Why a 'Data First' Mindset is Key to Making AI Work in HR

There is a silent barrier between the success and failure of AI initiatives in human resources, and it has little to do with whether employees fear the technology or have received sufficient training to master AI tools. HR data is foundational to artificial intelligence’s ability to generate accurate and reliable answers or recommendations on HR issues. And that HR data is often flawed in ways that blunt the technology’s impact and erode trust in AI as a decision-support tool.

Machine learning, generative AI (GenAI), and agentic AI have vast potential to save HR functions time, enhance decision-making, and make the department more future-ready. However, compared with other recent technological innovations, AI relies far more heavily on the existence of clean, accurate, and well-structured historical data to deliver on its promise. 

Data is the oxygen that gives life to algorithms. When HR datasets have problems, such as inconsistencies, anomalies, outdated information, or biases, they create flawed decisions, provide inaccurate forecasts, and degrade AI model performance. HR technology analysts say that reality should make data readiness — regularly auditing, cleaning, and structuring the raw HR data that exists in recruiting, performance, learning, or benefits databases so it’s suitable for use in AI models — a top priority for HR functions.

However, experts say that, for many HR leaders, getting their data in order is often a back-burner concern.

“Many HR professionals are still operating with the mindset created during the software-as-a-service technology environment of the past 10 years, where they just assumed their HR data could never be fully clean and they’d just have to make do with it,” said Lydia Wu, a former HR executive with Panasonic North America who now consults on HR data and technology issues. “But without data, AI doesn’t exist, and without quality data, the mistakes or poor decisions made by organizations will get repeated because AI is based on historical datasets and patterns.” 

Why Data Hygiene Matters

While investing resources in something as prosaic as data hygiene may not match the allure of purchasing a shiny new AI platform, experts have noted that data management is a critical step that is often overlooked at the peril of HR functions.

 “If an organization with limited resources is trying to decide where to spend its money, there’s a good argument for shifting resources they might spend on AI tools to data cleaning,” said Emily Rose McRae, a senior director analyst with Gartner who advises CHROs on the future of work. “You need good data management and governance to be able to take things to the next level in terms of creating better-informed decisions with AI.”

Michele Goetz, a vice president and principal analyst at Forrester who specializes in data management and business intelligence issues, said that ensuring good data hygiene is essential in the age of AI.

“Getting data ready for AI at the source level is increasingly important,” Goetz said. “That includes understanding your HR data lineage, improving governance across the data ecosystem, and figuring out which data sources are most trustworthy for the task at hand.”

Flaws in HR data occur for many reasons. As the data flows from its source in HR systems to a final destination in an AI platform — a process known as data lineage — it can transform and develop problems. These might include inconsistencies due to different formats or human error caused by manual data entry from one platform to another. 

Other common problems include mislabeled data, the existence of duplicates, missing values, outdated information, and biased data. HR also has to deal with a growing volume of unstructured data — such as written responses to open-ended survey questions, PowerPoint presentations, video files, podcasts, and emails — that is challenging to store, standardize, and analyze. 

Failing to address such data quality issues as part of AI initiatives is a cart-before-the-horse problem that can have significant ramifications for HR executives.  

“Data management is a project that, if you don’t pursue, you can get a phone call in the middle of the night from your CEO saying, ‘Why did this problem with our AI just happen?’ ” McRae said. “Then you’re forced to respond that you decided to push ahead with AI without good data cleaning and governance in place.”

The problem is far from theoretical. There are real-world examples of flawed data causing AI tools to draw biased conclusions or make inaccurate recommendations. The most well-known example is when Amazon discovered that an AI-driven hiring tool had discriminated against women applying for software engineering roles at the company. The tool was making recommendations based on analyzing volumes of resumes from men — not women — who had been hired for the role, and algorithms determined that being male was essential to success in the job. 

Wu said poor data hygiene can affect many of the ways that AI is used in human resources. She shared a real-life example in the compensation arena of a job role whose title had morphed from “analyst” to “associate” to “customer success manager” as a result of a series of company reorganizations. During this time, company recruiters also began sourcing jobs outside the Midwest, instead of the Northeast United States, and added a remote work option.

Because some of those changes were omitted from a compensation database and other data segments weren’t properly governed and cleaned, using AI to make a salary recommendation for that role became problematic. 

“If you use an AI tool designed to help value and price out that job role, do you really think the compensation range will be accurate given the conflicting data or the various keywords the AI will use?” Wu asked. “Those are the kind of data issues and nuances that can get overlooked.” 

Data quality isn’t a secondary concern for all HR leaders. Research shows that some are pausing their AI initiatives until they can get the HR data houses in order. A new study from Mercer found that 31% of HR teams that reported bypassing AI initiatives did so because they felt they weren’t ready due to data quality or infrastructure issues. HR has an important role to play in empowering workplace AI initiatives, and data hygiene is critical to that effort.

Establishing Data Ownership

HR technology analysts say another key issue in maintaining good data hygiene is establishing who has ownership of individuals’ data within the organization. 

