Human Review Closes The Gap Between A Match Score And A Qualified Candidate

Applicant tracking systems now return a percentage next to every candidate. The number arrives fast, looks and feels like a ranking, and is built from whatever the system can read in a document and a job posting. What it can't read is the hiring manager who has already decided a posted requirement is flexible, or the candidate whose last employer used a different title for the same work. A recruiter either catches those cases or the pipeline loses a qualified person without anyone noticing.

Working through that problem daily is Andrea Smith, People and Culture Business Partner at Salal Credit Union, a regulated financial institution in Seattle. Smith spent four years running talent acquisition operations there before moving into her current role, and she is Prosci Change Management certified. She also builds AI governance frameworks for client organizations through her own consultancy, which puts her on both sides of the question of what these tools should be allowed to decide.

"Human oversight can't just mean somebody clicks approve. It also means that after the AI gives you the recommendations, you're fully looking through those recommendations and the person as a whole," Smith says. She handles that review herself before any candidate reaches a hiring manager. She also checks the ranking against what the hiring team asked for in the intake meeting and what the candidate's history shows.

Where the match score falls short

The failure Smith sees most often is a vocabulary problem. Her organization is a credit union, and much of her qualified applicant pool comes from banks, where the same work carries a different job title. A system matching keywords against a posting has no way to know that two titles describe one role.

Posted experience requirements break down the same way. Smith runs intake meetings with hiring managers before candidates are screened, and those conversations regularly move the bar. A manager will say the posted year of accounting experience is negotiable if the candidate already knows the systems the team runs on. Details like that live in a conversation and don't necessarily reach the applicant tracking system.

Smith treats the match percentage as the beginning of the review for that reason. "I take all of those things into consideration, and not just the matching aspect of what that looks like within the applicant system," she notes. "I want to understand what problems this person has solved."

Oversight has to cost something

Smith is direct about why the shortcut is tempting. Reviewing a recommendation properly takes longer than accepting it, and the system gives no indication that accepting it was the wrong call. "Is that the easy thing to do? Absolutely," she says. "But is that the best thing to do? No."

Her framing is that the tool should widen what a recruiter can consider without narrowing who gets considered. She uses AI to draft job descriptions, to source, and to capture and summarize notes while a candidate is still talking, freeing up her attention for the conversation itself. "My approach would be to use AI to augment judgment, not replace it," Smith says. "I still want to keep humans accountable for the decisions that affect people."

The accountability point carries weight in a regulated environment, where a rejected candidate can now question the decision. Regulators and courts expect a selection procedure to be job-related and explainable, and a percentage generated by keyword matching is neither.

Governance in the right department

Smith's broader argument is that most companies have filed AI in the wrong place. Treating it as a technology question puts it with the team that can evaluate the software, but can't evaluate its effect on a candidate's opportunity.

She argues the function needs three owners at once: HR holds the workforce impact, risk and compliance holds the regulatory exposure, and technology holds the integration and the systems themselves. "AI just can't sit at the intersection within the IT department," Smith says. "It basically needs to sit under that HR umbrella, risk and compliance is going to be close, and then the technology piece of AI and integration."

Smith's own organization is working through this now, which she describes as an ongoing conversation with no settled policy yet. Adoption inside the company is uneven, and she estimates that the people using these tools fluently remain a minority of staff.

Candidate disclosure is key

Transparency is where Smith's governance position becomes a practical rule. Her organization doesn't currently permit managers to record or run note-takers on virtual interviews. Where managers struggle to capture answers, she has shown them the assistive tools she uses herself.

Smith's own practice is to ask the candidate for permission first. They are told a note-taker is running and asked whether that is acceptable, which she frames as the minimum an organization owes someone whose application it is processing. "I think just being transparent so the person knows is probably the best ethical use of AI right now," she notes. "Being honest and straightforward is probably going to get us way further than if we didn't say anything."

Smith is aware that plenty of people are skeptical of AI tools, and she treats that as a reason to ask before assuming. The standard she applies extends past hiring to any decision a tool helps make about a person. "The greater the potential impact on a person's opportunity, the greater the organization's responsibility for transparency, human oversight, and accountability," Smith concludes.