It's the AI era, why are you still using that outdated method to select candidates?
Organizations hold extremely high expectations for talent reviews, but the reality often falls short. A survey of over 300 global enterprises by the well-known HR consulting firm Talent Strategy Group shows that the average accuracy rate of identifying high-potential talents through traditional methods is only 44%. A survey of CHROs at large corporations by management consulting and organizational research firm Gallup reveals that while 64% of companies use the 9-box grid for succession planning, only 9% of CHROs consider it truly effective; even more strikingly, merely 3% of CHROs believe their organizations excel at identifying and selecting the right managers. This poor performance does not stem simply from managers' cognitive biases or HR inaction. The fundamental crux lies in the fact that traditional assessment methods cannot overcome the innate limitations of the human brain in processing complex information.
When the Old Brain Cannot Process New Talent
Cognitive neuroscience research suggests that the difficulty of talent reviews does not arise from the failure of a single link, but from the combined constraints of multiple limiting factors when the brain performs the task of "evaluating people".
Imagine a typical talent review meeting. The information managers need to process simultaneously includes at least: the multi-dimensional performance of the employee under discussion, comparison references with other employees, the criteria for each assessment dimension, and the organizational strategic needs behind these judgments. Cognitive science research shows that the capacity of human working memory is only 4±1 information chunks. When the number of information chunks in the discussion far exceeds this upper limit, the brain is forced to activate simplification strategies — either concluding by anchoring only one or two most prominent features, or giving vague scores that tend to be moderate. This perfectly explains why the 9-box grid, although often criticized for being simplistic, has been difficult to replace by more complex assessment frameworks in the past few decades.
Moreover, there are more hidden interferences lurking in the assessment process. Research by scholars from the Department of Psychology at the University of Alberta reveals that when managers verbally evaluate someone as "having strong leadership", because "leadership" frequently co-occurs with terms such as "strategic thinking", "judgment", and "communication skills" in daily contexts, the brain automatically activates these "semantic neighbors" and unconsciously raises the scores of other dimensions.
The core bottleneck of talent reviews is not the inadequacy of management processes or the lack of assessment tools, but that it is always constrained by the human brain's limited information capacity, cognitive architecture, and memory mechanisms. Since the crux of the problem is deeply rooted in the underlying mechanism of "how the human brain processes information", the real breakthrough should be to explore a completely different information processing paradigm.
Use AI to "Change the Brain" and Form a New Paradigm for Talent Reviews
Different from the human brain, which is constrained by limited attention, memory retrieval, and subjective integration, AI can process talent data with an alternative logic to improve the matching and prediction capabilities of enterprises in talent reviews.
Matching: Vectorized Matching Replaces Tag Management
Traditional talent reviews have an implicit principal axis of work: tagging, and the essence of tags is information compression. In the traditional model, managers and HR screen and compare based on these compressed tags to complete person-job matching. The biggest problem with this approach is that it uses preset, static tags to compress information. Traditional "tagging" is a top-down centralized behavior, where organizations usually need to first build a predefined skill library before assigning tags to employees. This system is only effective in a stable environment.
But in the AI era, the half-life of skills has sharply shortened. A report by IT research and consulting firm Info-Tech points out that the half-life of functional skills has dropped to 2.5 years; a survey by the World Economic Forum also shows that 39% of core skills will undergo fundamental changes by 2030. Compressing employees' characteristic information with outdated tags is no different from the blind men touching an elephant in the new era.
To solve this problem, it is not necessary for the organization to mobilize a large number of manpower to update and generate tags that meet organizational needs in real time, and then have HR assign tags to employees. Fundamentally, organizations will no longer need so many preset tags in the future. In the new matching paradigm, real-time "vectorized matching" is the core solution. First of all, a vector is a set of numbers used to describe a thing, and the spatial distance between points directly maps the similarity of things. By calculating the distance between vectors, the system automatically associates content with similar connotations, transcending the differences in literal expressions.