“In many cases, there isn’t a clear ‘owner’ of data that’s been identified between IT and HR functions,” said Stacia Garr, co-founder and principal analyst of RedThread Research, an HR advisory and research firm in Woodside, Calif. “IT is often responsible for technology systems but not the data within those systems.”

Unless HR has a robust and mature people analytics function, it often doesn’t step forward to assume responsibility for the quality of HR data, Garr said. “So the question becomes, ‘Who owns the data, and do they have enough remit to go in and fix the data when quality problems are identified?’ ” she said. “Good data management starts with establishing clear authority and accountability for HR data.”

Another challenge faced by HR is consolidating data from diverse systems into one place to drive improved decision-making. “Data is absolutely everywhere today, and most organizations are still trying to get their arms around how to bring different data forms together and simplify their data landscape,” Goetz said. 

Some HR leaders turn to third-party providers to help merge multiple data sources from areas such as recruiting, performance, and learning into a single view, with the goal of creating a “single source of truth” for people data. Master data management (MDM) software tools can also help centralize and standardize business data spread out across an organization. Examples of these types of software providers include One Model and IBM’s InfoSphere MDM product.

Top Data Hygiene Problems

Some HR data hygiene problems are more prevalent and pressing than others. Experts interviewed for this story identified the data quality issues to which HR leaders should give high priority.

Inconsistent data definitions. Often, data terms or metrics that have one meaning in one part of the organization can have a different meaning in another. Garr gives the example of employee turnover rates.

“Turnover in one part of the company might mean voluntary turnover, and in another part of the company, it includes involuntary turnover,” Garr said. “There can be small definitional differences that lead to dramatically different insights or conclusions drawn by the AI that’s analyzing the data.” 

Garbage in, garbage out with AI. With tools like GenAI, where the quality of the output depends squarely on the quality of the input — the prompts written by users to generate desired responses — the garbage-in, garbage-out problem can be pronounced.

“Dialoguing and interacting with an AI tool is itself a data flow,” Goetz said. “So if a prompt isn’t clear or detailed enough to produce a user’s desired output, that also becomes a data quality problem.” 

McRae also said it’s essential for HR leaders to recognize the limitations of what can be done to control the hallucinations generated by GenAI platforms, particularly when those tools produce fabricated responses. Experts have found that recent “reasoning” AI models are more powerful but also more prone to hallucinations. 

“You can’t fix all hallucinations with improved data quality,” McRae said. “There is a component to hallucinations you can’t program out of the tools. It’s not that large language models don’t have the data needed to produce accurate or reliable responses. It’s just that these tools can produce false or made-up answers on their own.”

McRae said that some HR executives she knows are pulling back on the use of HR virtual assistants due to concerns about hallucinations. “They have a growing discomfort about the prospect of hallucinations with those assistants,” she said. “It doesn’t matter how good their HR data is; these leaders feel they can’t risk having a tool that is seemingly authoritative providing wrong or fabricated answers.” 

Inherent bias in people’s data. When HR data used to train AI models contains bias, the AI will simply replicate those biases in its outputs and recommendations. Data that is incomplete; favors certain demographic groups, genders, or ethnicities; or only represents certain geographies can produce biased or discriminatory outcomes.

“Biased data can be an issue across the HR spectrum where it concerns AI use,” Garr said. 

Outdated information. Tools like GenAI need current data to generate reliable and accurate responses. “For example, if you’re feeding information on company policies into an AI-driven chatbot, you want to make sure you have the most recent policy data available,” Garr said. “Data recency can be a problem for some AI models.” 

Using Automated Tools to Boost Data Quality

The explosive growth of HR data means cleaning and governing it through fully manual approaches has become unrealistic and risky. Many software tools exist to help automate and scale the data-cleaning process. Experts say HR leaders should consider three categories of such tools to aid in the cleaning and governance of people data.

Data observability software. Also called data profiling tools, this software helps evaluate the current state of data and identify where problems may exist.

“These tools assess the quality of data, how data may be changing as it moves through the organization, and whether there are problems like anomalies or other issues,” Goetz said. Examples of data observability software providers include Monte Carlo and Telmai.

Data-cleaning tools. Once data problems are understood, data-cleaning software can be employed to fix them. These tools correct errors related to duplicates, missing values such as demographic information, invalid or outdated information, inconsistent formatting, and more. Examples of data cleaning software providers include OpenRefine, Talend, and Trifacta.

AI governance software. AI technology is governed by a growing number of global regulations and policies, and the data underlying HR systems must conform to specific standards as well.

“These automated tools determine whether AI is operating as expected within the parameters of those regulations or standards and whether data within those systems is complying from a risk [perspective] and responsible AI perspective,” Goetz said. Examples of this software include Credo AI, Holistic AI, and Monitaur AI. 

While a bevy of new software tools can aid in the cleaning, consolidation, or governance of HR data, Goetz said human expertise remains essential to the task. 

“You still need people to oversee the evaluation and cleaning of data to prepare it for integration into AI tools,” Goetz said. “The management and quality control of HR data is complex and requires a variety of different skill sets.”

Dave Zielinski is a Minneapolis-based business journalist who covers the impact of emerging technologies on the workplace. He is a frequent contributor to SHRM publications.