Editor's note: Posts published on From the Workplace are written by outside contributors and do not necessarily reflect the views or opinions of SHRM.
When Metropolitan Nashville Public Schools set out to reclassify nearly 1,000 job codes, we expected the hardest part would be getting the AI to align with our HR team on where roles belonged. We were wrong.
The AI performed within the expected range of human judgment, achieving approximately 77% to 79% classification accuracy compared to an 82% inter-rater reliability baseline among experienced HR professionals. The real discovery was not about the AI’s performance. It was what the process revealed: decades of accumulated structural ambiguity hiding in plain sight within our job architecture (a pattern consistent with what SHRM’s State of AI in HR 2026 report describes as the widening gap between AI capability and organizational readiness).
One finding stopped us cold: A single “Specialist” title was held by 435 employees across 24 different pay grades. They all had the same title, but widely varying compensation. That’s not just a data entry issue. It’s a systemic pay equity risk that no manual audit would have spotted.
To address this, we developed PRISM (Progressive Refinement and Intelligence Synthesis Model). PRISM goes beyond being an automation tool. It provides a governance framework. The system evaluates each role classification based on the likelihood of error and the potential cost of error, then organizes roles into three tiers:
- Low-risk: Roles suitable for streamlined review with periodic audit sampling.
- Moderate-risk: Cases requiring targeted spot checks.
- High-risk: Classifications where expert human judgment is essential.
Approximately 40% to 50% of roles fell into the low-risk tier. This allowed our compensation team to concentrate on the 20% to 30% of cases where ambiguity, risk exposure, and financial impact intersect. These are the decisions that truly require human expertise.
This shift fundamentally changed our compensation analysts. Instead of reviewing job descriptions individually, this framework now governs the system itself. They set escalation thresholds, investigate why certain role families consistently trigger high-ambiguity signals, and make the strategic decisions that no algorithm should make independently. This is the kind of human-centered AI implementation SHRM has been advocating for.
For HR leaders considering AI in job classification, the starting point should not be whether the technology is sufficiently accurate. The better question is whether your current job architecture can withstand the level of scrutiny AI will bring. In our case, the AI did not create new problems. It made existing ones visible and measurable for the first time.
The question is no longer whether to use AI in classification. It is whether organizations can afford not to understand what it would reveal.Wayne Birch is a Strategic Compensation and People Analytics professional at Metropolitan Nashville Public Schools.
Was this resource helpful?
Validate your HR expertise
Earning your SHRM-CP credential makes you a recognized expert and leader in the HR field.
Related Content
Learn how Marsh McLennan successfully boosts staff well-being with digital tools, improving productivity and work satisfaction for more than 20,000 employees.
The proliferation of artificial intelligence in the workplace, and the ensuing expected increase in productivity and efficiency, could help usher in the four-day workweek, some experts predict.
As artificial intelligence technology continues to develop, the demand for workers with the ability to work alongside and manage AI systems will increase. This means that workers who are not able to adapt and learn these new skills will be left behind in the job market.