Main content

Drinking from the AI fountain at H&M - democratizing AI in retail

Stuart Lauchlan Profile picture for user slauchlan June 14, 2021
H&M has set out to become the leader in applying AI tech and techniques in the fashion retail sector.

H&M store

Just as our founder Erling Persson once said he wants to democratize fashion, we are really about democratizing AI in our organization.

Retailer H&M has had some troubled times in recent years, but has been a pioneer in the application of AI in the sector. That mission statement from Persson, cited by Errol Koolmeister, Head of AI Foundation at H&M, is one that has its own updated iteration today. H&M wants to reduce time-to-market for AI use cases, make AI tools available for all product teams, improve skills literacy in  AI and increase model output availability.

It’s a goal that started to take shape some time ago, recalls Koolmeister:

We started really around 2015, 2016 when it became clear that our physical retail model was under attack. Digitalization was happening and we knew that we needed to do something. The question of course in management's head was, 'What are we going to do?'. Our business model was very successful for approximately 70 years, but we started seeing a decline, where our stock levels went up, and we started seeing potentially a harm in scaling more.

At this point the firm began to ask itself what it was doing in the field of AI? Having taken this question out into the organization at large, the responses were not encouraging, admits Koolmeister:

We were, of course, very good in some aspects, like CRM, like buying, controlling and the traditional aspects of retailing. In the past, when we were buying too much, we could just open a new store, but we wanted to be more precise. We wanted to minimize waste and really harness the value of AI.


Around 2017, this led to the launch of the first use cases, selected with two parameters in mind  - high value and high feasibility:

We wanted to pick the low hanging fruit, the biggest fruits. First, we didn't have the capacity to run these type of use cases internally. When you're just starting out, you don't have the amount of people that's required to drive an AI use case at that scale, so we went with an external consultancy firm. We got approximately 100 resources in and started focusing on value. That was the key aspect all along. Let's not build the most advanced models; let's build the ones that generate some sort of uplift and bring the value back so we can keep funding these type of initiatives.

Those first initiatives were so successful so far that they were self-funding in the first year and led to AI and Advanced Analytics becoming the first new corporate function in H&M in over 10 years.  This in turn led, in 2018, to a decision to start building these types of capabilities internally and become self-sufficient with the type of resources and skills and capabilities needed to become an AI and data-driven company.

By the following year, ambitions were scaling up. Koolmeister explains:

We started in recruitment, we started taking over the technical capabilities, we started developing for scale. We wanted to do not just seven or eight use cases, we wanted to do hundreds of them. So, in 2019 we really built this up and in 2020, just before the pandemic hit, it became quite clear that we were able to recruit over 100 people, replacing the consultants.

By this point, the firm had become sufficient enough to build new use cases itself, so when the pandemic hit, the company was able to adapt. It launched a new organization, merging AI with the old IT organization, becoming a new entity where AI now is a full part of the entire organization and enabling domain, he says, with a decision taken that all core operations decisions should be amplified by AI by 2025. The scope of AI application was growing, says Koolmeister:

We did a lot of these use cases in the beginning. We did allocation, of course. Where should our clothes go. We did  quantification, demand and how much should we buy. We did personalization and early recommendation and used these techniques to talk with our customers to make sure that we provided relevant offerings. We did mark down, so having the pricing at the end of the cycle.

But H&M realised quite fast that it had resource constraints, he says:

Even we put in a lot of money, we still wouldn't have the resources, because there truly is the war for talent. In the use cases that we were building, we realised there was some basic duplication. They were moving fast and they were agile, but given that we did seven or eight use cases, all running independently and autonomously to provide value, there was a lot of things being invented over and over again. We also saw in the beginning that there were long development times. Even though time-to-value was relatively short compared to some other cases I've heard, we thought that 12 months was too long. We wanted to bring that down into weeks

We also realised there was a lack of engineering knowledge. Of course you could hire a lot of engineers into AI, but if the organization wasn't really ready to take it over and run with it, that was hard. And we also thought that our process of starting new cases was relatively inefficient. We wanted to become a machine to start doing things again and again and again…What we realized as well was  there was a relatively low data and AI literacy across the organization.


Having identified all of these problems. H&M also realized that its delivery model wasn't really scalable. This was where, around the time COVID struck, that H&M came up with Fountainhead. Koolmeister pitches the mission statement here as:

We want to make sure that the things that we build can be used by everyone. We want to make sure that it's not just high tech data scientists and development times take a really long time and we're doing the neural networks that nobody's done before - that's not where our differentiation lies. Our differentiation lies in that we are able to extract value rapidly from AI. We are democratizing it, so the data analyst and product owner knows what to do when they're getting the hands on these type of capabilities.

The purpose is to accelerate the democratization of AI. It's easy to talk about democratization of AI,  it's easy to talk about democratization of data, but how do we actually do it? The way we go to market in our organization is that we actually do it by proving the value. We never think of a three year long project that is going to provide value in the end. We talk about agile, we talk about every sprint has delivery. It's about reducing the time to market for the AI use cases that we are building. It's about making  tools accessible for all the product teams. We also want to contribute to a higher AI skill set in this tech  upscaling the entire organization. And in the end, we want to increase the model output availability  for all.

Not re-inventing the wheel is a critical driver here, he adds:

The Fountainhead is about encapsulating AI capabilities, making sure that we actually rest on the solid ground. What we did when we sat down was that we started talking about what are we doing in our use cases? Everything we do rests on enablement is the key for delivering on these values that we're looking at. So it's around having proper Master Data Management, having a federated data model that enables all the product teams to work autonomously on the data catalog, data lineage, compliance etc, making sure that those are in  place. But when we talk about the Fountainhead, we talk about the capabilities…When you're doing ML ops, you shouldn't just start from scratch with something; you need to have building blocks in place. That's what we have done.

Now the basic idea is that every team in H&M can drink from the Fountainhead, he says:

Internal teams go to the Fountainhead and they take a platform and say, ‘We're going to do model development.’ Well, here are recipes and cookbooks for that. 'We want to set something up'. Well, here are common standards for setting it up. 'We need to upskill in our team in the literacy part' .Well, here are the trainings.

Developing some of these things centrally makes sense because if you remember how hard it was in the beginning -  the first time you started putting things into production, the first time you did a few models - you needed to write everything from scratch. We don't want that to happen. We want to take development time down from 12 months as it used to be, down to six months in just a few months. We also want to take it down to few weeks in the future as well.

The organization as a whole has gone from vertical to horizontal capabilities to be able to scale more effectively, he says:

We can't just add more people, it doesn't make sense. There's no way I can actually say to my stakeholders that this investment is going to pay off. So what we're doing now is, for instance, when we do pre-season demand forecasting...if we're using similar methodologies, why don't we just go to horizontal capability? This is also affecting the way we prioritize. We start investigating more and more.

And there is, of course a big cultural change angle here, as he noted:

How do we enable the talent to become even more sufficient?  We don't want to hire talented people and tell them what to do. We want them to come in and help us and tell us what we should focus on as well.  We're doing a full revamp of the technology architecture, again we are on version four now with our reference architecture. We are also launching large scale initiatives to enable the H&M group with the upscaling initiative across the entire region. We are going from vision to actions. We have done the proof of concept, we've done the proof of values, we are in production with our use cases. Now we're starting to develop the improving all of these components together.

A grey colored placeholder image