I’ve been writing several articles recently looking at how HR vendors are using AI.
The goal is to help HR professionals get a glimpse under the hood so that they can better understand what the AI component is doing, not seeing it as a black box.
Creating data
Whilst doing these pieces, I’ve had a chance to speak to Hariharan “Hari” Kolam, co-founder of the talent acquisition platform Findem, about how it uses AI.
My first question is always about data, specifically – “What data is the AI using?” because this is what helps ground my understanding in something tangible.
In the case of Findem, it turns out this is a very important question indeed.
Because the trouble with sourcing talent is that resumes, on their own, are a poor source of data. No matter how smart your AI is, if it only relies on resumes to find the best candidates then there is only so much it can do.
Findem’s approach, however, is to enrich a person’s profile with data about where they worked and when they worked there.
This data can tell you, for example, if someone was a software engineer at Google when it was a 500-person company vs a 50,000-person one.
Why does this matter? Weill, it’s because this gives the system much more to work with, to determine which people might be a good fit for a given job.
Creating attributes
As you can imagine, making sense of unstructured and unverified data about companies and careers is not easy.
To accomplish this though, Findem’s platform generates talent data from hundreds of thousands of sources.
Billions of person and company data points become more than 1 trillion attributes.
This data lake contains new talent data in the form of attributes for every person, team, and company.
These are the skills, experiences, and characteristics on which talent decisions are really based.
It’s worth noting that the amount of data we can handle with AI (millions, billions, etc.) is orders of magnitude greater than what we dealt with in the past.
Querying data
Once a data lake about potential candidates is created, recruiters will want an easy way to query it, to find promising candidates.
Findem has developed middleware that leverages several public LLMs, fine-tuned LLMs, and its own micro models to interpret a recruiter’s intent in natural language.
It separates interpreting the intent of a recruiter’s queries from the generation of the data about candidates.
Having the language model and intent separate from the service gives users in the middle the opportunity to confirm, reconfirm, and validate the results.
As with all LLMs, it’s possible to ask novel questions in natural language. For example, recruiters can search with queries such as: “I’m looking for a software engineer who is an expert in Java and worked at a fast-growing startup at some point in their career.” A candidate’s resume may not specifically note that they worked for a growing startup, but the AI can infer it from the history of the individual and whether they worked for pre-IPO companies that grew significantly during their tenure.
Another example of the sort of attribute that would have been impractical to answer without AI technology is: “We want an operations leader who has built loyal teams.”
Again, the AI can make sense of the query and look for leaders whose teams have followed them when they changed companies.
These examples show how a richer data set and the ability of AI to make sense of natural language queries is a big change from searching for resumes by keywords.
How do you know if it’s working?
A good question to ask of any AI-powered tool is: “How do we know if it’s working?”
This is related to the question of explainability – ie can the AI explain why it said, what it said?
In the case of Findem, the “is it working?” question can be answered at two levels:
- One level is the recruiter’s feeling as to whether the system is turning up good candidates who have the attributes they were looking for.
- The other level is to judge the system from the viewpoint of the talent acquisition process. T
The system provides data and analytics on each stage of the talent acquisition pipeline.
At this point it’s not about judging the AI per se, it’s about judging the whole system of which AI is one component.
Takeaways
Here’s what I think we can take away from all of this:
- What differentiates one AI-based HR tool from another may not be so much the AI itself, but what data the tool can access.
- Understanding the data an AI is using will be a key to understanding what the tool will be capable of
In an earlier article, we noted that Knockri was a vendor steeped in industrial-organizational psychology and was applying AI to build on that.
Findem is steeped in data management and applies AI to build on that expertise.
It will often be the case that vendors have expertise in something and use AI to leverage that.
I think it’s exciting when we see AI go beyond automating a process, enabling us to do something new.
In this case, it’s exciting to be able to search for candidates with attributes not specifically mentioned in their resumes.
Finally…
I asked Kolam how Findem uses AI internally.
He says it uses AI across the board from marketing, to coding, to HR.
The lesson is that if GenAI tools are not being used through your organization then you are at risk of falling behind.