Customer Stories
ANWB Leverages MLOps to Achieve Personalization at Scale
The Dutch mobility company partners with Xebia to finetune personalized recommendations for its members.
The Client
The Royal Dutch Touring Club ANWB is a commercially operated association that provides roadside assistance, insurance, energy contracts, and many other services to its members. It aims to make mobility easier, remove hurdles, and provide support when needed.
ANWB is deploying a growing number of data products across its numerous services. Some of these projects leverage GenAI, but others, like the one in this story, are an example of “traditional” AI used efficiently and with end-users in mind.
At a Glance
Challenge
The pipeline to create recommendations for ANWB members was hard to operate because of its slow speed and need for manual intervention and maintenance.
Solution
ANWB partnered with Xebia to restructure, automate, and speed up its recommendations data pipeline, making it easier to further improve and maintain it.
Results
Before the project, it took multiple days to produce new recommendations, which meant it was only run once every two weeks. After the project, the pipeline automatically runs daily and only takes a few hours. This speed-up allows for quicker iteration and therefore also resulted in increased click and conversion rates.
The Challenge: Personalization at Scale
Let’s consider an ANWB member who visits the website. For the first time since she registered, she searches for campervan accessories. That means there is a high probability she has recently bought a campervan. As a result, ANWB would like to offer her a personalized experience that includes recommendations about campervan insurance and similar products.
ANWB follows the principle of showing the right message for the right customer at the right time. For this personalization project, the goal was to show tailored banners on the website and send tailor-made emails to members. To achieve this, the marketing team prepared numerous copy variations. The challenge for the data team was matching the right message to the right customers.
Before the collaboration between Xebia and ANWB, recommendations were generated only once every two weeks. The pipeline took around three days to complete and contained many manual steps. As a result, recommendations often arrived too late to be truly relevant.
The Solution: Structure, Speed, and Automation
While ANWB had a clear PoC (Proof of Concept) for personalized recommendations, it had still not reached PoV (Proof of Value). This created a challenge: from a data perspective, it was difficult to reach PoV without investment into improving the codebase; from a business perspective, it was a questionable decision to invest more money in a pipeline that wasn’t delivering PoV. What to do?
The ANWB team liaised with Xebia consultants and decided on the following steps to gain control on this project:
- Limit the scope of the project to make it easier to measure and determine the value of the created solution.
- Structure the codebase by improving documentation and creating modules with clear responsibilities (e.g. all data loading happens in one place). Doing this made it easier to further improve and maintain the code by the ANWB team.
- Speed up the process to create fresh recommendations. By improving existing queries and leveraging use-case specific properties to make each step (pre-processing, model training and post-processing) more compute and time efficient.
- More automation using AWS Sagemaker pipelines. By modularizing the code and using Sagemaker to orchestrate and automate as much as possible, the recommendations can now be refreshed daily and on demand.
- Improve the MLOps way of working. By working together with ANWB’s internal MLOps team, leveraging their bespoke framework and also contributing back to that framework based on the learnings from this project.
This combination of structure, automation, and operational excellence created a recommendation engine that was easier to maintain, improve, and scale
"Xebia employees add value from the start using their technical knowledge, dedication, and collaborative mindset. They make sure that their knowledge and deliverables are integrated within the company"
Rogier Emmen
Data Scientist at ANWB
The Results: Faster Recommendations and Better Business Outcomes
Before the beginning of this project, personalized recommendations were computed every two weeks, and it took three days to run them. Now, they run daily in a couple of hours and are ready for the team every morning. Not only that, but the click-through rate increased for many products, some of them by up to 20%.
As part of the knowledge transfer to ANWB, Xebia consultants organized many workshops to bring the data team up to speed. The sessions included theory and plenty of exercises, very much like our MLOps on AWS, GCP, and Azure trainings. This ensured ANWB employees were able to take over the project after Xebia consultants were gone.
Looking Ahead: Building on a Strong MLOps Foundation
The ANWB engagement demonstrates how MLOps can transform machine learning initiatives from promising concepts into measurable business value. By focusing on automation, maintainability, and operational excellence, ANWB created a foundation for continuous improvement in personalization.
With a scalable recommendation pipeline, faster iteration cycles, and stronger internal MLOps capabilities, ANWB is well positioned to further expand personalized experiences across its services while maintaining full ownership of its machine learning solutions.
If you’re wondering what else MLOps can do for your company, be sure to check out the MLOps Beyond the Hype whitepaper and contact us.
Contact