ICA – How to Personalise Offers and Recipe Recommendations?
Other teams at ICA look to the RecEngine team as a good example of how to work together in a data-driven way.
Signe Medelius, Product Owner of the RecEngine team.
Today, RecEngine is the most extensive machine learning application at ICA Sweden, and the development is still ongoing — by a collaborative team of ICA and Avaus.
Martin Lennartsson, ICA Sweden
In late 2017, ICA realised the potential in using artificial intelligence (AI) to grow their market share, by better personalising marketing and online search based on all of their data (with regards to GDPR regulations). As a result, the Recommendation Engine (RecEngine) project was started to ensure that ICA could use the transactional data they had collected. With the help of AI, the team would personalise campaigns and specific offers — to give the customer segment unique offers and recipes for each household.
With millions of website visitors monthly, ICA required assistance in developing more precise analytical capabilities. As the Swedish retailer had access to both web and store interactions, there was potential to personalise the search, not only based on previous web interactions, but also on real-world behaviour.
Therefore, ICA was determined to advance its analytical capabilities to improve the customer experience and provide its clients with the best possible experience.
1. Transformation from Rule-based to Model-based
2. Education and Cross-competence Collaboration
= On the Fast Track to Results
Transformation from Rule-based to Model-based
There was no AI in any of the existing systems, since they were all rule-based. Avaus was brought on board to first migrate the existing rule-based systems to a new environment, and subsequently, replace the rules with AI to boost efficiency.
ICA chose Avaus over off-the-shelf solutions, because ICA wanted to tailor their solution more than what would have been possible with those solutions.
Eventually, a number of machine learning models were designed and implemented to predict user behaviour and preferences. They could be leveraged into different offer types to extend and improve the current offer type portfolio into an ensemble of models tailored for each customer.
Since 2018, over the course of a few years, the RecEngine’ team has provided undeniable business results. Embodying the Data x Algo x Action framework, several models were connected straight to the systems sending communication to customers.
Education and Cross-competence Collaboration is Key
A large part of Avaus’ mission has been to educate the organisation on how data science can be incorporated in marketing. Avaus has focused on ICA’s long-term viability and self-sustainability, by constantly teaching individuals who interact with the RecEngine team to promote data literacy across the organisation. On various occasions, the organisation’s marketing knowledge was incorporated into AI algorithms to improve their performance.
“It is beneficial to involve as many people from the organisation as possible in AI projects. Because employees are the ones that know the most about their organisation’s products and customers, including their knowledge into the models yields the greatest outcomes,” Eric Hörberg mentions.
“Using the models resulted in an increase in sales that was statistically proven,” says Eric Hörberg, Senior Data Scientist at Avaus.
“Having the team performing the tests and the team evaluating them be two separate teams is a healthy organisational strategy,” Eric mentions.
On the Fast Track to Results
ICA went from no AI models into a 10+ people team of Data Scientists and Data Engineers. “A leading star for how an Analytics Team should be run,” describes Martin Lennartsson, former Delivery Lead at ICA.
The outcome was predictive modelling and a recommendation engine, which allowed ICA to provide a more personalised user experience for its customers. ICA has established a significant lead in personalisation capabilities, and has shown that personalised marketing at the household level is a highly effective approach to improve sales, loyalty, and profit both in the short and long term:
- As a result of productionalising 10+ models, over the approximately 50 A/B tests conducted, sales increased a substantial amount. It increased so much that the effect of the model could be statistically proven.
- Loyalty is a KPI that has many definitions, but by the definition we used, we could see that loyalty also increase as a result of personalisation.
- Profit was also measured as a KPI, and it also increased by a substantial amount. This proves that the models did not only create “coupon hunters”, they increased customers’ overall baskets.
With several AI models, tweaked in even more A/B tests, their aggregate earnings were estimated to be a large increase in yearly sales across all channels, where the models are employed. A different, independent team at ICA calculated the later results and their statistical significance.
A grocery retailer in Sweden with around 1,300 stores and a market share of about 36%. ICA is the leading grocery retailer in the country and one of the Nordic region’s main players in Fast Moving Consumer Goods (FMCG).
Company Size: Enterprise (8,600 employees, ICA Sweden)
Project: Data Science for Personalised Campaigns and Offers & Recipe Recommendations