Over the past five years, user expectations in e-commerce delivery have evolved dramatically. It’s no longer just about speed — transparency, optionality, and control have become essential: when exactly will the order arrive, how much will it cost, and can I adjust the delivery options? The once-standard 3–5 day delivery window is still common, but increasingly outdated — a shift largely driven by the integration of machine learning technologies into logistics.
Same-day delivery is growing at a steady pace: according to Mordor Intelligence, the global market size for same-day delivery reached approximately $28.7 billion in 2024, with a projected increase to $32.44 billion in 2025 and $54.8 billion by 2030 (CAGR ~11.45%). But scale brings challenges: peak demand, limited courier supply, high numbers of empty miles, and operations that often run at low single-digit margins, depending on segment and geography.
“Delivery by lunchtime” is rapidly becoming the new baseline for marketplaces aiming to define customer standards for years to come. Yet every extra minute of courier time still costs the system more than it should. Major players like Amazon, Uber (Uber Eats and Uber Direct) address supply-demand imbalances through closed, proprietary pricing algorithms. As a result, the final delivery price — both for customers and retailers — is formed behind the scenes, beyond the comprehension of most market participants.
From an ML perspective, this creates a troubling reality: effective solutions do exist — but they’re kept private. Developers across the industry are solving the same challenges in parallel, with limited access to proven, production-grade approaches. This raises the barrier for regional and specialized platforms trying to compete.
Andrei Nartsev, a researcher, guest lecturer at the Higher School of Economics, and Head of Applied Machine Learning at Yandex, is one of the few experts who has openly published the logic and metrics behind his delivery optimization system. His model — built on batch delivery and flexible delivery windows — boosts courier efficiency and transparently converts time savings into price savings for the end user.
Following the release of his technical paper, we spoke with Andrei about how machine learning is shaping the future of delivery — more adaptive, precise, and scalable — and why such tools matter not just for tech giants, but for the health and openness of the entire logistics ecosystem.
1. What emerging patterns in user behavior are shaping the future of delivery services in marketplaces?
Same-day delivery was once a luxury. Today, it’s the default expectation. Thanks to advances in machine learning and logistics infrastructure, consumer behavior has quietly but fundamentally shifted: waiting is no longer acceptable. Deliveries are expected to be fast, seamless, and ideally, invisible.
What has changed is not just speed — it’s the scope. Users now rely on delivery platforms for everything: documents, flowers, furniture, keys, laptops, even jewelry. They no longer think about the type of item — they simply expect it to arrive quickly and at a fair price.
However, the continuous growth in demand puts increasing pressure on the system. When it exceeds supply — for example, during lunchtime peaks — delivery costs can rise significantly. At the same time, many users are willing to wait an additional 30–60 minutes for a noticeably lower price, and this is exactly why flexible delivery windows are rapidly gaining popularity. This is not a compromise, but a conscious, value-driven choice.
2. What do you see as the key bottleneck in scaling same-day delivery for marketplaces today?
The core challenge lies in the complexity that grows with scale. Modern logistics platforms handle narrow use cases well, like food delivery. It’s a relatively straightforward task: short routes, predictable demand, and uniform order types.
But as soon as a marketplace expands beyond a single category — offering delivery for groceries, furniture, documents, or electronics — it faces a surge in demand and supply heterogeneity that’s extremely difficult to coordinate in real time.
The issue isn’t courier speed or mapping accuracy. It’s about assigning thousands of highly diverse orders — with different urgency levels, physical constraints, and client priorities — to a limited pool of couriers in every city, at every moment.
One request needs to arrive ASAP. Another can wait two hours. A third comes from a large B2B client with its own SLAs. Simple heuristics don’t cut it.
What’s required is a system of algorithms that not only dispatches couriers but also anticipates downstream effects, manages trade-offs, and dynamically balances the interests of users, couriers, and the platform itself.
This is where machine learning stops being a support tool and becomes the core operating system of same-day delivery.
3. How is machine learning helping solve this bottleneck, and what’s unique about your approach?
Machine learning enables marketplaces to move beyond reactive logistics and operate with predictive control. In this context, it powers a core delivery optimization mechanism: forecasting batching efficiency — the benefit of grouping multiple orders into a single courier trip.