When applying this logic to the talent matching scenario, its role becomes very obvious. Taking the recruitment of "back-end engineers" by a business department as an example, traditional systems often can only retrieve based on the "back-end" tag, so employees who are not explicitly labeled are easily excluded from the candidate pool. However, in reality, many employees whose job titles are not "back-end engineers" have long been engaged in Python development, database management, interface design, and other work, and their ability structure highly matches the requirements of back-end positions. The key value of AI is that it no longer rigidly adheres to literal consistency, but can identify semantic associations of "different expressions but similar abilities", thus discovering potential candidates obscured by traditional tag systems.
This is not a theoretical assumption — leading enterprises have already put vectorized matching into practice. In 2023, HSBC cooperated with AI technology company TechWolf to identify semantic associations between work content and skills through vectorization technology. The system established a unified classification system covering 18 skill domains, 48 subdomains, and 398 skill clusters, mapping more than 2,500 job roles, and generating detailed skill profiles for over 30,000 technical employees. Hundreds of digital and technical talents who had long been hidden from management's view were systematically identified.
However, although the above practice introduced vectorization technology, it did not completely subvert the old model. Although semantic associations have been established for person-job matching, the final output still falls into the preset classification framework and tag management. In contrast, Eightfold AI, a human resources technology company serving enterprises such as Salesforce and Microsoft, has adopted a more thorough path: using vectors directly as the basis for matching, omitting the intermediate step of tags. It trains models on hundreds of millions of talent profiles and completes recommendations by calculating semantic distances. Under this model, person-job matching is completed in a vectorized manner, and tags only appear when it is necessary to explain the results to humans. According to an internal platform data analysis by Eightfold AI covering approximately 690,000 employees, employees recruited through AI screening with high matching scores have a nearly 50% higher promotion rate within two years of employment than the low-matching group, and their 12-month retention rate is also relatively higher. A comparative experiment published in Information Sciences in 2026 also shows that in multi-domain person-job matching tasks, the accuracy of semantic matching is about 2-5 times higher than that of traditional keyword tools.
Prediction: The "Meta-competence" of Individuals
The work accomplished by matching is essentially based on the present and looking back at the past. However, directly applying the same method to predict the future will encounter difficulties. The reason is not that vectorized matching fails in prediction scenarios, but that the target for prediction is not stable: in the AI era, the standards for defining outstanding talents are constantly changing. When the concept of outstanding talents continues to drift, any static fitting based on existing sample portraits may gradually become inaccurate.
Therefore, talent prediction cannot only answer "who is more like today's outstanding people", but must further answer "who is more likely to adapt to tomorrow's changes".
By analyzing the transcribed texts of 303 episodes of the well-known tech podcast Lenny's Podcast from 2022 to 2026, the author conducted a systematic analysis of the discussion on "what capabilities people should possess" in the AI context. The results show that among over 400 valid mentions, capabilities account for the highest proportion (about 55%), followed by traits (about 30%), and professional skills the least (about 15%). Among the capability-related descriptions, the most frequently mentioned ones are learning ability, strategic thinking and vision, and adaptability; among the trait-related descriptions, high-frequency terms include curiosity, taste, and intuition.
This result at least indicates that in the discussion of talents in the AI era, what attracts the most attention is no longer just a specific technology itself, but the higher-order capabilities that can support continuous learning, cross-situation transfer, and rapid adaptation.
We summarize this type of higher-order capability that transcends specific domain knowledge and can support individuals to continuously adapt and reconstruct their ability structure in new situations as "Meta-competence". This concept was first proposed by management scholar Burgoyne.
Among the many theoretical frameworks of meta-competence, "learning agility" is the core concept most widely applied in current management practices, and has been deeply embedded in the talent assessment systems of leading enterprises such as Tencent, HSBC, and Procter & Gamble. The founder of leadership development and talent assessment firm Lominger defines it as: the willingness and ability to learn quickly from experience and transform learning into successful performance when facing new, severe, or changing situations, which has been developed into five sub-dimensions: mental agility, people agility, change agility, results agility, and self-awareness.
