Ceci est une version HTML d'une pièce jointe de la demande d'accès à l'information 'Documents related to 19.09.2022 meeting with Wolt'.



Ref. Ares(2022)7025147 - 11/10/2022
Algorithmic 
Transparency
Report
2022


Wolt’s Algorithmic Transparency Report
I.     About 
Wolt 
         
1
II. 
 
Why 
write 
this 
Report? 
      2
III.  Our principles in building our products 
 
 
 
2
IV.  Algorithms 
powering 
Wolt      3
 A: 
Courier Partners – Task allocation algorithm   
 
  Non-discrimination
  Human 
oversight
  Risk 
management
 
B: Customers – Ranking & Personalisation algorithms   
7
  Non-discrimination
  Human 
oversight
  Risk 
management
V.   Help us learn! 
        11
I.     About Wolt
Wolt is a Finnish technology company that operates a food, grocery, and retail 
intermediation and delivery marketplace. We currently operate in 23 countries and over 
250 cities, serving millions of customers together with our merchants and courier partners. 
Wolt was founded in 2014 in Helsinki where our headquarter lies today,  and we employ 
more than 6 000 people in our offices across our markets. 
By operating a technology platform, Wolt brings convenience to customers, but also 
creates economic opportunities for local restaurants and retailers selling their goods 
through our platform, and for independent couriers providing services to Wolt. Simply put, 
when buying something on our platform, our customers can choose to have the products 
delivered to them. If they do, we then offer that delivery task to a courier partner who is 
logged onto our platform.
1

II.   Why write this Report?
We are committed to making the cities we operate 
in better places to live. We are also committed to do 
the right thing towards our teammates, partners, 
customers, and the general public. This is the basis 
for everything we do, so we figured we might as well 
be transparent about what we do and how we do 
it. That way, others are able to evaluate if we are 
indeed doing the right thing.
III. Our principles in building our products
We take the concerns around algorithmic 
transparency seriously. We have advocated for the 
explainability of the key operating principles behind 
Fairness 
technology. Recently, the European Commission 
We are focusing on building a platform that ensures 
put forward a draft declaration on digital rights and 
fair treatment of the people that interact with it. 
principles for everyone in the EU1. According to the 
There shall be no direct or indirect discrimination 
declaration:
based on personal attributes, such as racial or 
ethnic origin, sex, religion or belief, disability, age or 
Everyone should be empowered to 
sexual orientation.
benefit from the advantages of artificial 
intelligence by making their own, informed 
Meaningful impact
choices in the digital environment, while 
We are building a product organisation that works 
being protected against risks and harm 
at scale with autonomy, all the while optimising for 
to one’s health, safety and fundamental 
meaningful impact. Impact is the biggest driver 
rights.
of motivation for the product team, be it from 
business metrics, societal, environmental influence 
This makes a lot of sense. The same goes for the  
or personal growth journeys. At the core of this 
commitments outlined in the declaration. They are 
approach is a focus on the value we create for our 
in line with our thinking. For example:
customers, partners and society at large.
We commit to ensuring transparency 
about the use of algorithms and 
Transparency
artificial intelligence, and that people 
We acknowledge that this industry is new and 
are empowered and informed when 
intricate. By providing accessible information about 
interacting with them. 
how it works, like we do in this Report, we hope that 
it will build trust with our partners and customers, 
Finally, we wanted to write this Report so that we 
as well as the general public.
can learn. This is our first one, and we are sure there 
is a lot we can improve. We want to get feedback. 
We want to work with others.
Technology is not an end in itself. For us, it is a means 
to make the cities we operate in better places to 
live. We think technology can help in achieving 
this goal and hope that through transparency, the 
positive role of algorithms and platforms can be 
understood.
1     https:/ ec.europa.eu/commission/presscorner/detail/en/IP_22_452
2


