Employee Data Alerts

Context

Human Interest is a 401(k) platform serving tens of thousands of small and mid-size businesses. Every day, the platform ingests over 1 million employee records from 50+ payroll providers. That data feeds everything downstream: eligibility determination, enrollment timing, contribution calculations, compliance reporting.

Problem

Data quality was a blind spot. Nobody had a systematic view of what was coming in wrong, how often, or what it was costing the business. Discrepancies in payroll data caused cascading problems: missed enrollment windows, incorrect contributions, compliance exposure for clients, and a growing operational burden for internal teams triaging issues case by case.

The company didn't know the shape of the problem. I started by figuring that out.

Approach

• Diagnosed the problem through data. I analyzed patterns across the full volume of ingested employee records to understand what categories of discrepancies existed, which payroll providers were highest-risk, and where downstream failures concentrated.
• Built a discrepancy taxonomy. Mapped every type of data quality issue into categories by severity and frequency. This gave engineering and operations a shared language for the problem.
• Defined detection logic with engineering. Translated the taxonomy into automated detection rules that could flag discrepancies at ingestion, before they caused downstream failures.
• Launched Employee Data Alerts. Shipped the platform's first automated data quality product. Alerts surfaced issues directly to employer admins in real time, replacing reactive case-by-case triage with proactive, systematic detection.

Outcome

The alerts product cut client compliance penalty exposure and shifted the operations model from reactive to proactive. It also became the foundation for a follow-on self-service correction tool projected to save $100K annually in operational costs.

What I'd do differently

I underestimated payroll provider variability. We treated all 50+ providers roughly the same at launch, applying the same detection logic and alert thresholds across the board. In hindsight, tiering providers by data quality early on would have focused engineering effort on the highest-risk integrations first and reduced noise for employer admins receiving alerts from cleaner providers.