Issue #10: Building HR GPTs

This newsletter has been quiet in early 2024, but not because less is going on in the "future of work" arena or that it's less of a focus for me personally. Quite the opposite! I've been teaching a course on Algorithmic Responsibility to grad students at NYU SPS Human Capital Management , which just wrapped up this month.

This gave me a chance to dive even deeper into AI and the workplace — informed by focused research, connecting with great guest speakers, and the students' own questions and work. In the next few issues, I'll share some of that with you, with some topical future of work news mixed in as well.

In that spirit, this issue will focus on sharing results from the capstone assignment for the course: creating an HR-focused custom GPT using ChatGPT Plus.

Last November, Open AI released the ability to create Custom GPTs to paid subscribers, which meant they could make their own custom chatbots that would leverage ChatGPTs tech along with whatever instructions and data they provided it. The functionality got everyone very excited, leading to the creation of over 3 million custom GPTs as of January, although reviews of their collective quality and utility have been mixed (typical for any non-curated store).

Still, I was inspired by Wharton professor Ethan Mollick 's work with undergrad and MBA students to create custom GPTs that would (his prompt) "get you a job by showing a potential employer that you are a prompt engineer. It should automate a task in a job you want to do."

The exact "prompt" I gave to my students for their final assignment in our class:

• Using ChatGPT Plus, develop a working HR-focused GPT and share it with the class

• Along with the GPT, provide a write-up that details: 1. The concept/idea; 2. Responsible/ethical AI considerations: governing principles, assessment of risks; 3. Desired end outcomes and measurement of success; 4. Design and development: data and prompting strategy, change management plan, model training

My idea was that—after a semester of learning responsible AI principles and practices and critiquing the work of others (both HR tech vendors and the companies deploying AI-enabled solutions and processes)—the best way to get out of the ivory tower was to make something of their own that was useful and could stand up to the same scrutiny.

In this post, I'll share my key takeaways after reviewing the assignments, as well as showcase a few of the GPTs that students gave me permission to share.

My takeaways

1. Students have good ideas

I kept the ask intentionally broad ("HR-focused" can apply to a lot of things) and was impressed by the ideas they came up with.

In fact, several were things my colleagues and I are exploring right now, including:

• easy navigation to HR topic-specific resources and learning ( Alathoub Alqahtani ),
• improving HR professionals' business acumen ( Ana Gutierrez Barrios ),
• chat support for mental health ( Lorenza Agois, M.S. , Nidhi Gandhi ),
• drawing insights from exit interviews ( Pratiti Tapasvi )

A few of the ideas should really be on HR tech product roadmaps right now, if they're not already, e.g.,

• A GPT by Olivia Li and Bhavana Gowda supports not only performance review creation but also interpretation and identification of specific actions to take as a result—what exercises, training, and experiences should an employee who's received that feedback pursue, given their path and profile? This could help address the "last mile" problem that make performance reviews feel like a painful formality as opposed to something useful today
• Nola Dolan 's GPT scores job requirements on importance to identify opportunities to streamline JDs. Given that requirement bloat is a major contributor to a less inclusive process and pool (turning away prospective candidates that don't check 30 boxes and providing fodder for arbitrary screening), this feels like it would be a natural for Textio .

It's more evidence that the best ideas come from users and practitioners.

2. Generative AI unlocks creation from ideation

Like Professor Mollick's students, most of the Algorithmic Responsibility, Section 100 cohort did not have prior programming experience. They signed up for ChatGPT Plus and followed the process Mollick details here . I won't take up space here repeating the process steps; students largely had to use the tools and guides, and figure it out on their own.

The magical part was that they all successfully brought their HR chatbots to life in a matter of hours. The students indicated they spent the most time hashing out what they wanted the user experience and output to be, along with getting their hands on training data. The actual "programming" was the least labor intensive part, involving natural language prompting in ChatGPT to make refinements.

In a pre-generative AI world, the opposite would likely have been the case, and the need to program in a language would have increased the time and work needed by an order of magnitude... or even more likely, made it impossible for these creations to ever come to life. Unlocking these students' ability to create (instead of just ideate and hypothesize) is a game-changer.

3. The leap from existence to usefulness and productization is a big one

As great as many of these ideas are, and as exciting as it is to see them come to life, another realization for the class was that to make a viable, production-ready solution takes a LOT of work.

As much as I liked what the students produced as class assignments, I won't be petitioning our digital HR team to put them to immediate use today. Whether the audience is the broad employee base or a more narrowly defined group (e.g., the HR team), the use cases described above get pretty nuanced for different personas, and if you aren't addressing them specifically enough, the output won't be very useful. Despite the best efforts to feed instructions and data into the GPT creator, the experience often still feels like vanilla ChatGPT with a skin on top of it. This echoes many of the critiques of Open AI's overall GPT Store offerings.

Additionally, even if these GPTs worked really well as point solutions out of the box, most HR use cases are part of a collective system as opposed to a silo. Unless the experience and output is integrated into existing workflows and platforms, it will be too much of a pain to expect people will use it very much. This drives the advantage that providers like Workday, SAP, and Microsoft have in steering people toward their AI-enabled solutions, integrated strongly into their ecosystems.

That doesn't mean HR professionals would find custom GPTs fruitless. The ability of gen AI to help create (via corporate-friendly versions of this tech like Microsoft's Copilot Visual Studio) more quickly and in a less costly way remains exciting. But beyond exercises to learn (which is valuable in itself), it will take real investment in time and resources to create make things that will be useful and lasting for HR professionals' daily work.

4. Designing responsibly from the start makes AI solutions inherently more ethical and less risky

Students built smart, responsible features and checks into their GPTs. For example,

• A Mood Buddy solution was trained to recognize when the topic was getting into the territory of professional advice and guided the user to consider seeking the help of a medical professional.
• A GPT to draw insights from exit interviews required the user to upload data but recognized when identifiable information was provided and prompted the user to anonymize it first.

Given the course (and assignment rubric) centered on responsible AI, you'd expect my students to lean into ethical design principles and considerations. Regardless, building these in from the start translated to them being a much more organic part of the user experience and flow than if they were just slapped on at the end to check-the-box.

I'm proud of the students for diving headfirst into unfamiliar territory. The courage required to create and then put your work up for scrutiny (every student demo'd their GPT and answered questions in our final class) is great. And the lessons learned from going through it will sink in and stick longer than sitting through some lectures.

Check out some examples of their work below. If you're an HR professional or team reading this, try creating a GPT of your own — the tools to do so are broadly available, whether through ChatGPT Plus or a competing solution.

And if you're interested in learning more about the work the talented current/future HR professionals mentioned above/below have done, feel free to reach out to them directly. Many are graduating from NYU in a few weeks and searching for a full-time job or internship now!

GPT Example 1: ActionAid - a tool to create, interpret, and plan actions from performance reviews ( Olivia Li and Bhavana Gowda )

GPT Example 2: Employee Exits - a tool to draw insights and identify actions based on exit interviews ( Pratiti Tapasvi )

GPT Example 3: EmpowerHR - a tool to enhance HR professionals' business acumen ( Ana Gutierrez Barrios )