Good AI systems require good data: What to consider when looking to data in hiring decisions

Ferrari may have a really fast car, but it’s not really that useful without the fuel powering its systems.

Recruiting teams are racing to incorporate new AI tools into sourcing, screening, scheduling, and other steps along the TA process. As teams spend more time working alongside AI technology, the data underneath these systems becomes incredibly crucial.

As these sophisticated AI systems become even more entrenched in TA, recruitment leaders must fuel their systems with high-quality, clean data that’s structured for their unique purposes.

“As organizations are beginning to increasingly deploy AI in their systems, data becomes the foundation,” said Vishnu Shankar, VP of platform for the AI-powered talent intelligence company Draup. “That’s the reality: The AI is only as intelligent as the data that it relies on.”

Shankar told HR Brew that data structure in an AI era boils down to three pillars: the quality of the data, good governance and practices, and the AI systems using it. Shankar said he commonly hears complaints from HR pros that “my data is messy” and said Draup organizes data maturity into five different stages, from messy excel spreadsheets and no systems or architecture to high-performing, regularly maintained AI systems running in real-time.

Not so fast

“In this journey of AI-enabled talent acquisition, most people are ready to jump into acquiring AI tools, thinking it’s like a magic wand that will get the job done,” Shankar said. “But I think the most important step that we also emphasize [to] many of the customers that we work with is: start with assessment. Assess where we are. Assess do we have the right data, right governance, right technology in place.”

Shankar told HR Brew moving up on the maturity framework makes sure AI tools are set up for success. He added that while five stages of maturity may sound like a lot, “leveraging some external intelligences can help accelerate.”

Customers using larger AI-enabled HCM software may be able to leverage existing data inside their platforms or begin collecting and preparing data that informs specific AI outcomes with the flick of a switch. For customers of the enterprise software giant SAP, activating AI inside its suite of products, including its HCM SuccessFactors, is a “toggle switch,” according to its chief AI officer, Jared Coyle.

Coyle cautioned that there are cycles to data, especially in TA, so data collection over time will best capture the fullest picture, pointing to seasonal hiring in retail as an example. But customers could supplement that with historical data.

Part of the onboarding process for SAP customers, he said, is assessing the health of their data and figuring out the best way to mature it so AI can deliver its expected outcome. SAP’s copilot tool, Joule, he said, interacts with customers in natural language. They can tell the bot, in plain English, to ingest specific materials to build out the data for other processes.

“You’re using AI to generate the data that’s then going to use AI again on top,” he said. “It becomes this interesting cycle.”

Glass houses. “We are looking at challenges with two sets of data. One is getting the internal data, and this is the easier problem to solve, because you know what your business priorities are. You know what industry you operate in. You know why you need this role,” Shankar said. “It’s a matter of having the right processes in place to have a very good [job description] in place.”

Many job descriptions are in serious need of face-lifts, he suggested. The tasks they list don’t always mirror those performed by current employees, and new tasks, like those related to AI, may not even be included. Conversely, they may include tasks AI can do so well that they don’t belong in a 2026 job description.

This is why Shankir suggested a dynamic skills framework, so organizations can understand how the work is done and by who or what in real time.

“It’s extremely important to ensure that the data that we have is good enough to make really good recommendations, ensure it recognizes the right candidates and also factors in for rapid changes that is happening in the space, whether it’s new skills that are coming up, impact of AI that’s happening, and the context of the business,” he said.

Tony Castellanos, head of talent and comp at Nextdoor, has seen firsthand how AI adoption requires companies to rethink their information architecture.

“I think that this has really put a lot of pressure on how you design an effective information architecture,” Castellanos said. “We spent a lot of time building the foundations there to ensure that whatever we’re building on top of that…is something that’s going to be effectively informed. This is one of the interesting things I think about AI and automation, is you can’t sit these layers on top of how we’ve historically worked.”

At Nextdoor, that meant cleaning up central data sources that inform AI tools used by the company before deploying them for employees to interact with.

“What we have tried to do is really to do a big cleanup of…our core sources that inform company programs, so that when you have these agents that employees are engaging with, that it is going to the correct sources and that it’s delivering the content that we want to be delivered,” he said.

They’ve also adjusted how information is captured. Meetings are summarized in real time, ensuring decisions and context aren’t lost forever in someone’s notebook.

“What we are creating is something that is manicured or curated and more real-time reflective of what we are doing,” Castellanos said.

External context

Recruiting AI systems don’t just rely on internal workforce data. They also rely on sets of external candidate data—which can also get complicated.

TA teams must configure systems that can infer signals from external sources (candidates) that may look different from the taxonomy or structure inside their talent org. A candidate may list one skill that implies another. As the dataset matures, skills taxonomies become more sophisticated and systems better integrated, so context isn’t lost along the way.

“The external data, it’s what the candidate self reports, it’s what the candidate says they do,” Shankar said. “That’s where sometimes the candidate may be over-reporting what they are, under-reporting who they are, or may not be giving enough context about how they’re relevant for the job. That’s where it also becomes very important to leverage whatever information the candidate is given to interpret as much as possible.”

Do it right

Worth pointing out is that candidates entrust companies to responsibly protect and use their data when it comes to job applications and hiring. So beyond the context for the tools, good governance is important.

Coyle said the overarching imperative for data in the context of responsible AI requires “relevant data, reliably given, and responsibly executed.”

According to SAP’s public ethical principles guide, all the company’s development and use of AI tools inside the SAP suite and its offered externally as “a framework for how you can respect human dignity in all of this and make sure that you’re executing AI in an ethical manner, because it’s actually relatively complex to do that,” he said.

And AI could also help cull bias if deployers know and understand the outcomes they’re looking for and trying to avoid. Amazon’s head of manager enablement and inclusive hiring learning, Shabrina Davis, told HR Brew in February to leverage tools and data to reinforce desired outcomes as well.

“One of the positive ways of using AI in talent acquisition is identifying the gaps, identifying the areas where bias does exist, and then being able to tackle that uniformly and globally with a process or with something that’s automated,” she said. “We can remove this particular bias from the process, and the data can either tell us that we’re doing it or we’re not.”