IV.   Algorithms powering Wolt
We include vehicle type as a factor as it impacts 
both the delivery speed and delivery capacity. As 
As a digital platform, algorithms are a vital part 
Wolt has expanded into enabling the deliveries of all 
of our work and facilitate millions of decisions 
sorts of things beyond restaurant food – including 
everyday. Yet, today, in the digital age, many people 
groceries, convenience items, and for example, 
associate algorithms with concern due to their non-
flowers – the physical size of an order can vary a lot.
transparency. In this section, we aim to dispel that 
concern by increasing the broader understanding of 
Before a courier partner can be offered tasks, they 
how algorithms are – or sometimes are not – used 
need to go through the onboarding process, where 
at Wolt. We hopefully provide enough context on 
Wolt checks e.g. the individual’s right to work, and 
what happens as part of a delivery order to give you 
to download the Wolt Courier Partner app to their 
a good understanding of the whole process.
device (smartphone/tablet).
We have drawn inspiration from the UK 
A delivery task offer includes information on the 
Government’s Algorithmic transparency template 
pick-up location (i.e. restaurant or store) and drop-
published by the Central Digital and Data Office2, 
off location (i.e. customer’s home, office, or any 
and also the City of Amsterdam’s Algorithm register 
other location) and is always offered to only one 
for explaining the algorithms they have in use3. The 
courier partner at a time, so there is no competition 
following sections are based on the situation at 
on who accepts a task first or fastest. 
Wolt in February 2022 and how we partner with self-
employed couriers, i.e. courier partners, as they 
Once a task has been offered to a courier partner, 
form the majority of couriers on our platform.
they can decide either to accept or reject the task. 
If a task is rejected it will automatically be offered 
to the next potential courier partner based on 
A: Courier Partners - 
proximity, vehicle type and availability. The same 
applies with the automatic rejection of a task which 
 Task allocation algorithm
happens after 60 seconds of no response from 
the courier partner. If no one accepts a task, Wolt 
Task allocation algorithm
cancels the customer’s order without any negative 
One of the most basic functions of a delivery 
impact or consequence for the courier partner.
platform is to offer delivery tasks to couriers so that 
they can either accept or reject offered tasks, and in 
case of accepting, transport an order from point A 
to point B. The task allocation algorithm aims to find 
a courier partner who can deliver the order without 
it being terribly late. Therefore, the algorithm in 
general tries to minimise the kilometers to be driven 
and time for all deliveries with the constraint of 
ensuring  all orders are delivered within a target 
time. The more efficient the algorithm is, especially 
with regards to grouping orders, the more courier 
partners can also earn.
Wolt’s algorithm offers delivery requests, ‘tasks’, 
equally to courier partners based only on their 
proximity to the pick-up location (i.e. restaurant 
or merchant), the vehicle type, and their current 
activity status (available for deliveries, busy with 
ongoing deliveries or offline). The platform does not 
use any kind of ranking or rating to determine which 
courier partner is offered a delivery task, only the 
Courier partners are not obliged to accept offered 
location and the type of vehicle of courier partners 
delivery tasks and they can even change their 
who have chosen to log online are used. 
mind after accepting a task, i.e. they can ask to be 
unassigned from a task they originally accepted. 
2   https://www.gov.uk/government/collections/algorithmic-transparency-standard
3   https:/ algoritmeregister.amsterdam.nl/en/ai-register/
3


All courier partners are anonymised in connection 
proposed tasks at any given time and request to be 
to the automated task allocation and a courier 
unassigned from a task after accepting it.
partner’s identity or e.g. possible task rejections, 
does not affect how the delivery tasks are offered. 
Delivery route
There will be no adverse consequences for the 
Once a courier partner has accepted a task, they are 
courier partners for rejecting tasks or for asking to 
free to choose the delivery route. Wolt does not pro-
be unassigned from a task.  
vide any driving instructions for how to travel neither 
to the pick-up location nor the drop-off location. 
Delivery modes
Many partners use an external map service such as 
If a courier partner has marked themself available 
Google Maps, Waze or Apple Maps for this. 
to be offered delivery tasks in the Courier Partner 
app, they will be offered tasks within their proximity 
When the courier partner has picked up the order at 
and fit for their vehicle type. Courier partners 
the merchant they are asked to confirm in the app 
can choose between two different modes to 
that the correct order has been collected. Wolt then 
perform deliveries, they can either accept tasks 
provides the customer with a delivery time estimate 
in a so-called ‘Single task mode’, delivering one 
automatically based on a number of factors. These 
order at a time, or ’Bundled tasks mode’, delivering 
include historical data on the duration of deliveries, 
combined or multiple tasks at the same time.  
the delivery distance, traffic, time of the day, and ve-
hicle type within a specific area. The estimate is not 
In Bundled tasks mode, the app may offer the 
binding and does not oblige the courier partner to 
courier partner additional tasks before an earlier 
deliver the order within that time frame either. There 
task has been delivered to the customer, if there 
will also not be any punishment or any other adver-
would be additional tasks available for pick-up 
se impact on the courier partner for not meeting the 
and drop-off in the same general direction as 
estimated delivery time.
the first task. The benefit of bundled tasks is 
that it is more efficient overall, and for courier 
The customer can follow the delivery in real-time on 
partners they enable higher earnings as they are 
a map included in the Wolt app, so that the custo-
paid for each task as if they were individual tasks.
mer is ready at the drop-off destination when the 
delivery arrives. When the courier partner has deli-
In both Single task mode and Bundled tasks mode 
vered the order to the customer, the partner marks 
courier partners can always choose to reject 
the delivery as completed in the app.
4



