For decades, the Metro Nashville schools in Tennessee created district-level jobs organically as the system started programs, responded to evolving student needs, or secured new funding. The result has been an organizational chart that features many employees with similar job titles but markedly different responsibilities and salaries.
This year, as district leaders set out to assess how the system’s jobs and wages stacked up against other systems’ and ensure employees were receiving fair compensation, they quickly learned the project was too large for any one person—or small group—to realistically tackle.
Even if staffers took the project in bite-sized pieces, the 80,000-student, 12,000-employee district’s organizational chart had lost cohesion over time and presented a monumental task to sort through. One prime example discovered early on: 435 employees shared the title “specialist” but were spread among 24 pay grades and often had different job descriptions.
Manually reading each job description, analyzing whether the title and description make sense together, and recategorizing jobs that don’t align with their titles would take years of work—time the district didn’t feel it had to lose, said Wayne Birch, the district’s strategic compensation and people analytics professional.
That’s where artificial intelligence came in.
With the help of staff at Vanderbilt University’s Data Science Institute, Birch and his colleagues developed a custom AI system, called PRISM (Progressive Refinement and Intelligence Synthesis Model). The model has five “team members,” as Birch called them—essentially, five branches of the model that work together to assess, analyze, and flag both patterns and inconsistencies in job descriptions for roles with the same title. The AI tool assessed 992 titles that covered more than 1,000 non-instructional jobs, Birch said.
The models were expected to perform as accurately as a human resources professional with about two years of job-classification experience, Birch said.
The AI “team members” each analyzed the job titles and descriptions and decided whether they needed reclassification and, if so, to what. The models were designed to have “architected friction,” Birch said, meaning they were instructed to disagree with one another if needed rather than simply agree with and affirm the others’ work.
The friction is key, Birch said, because it creates “analytical checkpoints” that force each role to be examined through several lenses before it ever lands on a human’s desk for review. It’s the equivalent of a group of employees with differing expertise and viewpoints offering their takes—and disagreeing along the way to finding the best solution.
Metro Nashville schools’ use of AI offers a look at how a district is using the technology to solve a problem outside the classroom, where much of AI-related research and attention have focused.
And the problem the Nashville district is tackling isn’t unique, Birch said.
Most districts’ HR systems manage the same elements—primarily job classifications, compensation, pay equity, and pay-grade structures—and many districts have a job architecture that evolved organically over time and hasn’t been audited in years, if ever, he said. (Most Metro Nashville employees aren’t unionized, so titles and pay grades aren’t spelled out in contracts.)
“The focus on AI has been on student learning, and that’s understandable, but my argument is that in order to meet those goals, you have to have quality people to both educate the students and support the educators,” Birch said. “Now, this is also a good thing to point to and say, ‘We’re trying to take good care of the funds that we’re given in the district.”
The AI tool reclassified jobs that didn’t have descriptions that aligned with their titles correctly about 80% of the time, Birch said, which is about the same accuracy he would expect from a human doing the same work. For example, if someone is classified as a director but does nonmanagement work, that employee may be overpaid, Birch said.
When the AI tool reviewed higher-profile and higher-paid administrative positions, it flagged those cases for human review.
Now, all the tool’s work—reviewed by Birch’s team—has been handed over to senior leadership to review and decide what happens next.
“The goal of this was to provide the decisionmakers with the quality information they need to make decisions,” Birch said, “and that part has been very successful.”
