Evaluating AI in HR Tech: Building a Unified Evaluation Framework

Recruiters, hiring managers, and candidates rely on AI agents to screen, match, and source talent at scale. The hardest part of this adoption happens post-launch, especially at enterprise scale. A matching model can drift as hiring data changes. A sourcing agent that shines in a demo can struggle with real recruiter requests. These drifts have real-world consequences.  Qualified candidates can be scored unfairly, and recruiters can quickly lose confidence in their shortlists.

Phenom runs AI across the full hiring lifecycle, from intake and sourcing to screening, interviewing, and fraud detection. Most of these applications work in several steps, so an early error can quietly distort the final result. 

At AI Day, the Phenom engineering team shared how one unified evaluation framework keeps every hiring AI application accountable. 

What Is Phenom's Unified AI Evaluation Framework 

Our unified AI evaluation framework is a shared system for measuring how well hiring AI performs. It covers every model and agent, from early development through live production. Product teams don't have to build their own testing from scratch. They use reusable components and tailor them to their own hiring use case. For enterprise HR teams, the framework delivers three benefits.

• Proven improvements: Every change is tested against a benchmark built by HR experts before it's released.

• Early warnings: Live performance is monitored continuously, so quality drops and rising costs surface before users notice.

• Precise fixes: Each step of a workflow is evaluated, so teams can find the exact cause of a problem and fix that part only.

How Phenom's Framework Evaluates Hiring AI

Evaluating hiring AI takes more than a single test. It needs a trusted standard to measure against, a way to check quality at every level, and coverage both before and after launch. Our framework brings these together through three working parts, explained in the following sections.

A Benchmark Built by HR Experts

At its center is a golden dataset. It holds examples validated by HR domain experts across core recruiting tasks, including screening, sourcing, resume parsing, and job matching. Because experienced recruiting specialists review these examples, the standard reflects real hiring judgment rather than generic AI tests.

The dataset is versioned, so every change is measured against the same baseline. It's never used to train models. Its only job is to act as a quality gate. When a new failure shows up in live hiring workflows, the team labels it and adds it to the dataset. Over time, the benchmark grows stronger.

Four Levels of Quality

Every evaluation looks at quality from four angles.

1. Component: Did a single model, tool, or prompt get it right? For example, did the resume parser extract the right skills?

2. Trace: Did the sequence of steps and decisions make sense?

3. End-to-end workflow: Was the hiring task completed, such as delivering a relevant shortlist?

4. Production behavior: How are quality, speed, and cost trending as usage grows?

When a workflow score drops, the team can trace it back to the component that caused it.

Checking Before and After Launch

• Before launch (offline flow): The team tests the model or agent against the golden dataset and records every output. Since the correct answers are known, scoring is precise and automatic. Every prompt change or model update must pass this gate before release.

• After launch (online flow): An LLM judge, a separate AI model that grades responses and scores live interactions around the clock. When it's unsure about an answer, it sends the case to HR domain experts for review. This helps the team spot shifts in hiring data as they happen.

A modular pipeline sits behind both flows. It converts each application's data into one standard format, then runs shared metrics alongside use-case-specific ones. New models and agents can plug in without changes to the core system.

Fixing Underscored Candidates at the Source

The Phenom Fit Score helps recruiters grade candidates by measuring how well a candidate's profile matches a specific job. Behind it sit three models: a resume parser, a job parser, and the fit score model itself. If a match looks wrong, any of the three could be the cause. That's why the framework evaluates the full workflow, not just the final score.

In one example, HR domain experts noticed a pattern. Candidates in some specialized fields kept receiving lower scores than they deserved. At first, the fit score model seemed to be underscoring them. The trace data told a different story.

The team used an LLM judge to check the resume parser's output against the original resumes. Broken down by industry, the results showed the parser dropping key skills such as culinary arts, public speaking, and leadership. Those skills were on the resumes but never reached the fit score. The scoring model was working correctly. It simply had incomplete information.

The issue traced back to how the parser's skills section had learned from a larger teacher model. The team switched that one section from Gemini to ChatGPT and kept Gemini everywhere else for efficiency. They also added more industry-specific resumes to the golden dataset, along with new checks to catch similar gaps early.

Keeping the Sourcing Agent on Track

Phenom  Sourcing Agent helps recruiters fill open roles. It turns a search request into a ranked shortlist through a series of steps. These include understanding the request, building queries, splitting the role into search segments, fetching and ranking candidates, and planning outreach. The framework tests each step on its own, with dedicated test cases and metrics. Every run is logged end to end, and an LLM judge scores the steps that need recruiter-level judgment.

Two issues show why that matters.

1. Search segments that were too broad: The agent splits a request like "ICU nurse" into narrower segments, such as cardiac ICU or neonatal ICU. Step-level scores showed these segments were too broad, which hurt retrieval. The problem was missing context, not the model itself. The team added intake details and the recruiter's past hiring data to the segment step. Retrieval improved right away, and recruiters saw more relevant candidates.

2. Requests that changed mid-conversation: Recruiters often adjust criteria as they go. One might ask for ACLS-certified nurses in Austin, add five or more years of experience, then switch to Dallas. The traces showed the agent kept both cities and searched Austin and Dallas together. Once the issue was visible, the team quickly fixed how the agent updates each query.

How Consistent Evaluation Strengthens Every Hiring AI Product

The same framework evaluates every AI application across our applied AI. Each one is measured on what matters most for its role.

• Intake Agent: Guides hiring managers through a structured intake conversation and turns it into a job description, sourcing criteria, screening questions, and an interview plan. Evaluation checks whether job requirements are captured completely and correctly, so every artifact that follows starts from the right brief.

• Screening Agent: Assesses candidates against a role’s requirements and proposes an evaluation score for recruiters. Evaluation checks matching accuracy, evidence extraction, and proposed evaluation score correctness, so recruiters can rely on the scores they review.

• Voice Screening: Calls candidates and runs structured screening conversations, then gives recruiters a summary and score. Evaluation checks dialogue progression, question selection, response handling, and screening criteria, so every call stays on track and screens for what matters.

• AI Interview Agent: Conducts AI-assisted interviews and assesses candidates against the competencies a role requires. Evaluation checks interview progression, response interpretation, and the quality of candidate assessment, so interview results stay consistent and fair.

• Fraud Agent: Flags potentially fraudulent candidates by analyzing security signals and gathering supporting evidence. Evaluation checks the accuracy of its security signals, evidence gathering, and final decisions, so genuine candidates aren’t wrongly flagged, and real fraud doesn’t slip through.

• Foundational models: The Resume Parser structures candidate profiles, the Job Parser structures job descriptions, and the Fit Score measures how well the two match. Because the other products depend on them, these models are evaluated continuously, so an error here never quietly spreads downstream.

For enterprise hiring, this changes how AI is built and maintained. The old approach was reactive. Teams built, deployed, and waited for problems to appear. Phenom's framework replaces that with a continuous cycle of testing, monitoring, diagnosing, and fixing.

Each group benefits. Developers can compare versions and trace failures to specific components. Product teams see performance from a single step to the full workflow. HR and technology leaders get one view of AI quality across the hiring process.

As AI agents take on more hiring work, this discipline matters even more. It's what separates AI that impresses in a demo from AI that recruiters and candidates can trust every day.

Measured at every step. Trusted in production. Book a demo to see how Phenom evaluates and improves the AI behind your hiring workflows.