AI and Integration
The general release of ChatGPT in November of 2022 changed everything. Unique use cases for ChatGPT poured in every day, changing previously-held beliefs on what was possible.
Since connectivity enables everything, we knew OpenAI’s API would open up even more possibilities. So we began investing in creating our own OpenAI Connector for MuleSoft to make connectivity easy. We knew that while we were working on enabling connectivity, we’d be able to think of lots of great ways to use it when it was ready.
And we did. Below is our first organizational use case integrating into OpenAI’s services, and it’s been incredibly valuable to Green Irony already.
The Problem
Like many technology companies, Green Irony receives a huge volume of resume submissions to open job postings.
This can be both a blessing and a curse.
It’s a blessing because we have the reach and employment brand to attract hundreds of applicants to open positions within days of posting.
It’s a curse because someone needs to go through these resumes and whittle them down into a candidate pool for phone interviews. With techniques like keyword stuffing now combined with the likes of AI-generated resumes, properly screening resumes is getting more challenging by the day. Not only does it require a lot of person-hours to do properly, but it also requires heavy domain knowledge and experience in the roles being screened.
In short, doing it right required a lot of internal experts’ time and effort, and these experts were already very busy. We needed to get them some help.
The Solution
Fortunately, Green Irony’s R&D team, armed with the most disruptive technology we’ve ever laid eyes on, was up to the challenge.
Using Green Irony’s OpenAI Connector for MuleSoft, our team delivered a set of APIs capable of providing a candidate rating against a job post, reasoning for the rating, and a set of questions to validate the candidate and ensure gaps are filled. This is all done real-time, triggered by a webhook integration on every inbound candidate submission and performed asynchronously, writing scores and other relevant contextual information back into our Applicant Tracking System.
Real-time integration allows us to quickly alert our team of promising candidates, enabling us to get in front of them faster. Our solution was driven by a belief in the “human in the middle” concept, so our goals are to arm our team with the best information, not to disqualify candidates based on an AI-only recommendation.
The secret sauce of the solution lies in a few key areas:
- Easy, connector-based connectivity into OpenAI, enabling us to focus on delivering and testing what matters for the use cases
- Prompt engineering and A/B testing of various prompts, ensuring we mitigate any hallucination issues and receive the most accurate feedback possible about every resume
- Working up an accurate job posting that is very clear about what success looks like and then parsing resumes for key information that is relevant to scoring against a job posting
- Fine tuning of the OpenAI AI model and continuously sampling results with our human experts, ensuring alignment between human ratings and feedback and AI-generated ratings and feedback (in progress)
The Results
- Labor Bandwidth Savings: The AI provides a sophisticated scoring system that significantly reduces the time spent on evaluating each resume, from an average of 10 minutes to mere seconds. This labor savings comes from critical resources who are very busy with other work in a fast-growing startup.
- Increased Interview Quality: The AI system offers detailed insights on each candidates’ strong matches and gaps in relation to the job description. This has resulted in stronger selections for interviews and reduction in the number of necessary interviews in half from an average of 6 to 3. Increased quality of interviews leads to increased quality of onboarded new hires.
- Enablement of HR Resources: The contextual information delivered by our solution enables our HR resources to ask better, more technical questions to assess the quality of candidates sooner within our process. This results in a higher quality of candidate pool reaching our second interview stage.
- Reduction in Recruiting Timeline: The use of AI allows for swift identification and rejection of unsuitable candidates, reducing time spent in the recruitment process. This has led to a 5X+ reduction in the amount of time we spend thinning 100 applicants into 15 for phone interviews, so we’re able to talk to better candidates more quickly.
Key Takeaways for AI
If someone would’ve thought of this solution 12 months ago, our first thought would be that it would take millions of dollars worth of technology labor investment to deliver because of how time-consuming machine learning is.
The release of GPT-3.5 shattered this point of view. We now have access to an extensible LLM capable of performing this level of expert analysis with the right prompts and tuning. We’re in uncharted territory, and it’s up to leaders to find the right tasks for AI, tasks to take off the plates of overworked humans.
To us, our own resume screening process was the perfect type of scenario for delivering value with generative AI. We had a very tedious, time-consuming task that also required experts to perform it. This task was a scale inhibitor since not only does it require so much time that an organization like ours can’t possibly keep up, it’s also VERY important to get right.
Generative AI has been a revelation for our applicant screening process, and it’s just the beginning.