The assessment of meta-competence should shift the evaluation focus from "giving evaluations" to "measuring behavioral performance and competence transfer trajectories in new situations". Specifically, first of all, meta-competence needs to be decomposed into observable behavior patterns.
Taking "change agility" in learning agility as an example, its behavioral anchors can be summarized as "staying curious about new ideas and practices, being willing to face new situations, and actively seeking innovative even risky new methods". Different from traditional assessments that are easily interfered by expressive ability or a single outstanding performance, AI can extract continuous evidence chains from multi-source systems such as project experiences, performance documents, and collaboration records. It constructs trajectory models by extracting key events and behavior processes, transforming subjective perceptions into judgments based on objective facts. At the same time, situational simulations can also be introduced to focus on examining how candidates ask questions, identify constraints, and reconstruct solutions under limited information by constructing task scenarios with high uncertainty. Finally, the assessment results should not be compressed into a general "total potential score". A practical approach is to output a multi-dimensional profile including dimension scores, evidence summaries, and risk warnings.
5 Actions CHROs Can Take Immediately
The real implementation of vectorized matching, meta-competence assessment, and interpretable AI systems requires a lot of time investment. Before they are ready, CHROs can immediately practice the following five actions to change their thinking and habits, serving as the starting point for leveraging paradigm transformation.
1. Write a Job Profile That AI Can Understand
Most job descriptions are written for humans, not for machines. The upper limit of talent matching first depends on the clarity of talent standards and the semantic density of input information. The more ambiguous the input, the more distorted the output.
Select the most critical and hardest-to-recruit position, and spend 1 hour rewriting its job profile. There is only one principle: do not use vague adjectives, only write specific, observable, and verifiable requirements. For example, specific responsibility requirements, skill requirements, and the type of people you do not want. In the short term, it can be used for general large language models to initially screen candidates. In the long run, it is a key input for future vectorized matching systems.
2. Draw a Talent Data Map
List on a piece of paper where all "people-related" data in the organization is stored and who the managers are. The significance of this step is not to immediately connect the data, but to let the organization see for the first time how many data mountains lie between the present and the actual operation of vectorized matching.
3. Write Down Your Predictions About People, Seal Them, and Review Them Three Months Later
Select ten key talents and write down your judgments about them in the next three months: who will get promoted, who will leave, and who will stand out in the new project. Seal the written content, open it three months later, and verify it one by one with detailed factual data. This is the step with the lowest cost but the most discomfort, because it requires managers to test their own judgment models before promoting the organization to "use AI to improve judgments".
4. Select a Meta-competence You Truly Recognize, Decompose It Into Behavioral Anchors, and Find Three People to Run a Trial Evaluation
The first step is to select a meta-competence you truly recognize. The second step is to decompose it into observable behavioral anchors. The third step is to select three people — one you are highly optimistic about, one you have long underestimated, and one you have never been able to see clearly. Require writing evidence for each item before scoring. The value of this step is to test whether the organization's existing observation system is sufficient to support meta-competence evaluation.
5. Introduce an "Opposition Mechanism" in Talent Review Meetings
Find a talent review meeting and require that for each key talent under discussion, managers must provide two types of evidence at the same time: evidence supporting the current judgment, and evidence not supporting the current judgment, with no less than three items for each. In essence, this action artificially increases the weight of reverse signals at the lowest cost, to avoid the drawbacks of relying on subjective experience for prediction.
The purpose of talent reviews is actually to find certainty in an era of uncertainty. In the AI era, where does certainty lie? Perhaps the answer has changed. Certainty is no longer a clear anchor point, but a forward momentum. If in the midst of tremendous changes, it is destined to be impossible to see the future clearly, then the certainty we pursue may be hidden in every step that every individual strives forward.
This article is from the WeChat public account "Harvard Business Review" (ID: hbrchinese), written by Shen Kaijie, and published with authorization from 36Kr.