Earnings
Courier partners are paid per each completed de-
livery task. The fee per task consists of two compo-
nents, the fixed ‘Task Fee’ and the flexible ‘Distance 
Fee’. The Task Fee includes the distance from the 
courier partner’s location to the pick-up point of 
the order and the minimum drop-off distance tow-
ards the customer. The Distance Fee is paid out in 
addition to the Task Fee for long drop-off distan-
ces. The longer the additional drop-off distance, 
the higher will the Distance Fee be that the courier 
partner receives.
Courier partners with a car or van can receive large 
orders, i.e. orders that contain many items and are 
ordered by groups. As large orders may take longer 
to deliver, courier partners earn extra from those 
orders. 
In case of bundled tasks orders, the courier part-
ner earns a fee per each task delivered, even if they 
deliver multiple tasks from the same merchant to 
customers in the same building. 
Courier partners can at all times view their accumu-
lated earnings in the Courier Partner app. 
Non-discrimination
As already described above, all courier partners are 
anonymised in connection to the automated task 
comes up during the delivery. Common questions 
allocation and a courier partner’s identity does not 
that the Wolt Support gets include the accuracy of 
affect how the deliveries are offered. A delivery task 
the delivery estimate and missing or faulty items of 
is always offered to only one courier partner at a 
the order. Drop-off is another part of the delivery 
time, so there is no competition on who accepts a 
experience where human intervention is someti-
task first or fastest.
mes necessary – be it due to missing address details 
or other difficulties in navigating the last 50 meters.
We also review the balancing on vehicle types regu-
larly to ensure the courier’s choice of vehicle only 
Risk management
matters when it makes a crucial difference for the 
As location data is the most important criteria to the 
delivery based on the delivery distance or the size 
algorithm that offers tasks to courier partners, this 
of the ordered product.
can lead to variance between couriers using diffe-
rent devices with different geolocation accuracy. 
Human oversight
It is in every party’s interest that the system is as 
The algorithm is being continuously maintained 
accurate as possible as this influences the accura-
and updated to ensure it works as desired, whilst 
cy of the offered tasks. We also have technical and 
also making the algorithm faster in offering tasks to 
organisational measures in place to ensure proper 
couriers. We also look to maximise courier earnings 
level of privacy and security compliance with regard 
by developing the algorithm to offer more bundled 
to geolocation data, including regular review of the 
tasks when possible. This work is fueled by aggre-
necessity, granularity and retention of such data.
gated data from actual deliveries and operational 
considerations. 
Courier partners move in traffic and use smart-
phones. This includes inherent risks, such as 
Operationally, the Wolt Support can help courier 
looking at your device for navigation informati-
partners, merchants, and customers if anything 
on while being on the move. Thus, new tasks are 
 5




