We hired 200 people last year.

This year we've hired 3.

What changed? We implemented an AI resume screener to "improve candidate quality."

The promise: "The AI learns from your best hires and automatically filters out unqualified candidates. No more wasting time on bad resumes."

What actually happened: The AI rejected 100% of applicants for three months straight, including internal transfers, employee referrals, and our CEO who applied as a test.

How We Got Here

Our VP of People Operations attended a conference.

Conference speaker: "Manual resume screening is dead. AI can screen thousands of resumes in seconds and only surface the best candidates."

Our VP: "We get 500 applications per role. This will save us hours!"

We bought an AI resume screening tool.

The setup:

1. Upload resumes of your top performers
2. The AI learns what "good" looks like
3. The AI automatically rejects anyone who doesn't match the pattern
4. You only see the "best" candidates

Simple. Data-driven. The future of recruiting.

Week One: Zero Candidates Passed

We posted a Software Engineer role.

147 people applied.

The AI screened all 147 resumes.

Results: 0 candidates passed the initial screen.

Me: "That can't be right."

I manually reviewed 20 random resumes.

What I found:

• Senior engineers with 10+ years experience at top companies
• Candidates with the exact tech stack we needed
• People with Computer Science degrees from MIT, Stanford, Berkeley

All rejected by the AI.

Me to the AI vendor: "Your tool rejected everyone. Including people who are clearly qualified."

Vendor: "The AI is working correctly. Those candidates don't match your success profile."

Me: "What's wrong with them?"

Vendor: "The AI doesn't provide explanations. It just flags candidates who don't match patterns from your top performers."

Me: "So we have no idea why it rejected them?"

Vendor: "Correct. But trust the data. The AI knows."

Week Two: We Lower The Standards (It Gets Worse)

I called the vendor again.

Me: "We've received 200+ applications across 4 roles. The AI has passed exactly ZERO candidates. This can't be right."

Vendor: "Let me check your settings... ah, I see the issue. Your 'match threshold' is set to 95%. That's very high."

Me: "What should it be?"

Vendor: "Most clients use 70-75%."

Me: "Okay, change it to 75%."

Vendor: "Done."

We waited.

Results after lowering threshold to 75%:

• Still 0 candidates passed

I called back.

Me: "We lowered the threshold. Still zero candidates."

Vendor: "Interesting. Let me run diagnostics."

10 minutes of hold music

Vendor: "I found the problem."

Me: "What is it?"

Vendor: "Your training data was too homogenous."

Me: "What does that mean?"

Vendor: "The AI learned VERY specific patterns from your top performers. If a candidate deviates even slightly, they get rejected."

Me: "How do we fix it?"

Vendor: "You need more diverse training data."

Me: "We uploaded resumes from our 20 best engineers."

Vendor: "Right. That's the problem. They're all too similar. The AI learned to only accept people who look exactly like them."

Me: "So our AI resume screener... only accepts clones of our existing team?"

Vendor: "Essentially, yes."

Week Three: The CEO Test

Our CEO heard about the zero-candidates problem.

CEO: "Let me test it."

He created a fake resume. Used his real work history, skills, and education. Changed his name.

He applied to one of our open roles.

Result: Rejected by the AI.

CEO: "The AI rejected me. And I run this company."

Me: "I know. It's rejected everyone."

CEO: "What is it looking for?"

Me: "We don't know. The AI doesn't explain its decisions."

CEO: "This is insane."

We called an emergency meeting.

The AI's "Logic" (If You Can Call It That)

We finally got the vendor to pull diagnostic reports.

Why the AI was rejecting everyone:

Reason #1: Font choice

Our top performers' resumes happened to be in Arial font. The AI learned that "good resumes use Arial."

Anyone who submitted a resume in Calibri, Times New Roman, or any other font got auto-rejected.

Reason #2: File naming

Most of our top performers named their resumes "FirstName_LastName_Resume.pdf."

The AI learned this was the "correct" format.

Candidates who submitted files named "Resume.pdf" or "John_Smith_Software_Engineer.pdf" got rejected for "non-standard file naming."