A metric called SH economy (Supply-Hour Economy) was introduced — it measures how much courier time is saved when orders are delivered in a batch rather than individually. For example, if two separate trips take 33 minutes and the batched route takes 26 minutes, the SH economy is 21%. This value is predicted by a machine learning model in real time.
What happens next is key: that predicted efficiency is translated into a transparent discount for the customer. The model doesn’t just optimize operations behind the scenes — it directly shapes the delivery offer shown at checkout. This is where ML becomes a driver of both user experience and pricing logic.
What makes the approach unique comes down to three principles:
- Category-agnostic design — the system supports food, electronics, documents, furniture and more, all under one architecture.
- Continuous adaptability — the model retrains on a rolling basis to reflect seasonality, behavioral shifts and regional patterns.
- Production-grade deployment — the system is already live and powering pricing and routing decisions under real-world conditions.
While many systems are built for a single vertical, this one was designed from the ground up to support a mixed-demand dynamic marketplace. This approach enables greater flexibility, resilience, and scalability.
4. Why do batch delivery and flexible delivery windows play such a central role in your system?
Flexible time windows and batching unlock optimization opportunities that are simply not possible in rigid, one-order-at-a-time delivery models.
When every order must be dispatched immediately, the system loses the ability to coordinate — requests are routed one by one, often inefficiently. But if a user is willing to wait just a bit longer — typically within 30 to 60 minutes, or at most a few hours — the platform gains a valuable window to identify a second nearby order and fulfill both within a single route. This dramatically reduces courier workload, without compromising speed or reliability.
Importantly, we are still solving the same-day delivery problem. This is not about postponing orders to the next day — it’s about introducing a small, intelligent buffer that enables better decisions in real time.
This is where batching becomes viable, and SH economy — the metric for courier time savings — becomes measurable. Data consistently shows that even slightly wider windows lead to better route efficiency and substantial time savings for couriers.
In fact, the deployment of our ML model built on these principles has already shown statistically significant results: delivery costs for batched orders decreased by 9% on average, while courier earnings remained unaffected. For customers, that translates into lower prices — and for the platform, into stronger user acquisition potential.
And that efficiency doesn’t stay within the system — it’s passed on to the user as a transparent, dynamic discount: the customer waits a little longer, pays less, and the platform operates more sustainably.
In this sense, flexible windows aren’t about slowing things down — they’re about making fast delivery smarter and more scalable. This mechanism creates tangible benefits for all participants in the logistics chain.
5. What limitations do current ML technologies still face in this domain?
There are three major technical challenges.
First, the target metric — SH economy — is highly volatile. It fluctuates with seasonality, geography, and user behavior, making it difficult to model with traditional assumptions about data stability.
Second, in calculating SH economy values, the model uses both actual data (how much time was spent delivering orders in a batch) and predictive estimates: how much time a courier would have spent if an order were delivered solo. That requires careful reconstruction of counterfactuals — estimations that are both unbiased and statistically sound.
Third, the model influences user behavior, which in turn feeds back into the target. As more people choose flexible delivery due to discounts, overall batching dynamics begin to shift. This is handled through regular retraining and keeping the model calibrated over time.
6. What’s next? Where do you see the future of dynamic pricing and delivery optimization?
Delivery systems are becoming increasingly personalized and adaptive. Today, they already respond to external signals like traffic, weather, and local events. But the next step is deeper integration of real-time user behavior — urgency, flexibility, habits, and preferences.
The future lies in systems that dynamically adjust not just pricing, but routing and delivery logic based on individual user context and live conditions.
And it won’t just be about smarter algorithms. AI models will orchestrate the marketplace, but the actual execution of delivery will increasingly shift to robotic systems. This will reduce costs, increase precision, and minimize dependency on human labor.
The long-term vision is a self-tuning marketplace where logistics becomes seamless, automated, and practically invisible to the end user. Smart models will govern the flow, while physical execution becomes more autonomous. This is the future of logistics.
