AI in HR: from promise to practice
Published June 16, 2026
- Data & AI
Key takeaways
- AI in HR has moved beyond the POC stage, but has not yet scaled: use cases are multiplying, but their industrialisation is still hindered by data quality, process robustness, and regulatory challenges, far more than by the technology itself
- Value is not everywhere: it is all about prioritisation: Recruitment, training, and HR services deliver quick wins, while more advanced use cases (attrition, weak signals, etc.) require a solid data foundation and strong governance before being activated
- Scaling requires a shift in approach: from “test & learn” to product-driven management. Mapping use cases, identifying the right ones, structuring data, and redefining governance become key conditions for turning trials into success
All HR departments have their own proof(s) of concept (POCs), in varying numbers. However, use cases deployed at scale across organizations are still not the norm. Technology is rarely the limiting factor, as vendor capabilities have never been as advanced. Instead, data, processes, regulatory risk, and change management are the main barriers.
This study, the second instalment following “Work and organizations: what is AI reshaping”, examines the HR function from within. It draws on around twenty interviews with CHROs and HR Tech vendors. A common thread runs through it: distinguishing what creates value in production and identifying high value-added use cases through a systematic review of the HR function’s scope. Six guiding principles emerge for the deployment of AI within the HR function.
6 learnings
Four main pathways to AI are available in HR: general-purpose LLMs, native features of HR information systems, specialized solutions from HR Tech providers, and in-house development. They already coexist, without necessarily being known to everyone. The right selection criterion is not technological but operational: the need determines the pathway.
Not all HR processes have the same potential. Recruitment and learning are already well-established areas. People Operations & HR Services offer the promise of the fastest and most measurable gains. Compensation & Benefits and employee relations still call for caution. More ambitious use cases, attrition prediction, weak signals of disengagement, talent marketplaces, benefit from waiting until a robust data and governance foundation is in place. In all cases, selecting the right use case and the right HR process will depend on the maturity of your organization and on the conditions outlined in the following paragraph.
A strong HR use case is not necessarily the most impressive. However, it almost always starts from a concrete operational pain point. Five non-exclusive conditions can help determine whether the identified use case is a strong candidate: a real business pain point or a clear gain for HR and/or for employees or managers; usable data; the ability to be deployed quickly; and a certain level of regulatory and social acceptability. Naturally, not every use case will meet all these criteria, but they can serve as a compass to move beyond a simple POC.
AI will amplify dysfunctions or temporarily mask an issue, but it will not fix it. A flawed process augmented by AI industrializes its own shortcomings. Recruitment with unclear criteria multiplies its biases at scale. Before anything else, a central question should be asked: should this process be accelerated as is, or fundamentally redesigned?
HR data is more fragmented than in any other function. Introducing, for example, an AI agent can quickly make this visible. Some use cases identified in this study have shown it: solid work on data and document repositories is a major prerequisite for any AI implementation in HR.
In the future, as soon as an AI agent operates within HR processes, the model “IT deploys, HR uses” will no longer hold. The HR function will move from being a “simple” user of tools to a true designer, managing AI products. Establishing governance early creates the trust infrastructure on which each deployment depends. Defining the modalities of human–machine interaction and determining what is delegated, or not, will become expected from the HR function, both for its own use cases and for those of the entire organization.