only offered when a courier is not already on a task. 
Online hours
This way, they can accept the next task, for examp-
Self-employed courier partners can connect them-
le, only after they have dropped off a delivery to a 
selves to Wolt’s Courier Partner app and offer their 
customer. This limits the situations where the app 
services whenever they want, without any obligati-
sends a notification that the courier partner might 
on to work. This allows them to have maximum fle-
want to react to. If the courier partner is in a situa-
xibility and log online or offline on the Wolt Courier 
tion where it is riskier for them to interact with their 
Partner app as they like. They can even choose to be 
device, they do not have to do anything; the task 
online without any intention to perform any delive-
is declined after the 60 second timer and courier 
ries, and can work for and with other companies, in-
partners are not penalised in any way for not ac-
cluding our competitors at the same time as for us. 
cepting or declining tasks.
Courier partners do not need to inform Wolt in ad-
Performance
vance if or when they intend to not perform delive-
Wolt does not use data such as rejection rates, 
ries for any period of time. In practice, as opposed 
amount of performed deliveries, speed of delive-
to an employee, courier partners choose at their 
ries etc. on courier partners. Rejecting a delivery 
discretion when to take a break or whether to remain 
request does not have any impact whatsoever on 
offline for a longer period without needing to inform 
courier partners’ possibility to get future delivery 
Wolt or receive Wolt’s approval. 
requests offered through Wolt’s platform, or their 
contractual partnership with us, nor does it result in 
Wolt does not apply any algorithms to monitor on-
any disciplinary measures.
line hours of the individual courier partners and let 
that impact the contractual partnership or task-al-
There are no ratings on courier partners. After a 
location. When we analyse online hours it is on an 
completed delivery, customers rate the delivery 
aggregated and anonymised basis to better under-
experience as a whole. Thus, when giving this rating, 
stand our operations. 
the customer signals their opinion of the various 
aspects that form the delivery, for example, the de-
A recent pan-European study by Copenhagen Eco-
livery time estimates generated by Wolt, the packa-
nomics found that a majority of couriers (72%) say 
ging of the goods and so on. The rating customers 
platform work is a complementary activity, with 34% 
give on the delivery experience is only a tool for our 
delivering while studying and another third (34%) 
Support Team to know whether there might be an 
saying they access platform work to top income 
issue with an order and if they should reach out to 
from other full or part-time work4. At Wolt, around 
the customer. 
40% of courier partners delivering via our platform 
work less than 7,5 hours per week on average.
4   https:/ copenhageneconomics.com/publication/study-of-the-value-of-flexi- 6
ble-work-for-local-delivery-couriers/


B: Customers – Ranking & Personalisation  
algorithms
The purpose of providing personalisation is to 
let customers discover different restaurants and 
stores and help them find what they need, be it 
chocolate chip cookies to feed a sudden crave, 
new fuses to get the lights back on, a surprise flower 
bouquet, or their weekly groceries. Ranking and 
personalisation on the Wolt platform happens in 
four main sections on the app and website:
 
1.    Discovery page
 
2.   Restaurants page
 
3.   Stores page 
 
4.   Search page
1. Discovery page
The Discovery page is the front page of Wolt and 
the first thing users see when they open the app or 
website. The ranking of content on the Discovery 
page uses a mix of manual and automated sorting. 
For example, ad hoc promotions and campaigns 
are manually added to the page by our marketing 
teams. 
We divide venues on the Discovery page into 
categories and carousels using automatic tools. 
Category Carousels: We tag venues into different 
Venue Carousels: Since we have a lot of different 
types of categories, such as ‘pizza’, ‘burgers’, ‘salad’ 
categories, some venues may show up in multiple 
etc. The order of these categories is automatically 
carousels. We want to show the customer venues 
determined by assessing the popularity of each 
that are relevant, therefore the venues on the 
category by aggregating previous orders. 
Discovery page are ranked using the First-Time 
User or Returning User Model explained below. A 
venue that has been ranked high will be given a more 
prominent spot.
7


First-Time User Model
Returning User Model
If the customer is an unregistered user or a registered 
The second tool we use in ranking content is 
user with no purchases, the sorting is done through 
the Returning User Model. This model is applied 
a simple static model. The model factors in three 
to registered users who have made at least one 
criteria. The first is the delivery estimate, i.e. the 
purchase with Wolt. It can be explained as a way of 
time it would take for the customer to receive the 
recommending the user something based on the 
order after completing their purchase. The second 
purchase history of other similar customers. The 
criteria look at the venue relevancy and popularity 
ranking looks at past purchases of an anonymised 
based on how many purchases other customers 
user without identifying the user personally and 
have made with the venue. This criteria also factors 
recommends venues based on purchase history. 
in venues that are new on the platform to make sure 
For example, if an anonymised Customer A 
they are also promoted initially. The third factor is 
purchases something from Venue 1 and Venue 
the venue price range that looks at the price of 
2, we look at other anonymised customers who 
menu items in the restaurants and tries to promote 
have ordered from the same venues. If the other 
a wide range of venues with different prices to 
customers have purchased something from Venue 
customers. 
3 we may recommend that venue to Customer A. 
The First-Time User Model works with aggregated 
Personalisation algorithms are trained every 
data. For example, two first-time users opening 
night. If you make a purchase during the day the 
the Wolt app in the same location will see the 
personalisation will be applied the following day. 
same venues ranked in the same way and no 
If you are a first-time user and make your first 
personalisation is applied. 
purchase, the next day the user experience may 
change since it is now based on the Returning User 
Model. 
8



