In my previous article, I explored the idea that AI, like raising a child, requires thoughtful guidance, patience, and consistent support to realise its full potential. This nurturing analogy highlights a key truth: AI starts in its cradle with basic capabilities but needs structured care to grow into a capable partner for HR. Transforming AI into a thought partner requires us, as HR practitioners, to go beyond its generic training and provide the tailored foundation it needs.
When we hear that a model is trained on billions of parameters, it sounds impressive. However, this training is often based on vast, generic data, giving AI surface-level knowledge across varied topics. While this enables it to handle simple tasks such as summarizing engagement surveys or tracking headcount trends, lacks the depth required for fields such as HR, where decision making is complex. Additionally, much of this data includes synthetic or randomly generated information, artificially created to increase dataset size but often lacking real-world context. For example, synthetic data might simulate employee feedback trends or hiring patterns but fail to reflect the unique nuances of specific industries or organizational cultures. These inaccuracies and biases corrupt the model’s ability to provide actionable insights grounded in real-world HR contexts.
To develop AI from cradle to truly capable, we must train it with high-quality, HR-focused data and techniques including Chain-of-Thought (CoT) reasoning*. By building a tailored model, we enable AI to think through HR tasks step-by-step, offering deeper insights to support complex decisions.
This journey goes beyond basic AI implementation and focuses on i) feeding high-quality data, ii) developing a nuanced vocabulary, iii) fine-tuning responses, and iv) monitoring / checking for biases.
Key Milestones in AI Prep Journey
Here’s a closer look at how we can set up AI for long-term success in HR.
i) Feeding High-Quality Data to get Reliable Insights
Think of high-quality data as clean, well-prepared information that goes beyond simply gathering siloed content. It must be accurate, consistent, complete, timely, and relevant for meaningful analysis and insights (Think ACT – Accurate, Timely and Complete). Add data that you think will bring in value and remove what will muddle it. Start by filtering out outdated or irrelevant data, removing blanks and duplicates, and standardizing formats. For example, feeding org structure data from 10 year ago may not be helpful. Even worse if we have functions and BUs with old names that have no link to the present org structure. AI will be as clueless (or more - as it does not know what to ignore) as us in using this particular data.
ii) Building a Vocabulary for AI to Understand HR Language
Developing a specialized vocabulary goes beyond listing terms. It involves defining each term within the specific organizational context. For example, DE&I can encompass a variety of dimensions, but in an organization context, it primarily focuses on gender parity, inclusive hiring, and cultural representation, among other areas. Similarly, engagement scores may include custom survey questions unique to organization goals, such as measuring career growth opportunities or work-life balance. By curating a glossary with detailed definitions, we give AI a nuanced understanding of our HR language, ensuring it interprets terms accurately within our framework. Good, clean, and comprehensive documentation is the elixir that supports this process and provides AI model with a reliable reference point. It standardises interpretations and creates consistency, allowing AI to respond effectively and accurately, especially as terms and organisation practices evolve.
iii) Fine-Tuning AI for HR Applications
Fine-tuning builds on the vocabulary by training the AI with HR-specific data that reflects real-world scenarios unique to an organisation. While the vocabulary provides a foundational understanding of key terms, fine-tuning takes it further by using actual performance review data, engagement survey results, and other HR records to help the model recognise patterns and apply terms accurately. For example, after developing a glossary that defines “DE&I,” fine-tuning would involve training the AI with past DE&I reports, allowing it to recognize different diversity initiatives, measure progress accurately, and generate insightful analysis. We are teaching AI to apply terms in practical contexts, making it more responsive and capable in its HR applications.
iv) Monitoring for Bias and Ethical Considerations
AI models are often biased to start with because they learn from historical data that reflects existing societal biases or imbalances, and they lack the contextual understanding to identify or correct these patterns. For example, recruitment models are shown to favour male candidates for technical roles or rate specific demographics lower in performance reviews. By regularly reviewing AI recommendations, HR can spot the biases early on and adjust inputs as needed. Training AI on a broad set of examples representing all demographics and job functions can help reduce biases in recommendations. This proactive approach not only improves AI’s accuracy but also reinforces the fairness and inclusivity we aim to promote in our workplaces.
Just as we nurture talent within our teams, nurturing AI allows it to grow into a reliable partner that empowers us to make data-driven, people-centric decisions. As HR practitioners, we have the unique opportunity to shape AI into a tool that truly supports and enhances our work.
What steps have been most effective in preparing AI models for your HR teams? Would like to hear what has worked well and what challenges still remain.
* We will discuss Data Cleaning and Chain-of-thought Reasoning in detail in the subsequent articles. (Updated: Here is the article on Data Cleaning)