically isn't is a quiet but persistent revenue drain. The faster a return is logged and inventory restored, the more orders that stock can fulfil.

5. Peak-hour collapse

Manual routing holds up well enough when order volumes are manageable, but it starts breaking down exactly when you need it most i.e. during a sale, a festival weekend, or a busy Saturday night when three times the usual orders come in at once. And when it breaks, the costs don't just add up, they compound. Deliveries fail, refunds go out, riders clock overtime, and someone has to scramble to re-assign orders on the fly. By the end of the night, all that chaos can quietly wipe out the margin from every order that actually went smoothly.

What makes this especially damaging is the timing. These demand spikes are the moments when the most customers are on your platform; people trying it for the first time, or coming back after a while away. A bad experience during a sale isn't just the cost of one failed delivery. It's the experience people remember, and more often than not, the reason they don't come back.

6. Takeaway?

All these problems are connected. When an order is sent to the wrong store, it creates delays and puts extra pressure on riders. As orders pile up, manual rider assignment becomes less efficient, making deliveries slower and causing ETAs to slip. When promised delivery times are missed, customers lose trust and are less likely to order again. At the same time, slow return processing keeps good inventory unavailable for new orders, reducing sales. During busy periods like sales or weekends, these small inefficiencies build on each other, leading to delays, cancellations, higher costs, and unhappy customers ultimately hurting both profits and customer loyalty.

What AI routing actually does differently

The term "AI" gets applied loosely in logistics technology, so it's worth being precise.

A rule-based system follows a fixed decision tree: assign to the nearest store, pick the fastest road, choose the rider with the lowest active task count. These rules function reasonably when conditions are stable. In quick commerce, conditions are never stable. Inventory levels, rider positions, traffic, order volumes, and customer locations shift continuously and simultaneously and the interactions between these variables are too complex for a static ruleset to handle well.

An AI routing engine processes all of these signals in real time and finds the combination of decisions such as, which store, which rider, which route produces the best outcome across all dimensions simultaneously. It doesn't pick the nearest store. It picks the store that, given current inventory, current backlog, current rider positions, and current traffic, will produce the fastest and most reliable delivery for this specific order, right now.

The practical results of this shift are measurable. Companies that deploy AI route optimisation report 15–25% reductions in transportation costs and 10–20% fuel savings compared to manual planning. One AI routing study in a fast-delivery context showed average delivery time dropping from 31.2 minutes to 25.4 minutes while on-time delivery improved from 78% to 92%. Customer retention in the same study rose from 74% to 89%.

The other advantage is learning. A rule-based system only improves when a human updates its rules. An AI routing engine notices patterns that orders from a certain pin code placed after 7 PM tend to require a return call, that a specific route becomes unreliable in evening rain and incorporates those patterns into future decisions automatically. Every order makes the system incrementally more accurate.

The 3 routing decisions that define your unit economics

Not every routing decision affects your bottom line equally. Three decisions, made on every single order, account for most of the difference between an operation that's profitable and one that isn't.

1. Which store fulfils the order

The nearest store isn't always the right one. If it's backed up with orders or running low on stock, your delivery is already late before a rider has even been assigned. AI looks at queue depth, stock availability, and packing speed across all eligible stores and picks the one most likely to get the order out fastest, not just the one closest on a map.

2. Which rider picks it up

At scale, manually matching riders to orders leads to avoidable delays; a rider two minutes further away gets assigned simply because the better option wasn't checked in time. AI tracks every rider's live location and current workload and makes the optimal match across all open orders at once, cutting down delivery times and reducing fuel costs without any manual intervention.

3. Which route they take

Last-mile delivery eats up 28-53% of total fulfillment cost per order, which means even small route inefficiencies add up quickly across hundreds of daily deliveries. AI doesn't just plan a route at dispatch and leave it, it updates the route in real time based on live traffic, and can even combine two nearby orders into a single run when it makes sense to do so.

The metrics that tell you whether your routing is working

If you've upgraded to AI routing, four metrics will tell you whether it's working long before the impact shows up in customer satisfaction scores.

1. On-time delivery rate 

Good routing should keep the on-time delivery rate high even during peak hours, not just on quiet afternoons. If your on-time rate drops on Friday evenings or during a sale, your routing is struggling exactly when it matters most. Fynd Quick maintains a 90% on-time delivery rate across its platform.

2. Order-to-dispatch time 

Order-to-dispatch time measures how long it takes from a confirmed order to a rider leaving the store. This number should be short and consistent. If it keeps varying, you likely have a store assignment problem, an inventory accuracy problem, or both.

3. Cancellation rate 

The cancellation rate is worth watching closely, because a surprising number of cancellations trace back to routing; the wrong store was assigned, the item wasn't actually in stock, or the ETA slipped so much that the customer gave up. Fixing your routing is often the fastest way to bring this number down.

4. Return processing time 

The return processing time matters most in fashion and pharmacy. Every hour a returned item sits unprocessed is an hour it can't be sold to someone else. When your systems are connected and returns are routed intelligently, this number stays low and your inventory stays active.

Why getting the infrastructure right matters

Quick commerce in India is still early. The brands that get their operations right now before the market tightens and margins compress further will be the ones still standing when it does. AI routing used to be something only the Blinkits and Zeptos of the world could afford to build. That's no longer true. The technology is accessible, the ROI is proven, and the window to build an edge is still open. The only question worth asking is whether your operation is capturing these gains or quietly bleeding margin on every order you fulfill.