2. Restaurants page
The Restaurants page features restaurant venues 
that are delivering to the customer and are sorted 
by their availability, meaning if they are open and 
accepting orders they will be displayed before 
closed venues. The available venues are then 
re-sorted using the First-Time User Model and 
Returning User Model.
3. Stores page
In our Stores page we display the selection of retail 
and grocery stores available to the customer. We 
apply a different kind of sorting algorithm for our 
Stores page as it works better for the customer 
experience. 
Stores Page Model
This model looks at three different features for 
ranking content on the Stores page. 
•  Popularity. This criteria also factors in venues that 
are new to the platform to make sure they are also 
surfaced to the customers. 
•  Customer rating
•  Likelihood of customer placing an order
The algorithm assigns each venue a rank according 
to all three criteria separately. Then it combines 
all rankings and re-sorts the venues based on the 
result. Thus we ensure that the venues that are 
displayed on the top of the page receive the highest 
ratings of all three criteria. 
9


4. Search page
If users look for a particular type of restaurant, 
store or item, they can use the search feature 
to find that. At Wolt, we have two main types of 
search functions, the Global search and the In-
venue search. For both search functions we use 
the customer’s typed query which includes any 
characters regardless of the language and factoring 
in synonyms. In the case of Global search, we also 
use the user’s delivery address, if the user has 
provided it to us. If the address is not specified, we 
use their current location if they have granted us 
their permission for this.
Global search
The Global search helps users search the platform 
for venues that match the users’ criteria and is 
within the delivery area of the customer. 
When users type their search in the search bar, 
the search request is sent to a back-end service, 
which factors in both the typed search text and 
the user’s location to build a search query. That 
search query is then matched against an existing 
index of metadata containing information from the 
merchants and their menu or product items. This 
produces a first list of results of venues that are 
corresponding to the search query. 
The search results are sorted by an algorithm 
determining the relevancy of the venue as well as if 
the venue is open or not. For example, venues that 
match the search query, but are not open, will be 
displayed at the bottom of the list. 
In-venue search
The second type of search feature is the In-venue search. When a user is looking at a particular restaurant or 
store they can search that venue for specific items. This search function works differently on the web and in the 
app in order to improve the user experience as some stores might include hundreds of items. 
Web
Mobile
When a user types a search query within a  specific 
When a user types in a search query within a specific 
venue’s page on the website, a filtering is applied 
venue’s page on the mobile app, a search request 
to hide the items that do not match the search 
is sent to a back-end service. The back-end tool 
query. The filtering is based on the menu item title, 
matches the search query against an index based 
description and category.
on the venue’s metadata, sorts it and returns the 
products as search results to the users. 
10

Non-discrimination
Risk management
Wolt’s ranking and personalisation algorithms 
Similarly to task allocation, location unsurprisingly 
do not process or use personal information that 
plays a role also on the customer side. Customers 
would have been identified as potential sources 
see venue options that are nearby based on their 
of discrimination against people impacted by 
approximate or city-level location. Customers also 
the automated process. In fact, we do not collect 
may choose to add a delivery address to the Wolt 
any data on gender, sexual orientation, racial or 
app in order to receive their order to the location 
ethnic origin, national origin, religious affiliation or 
of their choice. Inaccuracies in location data could 
disability. 
impact the selection of venues the customer sees, 
although this is more theoretical given we use less 
When training the algorithmic tools, we only use 
granular approximate data, and there are clear ways 
aggregated and anonymised data. 
the customer can ensure that their desired delivery 
location is used.
Human oversight
We mitigate and balance the algorithms to also 
The ranking and personalisation algorithms 
surface new restaurants and stores to customers. 
are in continuous development. As described, 
This balancing is phased out after a while.
specific sections in the app and website, such as 
promotions, are manually created.
V.    Help us learn!
Thank you for reading our first Report on Algorithmic Transparency. We made this Report partly 
in order to learn. We want to understand what was interesting, where we should have explained 
our algorithms better, and any other perspectives you think we should have covered. On our 
side, we continue to improve our platform and the algorithms that make it tick. If you have ideas 
that could help along the way, let’s chat!
You can get in touch with us via a new, dedicated point of contact for all things algorithms and 
transparency: xxxxxxxxxxxx@xxxx.xxx. The alias is managed by Wolt’s Public Policy team, who 
will liaise internally to gather the right people for questions, feedback and other input you might 
be interested in sharing with us.
11


http://www.explore.wolt.com/transparency