Sander on BUILDERS: AI at retail scale is an organisational problem before it is a technical one

Sander on BUILDERS: AI at retail scale is an organisational problem before it is a technical one

Sander, co-founder of Ardinois, joined BUILDERS, the podcast by Proxify, for the episode “How AI is transforming retail at H&M”. The conversation draws on his time leading AI and machine learning at H&M and earlier roles at Nike and bol.com, and moves from data foundations and team design to demand forecasting, personalisation, and where the next wave of value will come from.

Governance, not infrastructure, is the constraint at scale

The technology to handle retail-scale data exists. Hyperscalers and modern data platforms have solved the volume problem. What they have not solved is agreement: a customer is defined differently by finance, by e-commerce, and by the ERP. Data must be granular enough for data science and harmonised enough for reporting, and holding those two requirements together is a question of governance and organisational process, not of platform capability.

A model has no value on its own

Value is created when someone takes a different decision because the model exists, or when the model is embedded in an automated decision. Everyone in the delivery chain, from data scientist to engineer to product manager, needs to understand how that value mechanism works for the solution they are building. Without that shared context, teams over-engineer features that do not matter and under-invest in the ones that do.

The translator function is now critical

Every enterprise software vendor is embedding AI in its product, and every business stakeholder arrives with expectations shaped by conferences and LinkedIn. Organisations therefore need people who pair business fluency with a working understanding of what AI can and cannot do: typically a product manager and an engineering lead operating as a pair. Their role extends beyond building. They guide buying decisions, separating what is real from what is overpromised in vendor conversations.

The language model is a new employee on a permanent first day

An experienced analyst knows which table holds the clickstream data and which one to avoid. That knowledge was built over years and lives nowhere except in their head. A language model arrives with none of it, and in most deployments it does not learn over time. The conclusion is structural: the organisation’s information landscape must be made accessible and well described, because the model cannot compensate for a landscape that is not. Preparing data for a single use case works, but it does not scale. The shared layer does.

What comes next

Embedding models are quietly solving the cold-start problem in demand forecasting for new products. Hyper-personalisation is bringing the in-store conversation to online channels. And shopping through assistants such as ChatGPT and Claude may become a sales channel retailers have not planned for. The common thread: the technology keeps improving, but capturing its value depends on how organisations adapt their processes and people. That takes time, and the competitive question is who converts fastest.


Listen to the full episode of BUILDERS by Proxify on YouTube or Spotify.

Frequently asked questions

What is the real constraint for AI at retail scale?

Governance, not infrastructure. Hyperscalers and modern data platforms have solved the volume problem. What they have not solved is agreement: a customer is defined differently by finance, by e-commerce, and by the ERP. Holding granularity for data science and harmonisation for reporting together is an organisational process question, not a platform capability question.

What is the translator function and why does it matter?

People who pair business fluency with a working understanding of what AI can and cannot do — typically a product manager and an engineering lead operating as a pair. Beyond building, they guide buying decisions, separating what is real from what is overpromised in vendor conversations.

Why do language models struggle in enterprise deployments?

A language model is like a new employee on a permanent first day: it arrives without the tribal knowledge an experienced analyst built over years, and in most deployments it does not learn over time. The organisation's information landscape must be made accessible and well described — the model cannot compensate for a landscape that is not. A shared information layer scales; per-use-case data preparation does not.

Where is the next wave of AI value in retail?

Embedding models solving the cold-start problem in demand forecasting for new products, hyper-personalisation bringing the in-store conversation online, and shopping through AI assistants such as ChatGPT and Claude becoming a sales channel retailers have not planned for.

Let's talk about
where you are.

30 minutes. No pitch deck. No obligation.