Algorithmic Transparency: Courier Partners

At Wolt, we partner with thousands of self-employed courier partners who can decide when, where, and how they want to perform delivery services. The following post explains the technology that the courier partners interact with and how the delivery journey works and can be found in the Wolt Algorithmic Transparency Report 2024. 

Courier partners play a crucial role in connecting consumers with local merchants by making the deliveries. With over 60,000 orders occurring simultaneously at times, ensuring every order is delivered is no small feat. 

That's where our task algorithm comes into play – it's designed to offer delivery tasks to the most suitable courier partner available based on four main parameters:

• Availability: The algorithm needs to know that a courier partner is online and has marked themselves as ready to accept deliveries or if the courier partner is not available due to being offline or busy making another delivery.
• Location: The algorithm needs to know the location of the courier partner to understand their proximity to the pick-up location. It receives the courier partner’s location data when they are online in the Wolt Courier Partner App and not when they are offline.
• Delivery vehicle: The algorithm needs to know the delivery vehicle the courier partner is using to understand their capacity for larger orders and the vehicle's speed. We currently factor in three different vehicle types—bicycle, motorcycle, and car—and the data is based on the courier partner’s own reporting. In 2024, we are planning on adding more vehicle types, such as e-bikes, to accommodate the variety of capacities and speeds these vehicles offer, ultimately improving the accuracy of the algorithm.
• Special capabilities: Some courier partners have capabilities to deliver special kinds of orders based on training or interest on their behalf. For example, only some courier partners are able to deliver pharmacy orders or handle cash in markets where this is available.

Delivery offers are only made to one courier partner at a time, and courier partners are offered the choice to either accept or reject the task at hand. Deliveries may also be offered as bundles in case there are multiple orders to pick up at the same venue or multiple venues in the same area and then subsequently drop off to different consumers who may live generally in the same area.

To help inform courier partners about the offer, every offer comes with information about: 

• The name of the venue

• Pick-up location

• Drop-off location

• Estimated distance the courier partner needs to make to complete the delivery

• Expected amount to be earned (the total may be higher, for example, due to customer tips, but it will never be lower than what is shown in the delivery offer).

If the courier partner does not respond to the delivery offer, it is automatically rejected within 30-60 seconds (range depending on market). Based on the same parameters above, the task algorithm then calculates the next suitable courier partner. There are no penalties for rejecting delivery offers, and this does not impact the courier partner’s future delivery offerings.

At Wolt, we do not use any type of performance monitoring or rating to factor into the task algorithm. In fact, courier partners are anonymized in connection to the automated delivery offering so that no type of subjective criteria, like personal characteristics or traits, could be factored in the algorithm. 

When you open the app, you might wonder how we predict the delivery duration displayed for each venue. For consumers, this pre-checkout delivery estimate may play an important role in choosing where to order from. 

The pre-checkout delivery time estimate is computed using a mix of data sources, which include:

• Delivery distance between venue and consumer
• The ratio of ongoing deliveries and active courier partners
• Average delivery time in the city in the past 30 mins
• Average venue order fulfillment time in the city in the past 30 mins
• Average venue order fulfillment time in the city based on different aggregations of historical data
• Pickup estimate in the city based on the past month’s pickups
• Current weather conditions in the city

You've placed your order, eagerly awaiting its arrival, while the delivery estimate countdown clock ticks away in the app. But have you ever wondered how we calculate these delivery estimates after you’ve placed your order? 

The order tracking estimates consider factors such as:

• Venue preparation time
• Courier partners’ availability
• Courier partners’ estimated travel time incl. type of vehicle, as well as the time needed to find the venue or your location.

One of the factors for calculating the delivery estimates is the estimated travel time for the courier partner to make the delivery. Ensuring the accuracy of these estimates is crucial, as it not only reduces the average delivery time but also lowers the costs associated with service quality by minimizing the likelihood of late deliveries. It’s important to note that courier partners are not required to complete the delivery within the estimated time frame. Deliveries made later than what was estimated are often due to circumstances outside the control of the courier partner, such as difficult traffic or weather conditions.  

