Retail AI Roadmapping: A Data-Driven Portfolio Decision Engine

Abstract

Retail companies face a growing number of possible AI initiatives, but the harder problem is often not idea generation. It is deciding which use cases should be prioritised, which ones fit the company context, and which ones are realistic to move beyond experimentation. This thesis addresses that problem by developing a structured decision-support framework for AI use-case prioritisation in retail, based on an organisation × use-case representation that combines company-side attributes, use-case-side attributes, and computed fit logic. The empirical material comes from interviews with 33 Nordic retail companies across Sweden, Norway, Denmark, and Finland. The sample was weighted towards larger retailers, with 16 companies in the SEK 1–10bn annual revenue range and 12 above SEK 10bn, and included senior respondents such as CIOs, CTOs, CDOs, and heads of data, technology, or AI. Across these interviews, respondents listed 151 AI use cases, of which 55 were coded in greater depth with structured attributes and managerrated impact assessment. The study combines descriptive portfolio analysis, baseline impact modelling, model interpretability analysis, cluster-strength analysis, and qualitative interpretation. The baseline impact models showed some predictive signal in the structured interview attributes, especially for focused non-fit feature sets with tree-based ensemble models. However, explanatory performance remained modest, and predictions were compressed towards the middle of the impact scale. The models should therefore be read as support for prioritisation rather than as precise impact predictors. The results suggest that AI use-case prioritisation in retail should not be reduced either to GenAI adoption or to one universal ranking of use cases. GenAI adoption was widespread, but traditional ML use cases often showed clearer operational impact in the observed material. Data availability was generally stronger than data quality and accessibility, while process-change burden appeared more prominent than technical integration burden. In the observed-only cluster-strength results, internal delivery, high data readiness, high change capacity, and forecasting or inventoryplanning use cases appeared as recurring strong patterns, although these should be interpreted as signals rather than general rules. Overall, the thesis supports a context-sensitive playbook logic. Retailers should prioritise AI use cases by asking where the operational pain is, whether the relevant data is reliable and accessible, what portfolio role the use case should have, what delivery route fits the context, and what operating model is needed after launch.