High-Volume Resume Screening Is Broken: Why LLM-Driven Parsing Fails and How to Fix It

A single opening can pull in thousands of resumes, and someone still has to read them. Handing the whole pile to a language model looks like the fix. In my view it swaps a slow problem for a riskier one.

LinkedIn reports about 8,200 job applications submitted every minute (LinkedIn, 2026), and the median non-executive role takes 39 days to fill (SHRM, 2026).

1. Why "send every resume to an LLM" fails

Bias. A University of Washington study found the language models it tested, ranking resumes, favoured white-associated names 85% of the time and female-associated names only 11% (UW, 2024). Embeddings inherit this too, which is why section 4 matters.

Manipulation. In about 200,000 real resumes, roughly 1% hid instructions aimed at AI screeners (Zhang, 2026).

Inconsistency. Ask twice, get two rankings. You can't defend that to a rejected candidate.

Cost isn't the real problem; it's small either way (OpenAI, 2026):

At ₹96 to the dollar. The saving matters less than what the model never gets to decide.

2. Rules first, for the must-haves

Location, notice period, work authorisation, a required licence: these are yes-or-no facts. Check them with plain code. It's instant, explainable and identical for every candidate. Filter only on what the job truly requires.

3. Then match on meaning

Retrieval-augmented generation pairs a search index with a language model (Lewis, 2020). Embed the job description and the remaining resumes, retrieve the closest matches, and let the model write a short summary quoting the resume lines behind each match. The model explains; it never rejects.

4. A recruiter makes every decision

The EU AI Act lists systems that filter job applications as high-risk (EU, 2024). New York requires bias audits of automated hiring tools (NYC, 2023). India's DPDP Act needs a notice and clear consent before you process personal data (DPDP, 2023).

My target is cutting first-pass review time by 80% against your own baseline.

Is your screening ready?

Score each 0 (no), 1 (partly) or 2 (yes):

• Must-haves: written down and checked by code.
• Hidden text: stripped before any model reads a resume.
• Evidence: every match quotes the resume.
• People: a recruiter makes every reject decision.
• Audit: outcomes reviewed for bias each quarter.

Your next step

Time how long your team spends on first-pass screening for one open role this week. That number is your baseline, and the only honest way to prove any tool helped.