In 2023, we introduced a machine learning (ML)-based travel time system to enhance the accuracy of travel time estimates for the delivery estimate. With our new ML-based model, we learn from past deliveries how the travel time varies over time. This more dynamic approach, updated hourly, better adapts to traffic fluctuations throughout the day, such as peak lunch and dinner times.

Similarly, we leverage ML-models to predict the time needed by courier partners to locate the venue or the delivery location once in a given geofence. These pickup and dropoff service times correspond to the time needed by courier partners to park and walk to the target location. 

The pickup and dropoff service times are estimated considering information, such as the average delivery execution time in the neighborhood, the order size or the vehicle type used by the courier partner. In 2023, accuracy improvements were achieved through using a modified algorithm (based on decision trees), training on larger time windows, and incorporating additional features like the courier partner's vehicle type. These enhancements ensure that our predictions reflect the real-world challenges faced by courier partners and contribute to a smoother and more efficient delivery experience for everyone involved.

Algorithms play an important role in enhancing safety in the digital space. With thousands of people who deliver on our platform on a daily basis, we take every action to prevent the misuse of our platform and protect it from being a place to treat anyone unfairly.

We already check every courier partner’s ID and working permits when they join the platform, but we also use facial verification technology to ensure that the person making deliveries is who they are supposed to be. In practice, this means that a courier partner can get a request to take a selfie on the Wolt Courier Partner App. The facial recognition technology then compares the selfie to the photo in the ID documentation on file, to ensure that the two images match. 

Automated facial verification improves partner privacy by standardizing biases across checks and minimizing data exposure within our systems, compared to manual verification. This helps mitigate potential biases and keeps data transfers to a minimum, supporting our principles of privacy and data protection.

Our initiative stems from a desire to responsibly manage platform access, safeguard customer privacy, and maintain the integrity of our courier partner community. We will continue to develop our technology, platform and operations to ensure that Wolt is a safe service for everyone.

At Wolt, we are committed to ensuring that courier partners engaging with our platform have access to attractive earnings. Courier partners are entitled to a delivery fee for each completed delivery they choose to accept. Based on the experiences gained from millions of deliveries since our very first in 2015, we understand that no two deliveries are alike. The same trip can be very different on different days due to elements out of courier partners’ control, such as traffic or weather conditions. Our ambition is therefore to make each task equally attractive for courier partners to accept. We want to ensure that each task is priced based on the effort needed to complete it.

The model for pricing each individual delivery is based on the following factors:

• Delivery distance: The actual estimated traveled distance (not based on straight line distance) by the courier partner from the pick-up location to the customer (and each successive consumer in case there are bundled deliveries).  
• Pick-up distance: The estimated distance traveled to reach the pick-up location (if applicable).

Moreover, to make the estimation even more accurate, the model may also take into account:

• Consumer and merchant location type: The locations might be difficult to reach (e.g. limited parking space).
• Weather: Difficult weather conditions may also impact the delivery. In case of severe weather, operations will be closed.
• Type of order: Bigger orders in terms of volume and weight, such as large grocery orders, usually impact the delivery and should therefore be reflected in the pricing.
• Other factors: Other conditions, such as difficult terrain or poor parking, higher need for courier partners to come online due to time of the day or other special events may also influence the estimation.

It is important to note that the delivery fee and service fee consumers pay Wolt is usually just a fraction of what Wolt pays courier partners for their services. In fact, there is no direct link between what consumers pay and what we pay courier partners for their delivery services, e.g. promotional fees or discounted delivery fees do not negatively impact earnings. In addition to the delivery fee consumers pay, Wolt funds courier partner earnings from the fees we charge merchant partners and other revenue sources.

The information on Wolt’s products and algorithms in this report are based on our operations as of February 2024 in Austria, Azerbaijan, Croatia, Cyprus, Czech Republic, Denmark, Estonia, Finland, Georgia, Greece, Hungary, Iceland, Israel, Japan, Kazakhstan, Latvia, Lithuania, Luxembourg, Malta, Norway, Poland, Serbia, Slovakia, Slovenia, Sweden. For more information, check out our Transparency web page.