Reason #3: Exact job titles

Our top performers had job titles like "Software Engineer II" or "Senior Software Engineer."

Candidates with titles like "Software Developer" or "Full Stack Engineer" got rejected because the AI didn't recognize those as equivalent.

Reason #4: Resume length

Our top performers' resumes averaged 1.8 pages.

The AI rejected anyone with a 1-page resume (too short) or a 2+ page resume (too long).

Reason #5: Education section placement

Most of our top performers listed education at the bottom of their resumes.

Candidates who listed education at the top got flagged as "junior" and rejected.

The AI wasn't evaluating skills, experience, or qualifications.

It was screening for formatting quirks.

The Internal Transfer Disaster

Meanwhile, one of our product managers applied for an internal transfer to a different team.

She'd been with the company for 3 years. Consistently high performer. Great reviews.

The AI rejected her.

HR: "Why was she rejected?"

Me: "Because her internal application resume was formatted differently than the training data."

HR: "She works here. She's already been hired and vetted."

Me: "The AI doesn't know that."

HR: "Can we manually override it?"

Me: "Technically yes, but then what's the point of the AI?"

HR: "Good question."

Week Four: The Referral Incident

Our engineering manager referred a former colleague.

Manager: "This person is brilliant. We worked together for 5 years. I want to hire them immediately."

The candidate applied.

AI result: Rejected.

Manager: "What? Why?"

Me: "The AI flagged them as unqualified."

Manager: "They have 12 years of experience and built three products from scratch."

Me: "Their resume was in Calibri font."

Manager: "WHAT?"

Me: "The AI learned that good candidates use Arial."

Manager: "This is the dumbest thing I've ever heard."

Me: "I agree."

Manager: "Can you manually review them?"

Me: "Yes."

Manager: "Then what is the AI even doing?"

Me: "Excellent question."

We Disabled The AI

After 3 months of zero hires, we turned off the AI resume screener.

Final stats:

• 1,247 applications received
• 0 candidates passed AI screening
• 100% rejection rate

When we manually reviewed the rejected candidates:

• 43 were highly qualified and got interviews
• 12 received offers
• 8 accepted

We hired 8 people from the pool the AI rejected.

The Vendor's Response

Me: "Your AI rejected everyone, including our CEO. We're canceling our contract."

Vendor: "I'm sorry you didn't have a good experience. AI resume screening requires careful tuning."

Me: "We spent three months tuning it. It never worked."

Vendor: "Perhaps your training data wasn't representative."

Me: "We used our best performers."

Vendor: "Right, but AI needs diverse data. If your top performers all have similar backgrounds, the AI will only accept similar candidates."

Me: "So your product is designed to reinforce hiring bias?"

Vendor: "That's not how I'd phrase it."

Me: "How would you phrase it?"

Vendor: "The AI learns from patterns in your data. If those patterns aren't diverse, the outputs won't be diverse."

Me: "So you sold us a tool that's guaranteed to reduce candidate diversity?"

Vendor: "...we recommend clients use it as one input, not the sole decision-maker."

Me: "You literally marketed it as 'automatic candidate filtering.'"

Vendor: "Right. But it should be used thoughtfully."

Me: "We're done here."

What We Learned

AI resume screening sounds great in theory:

• Save time
• Reduce bias
• Surface the best candidates

In practice, it:

• Rejected everyone
• Screened for formatting quirks instead of qualifications
• Reinforced bias by only accepting candidates who looked exactly like existing employees
• Provided zero transparency into why candidates were rejected

We're back to manual resume review.

Yes, it takes longer.

But at least we're hiring people.

The Bottom Line

If you're considering AI resume screening:

1. Test it extensively before going live (we didn't)
2. Manually review rejected candidates (we eventually did)
3. Ensure the AI explains its decisions (ours didn't)
4. Use diverse training data (we didn't)
5. Don't trust "the AI knows best" (it doesn't)

Or just stick with humans.

We read resumes slower than AI.

But we don't reject the CEO.