AI Screeners Agree 14% of the Time. Median Ghost: 75 Days.

Two numbers landed within a week of each other in July 2026, and together they end the argument about whether inbound recruiting still works. Greg Savage's industry testing found that running the same AI screening tool twice on identical candidate data produced only 14% overlap in the shortlists. Pin's 2026 Employer Ghosting Index, built on 200,000+ recruiter-candidate exchanges, found that 72% of active pipeline candidates wait 30 days or more for a follow-up, with a median silence of 75 days.

If your ranker is a coin flip and your follow-up is a coma, you do not have a funnel. You have a lottery that the candidates already know is rigged.

The 14% number is worse than it looks

A shortlist overlap of 14% on identical inputs means the AI screener is non-deterministic, not merely biased. Two candidates with the same resume can get opposite outcomes on a Monday versus a Tuesday, and there is no legal defense of that shortlist when an auditor asks how the decision was made.

Most vendor conversations about AI resume screening accuracy focus on bias: does the model favor Ivy League names, penalize career gaps, downweight non-Western surnames. Those are real problems. But 14% consistency is a different class of failure. Bias at least produces a stable, wrong answer you can measure and correct. Non-determinism produces a different wrong answer every run.

Three consequences follow directly:

• You cannot A/B test the screener. If run 2 disagrees with run 1 on 86% of the pool, any downstream metric you compare is measuring noise.
• You cannot defend a rejection. Under EU AI Act enforcement and the growing US state-level ADS statutes, "the model said no" is not a legal answer when the model would have said yes an hour later.
• You cannot improve the model with feedback. Recruiter thumbs-up on a shortlist teaches the system nothing when the shortlist itself is stochastic.

Savage's test priced what vendors marketed as a feature. Adoption of AI resume screening doubled from 26% to 43% of HR teams between 2024 and 2025 per SHRM data, and 44% of organizations now use AI specifically to screen resumes. Most of them have never audited the base-rate accuracy of what they bought.

75 days of silence is not slowness. It is the absence of a decision.

Recruiters are not slow. When a decision gets made, it gets made almost immediately. The problem is that most conversations never reach a decision moment at all.

The Pin index measured 72% of active pipeline candidates going 30+ days without a single logged touchpoint, with median silence at 75 days on stalled prospects. But the same dataset shows that when a conversation does advance to interview, the median elapsed time is 2 days, the 25th percentile is zero days, and the 75th percentile is 9 days.

The pipeline is too wide to review, so it does not get reviewed. Candidates interpret that as ghosting because functionally it is. And the industry's response has been to buy dashboards, set SLAs, and nudge recruiters. None of that addresses the mechanism. You cannot SLA your way out of a queue that grows faster than you can process it.

75

The candidate-side numbers confirm the same story from the other direction. Criteria Corp's 2026 research, reported by Fortune in March, found 53% of job seekers were ghosted by an employer in the past year, a three-year high up from 48% in 2025 and 38% in 2024. Greenhouse's December 2024 State of Job Hunting Report put post-interview ghosting at 61%, up 9 percentage points in a single year. The trend line is not bending.

The math of inbound stopped working at 25:1

Inbound recruiting fails in 2026 because there are structurally too many applicants and too few humans to review them. In Refolk's index of professional profiles, roughly 554,655 US professionals hold a current title of Software Engineer, Senior Software Engineer, or Staff Software Engineer. Against that, roughly 21,806 US professionals work as Technical Recruiters or Sourcers. That is a 25.4:1 engineer-to-technical-recruiter ratio, and it is the ratio before you factor in the Employ Inc. 2026 Hiring Benchmarks Report figure of 257.6 applications per posting, up from 207.2 in 2024.

Here is the full dataset behind the argument:

The 25:1 ratio is not a staffing problem you can hire out of. If every US technical recruiter were cloned tomorrow, the ratio would still be 12:1 against a rising application volume, and the screener would still be non-deterministic. Adding humans to the top of a broken funnel does not fix the funnel; it just distributes the same 14% coin flips across more desks.

The only intervention that changes the math is narrowing the top of the funnel before the applications arrive. That is what outbound sourcing does. Instead of triaging 257 applicants per role, you start with the specific population you actually want.

The Rust example: 1,088 people you will never meet through inbound

The candidates you most want to hire are the ones least likely to apply. In Refolk's index, only 1,088 US Software, Senior, or Staff Engineers currently list Rust as a skill. That is 0.2% of the US software engineer population.

The top employers of that pool tell you everything about why inbound will never surface them:

• Oxide Computer Company
• Shopify
• Figure
• Uniswap
• Helius
• Reclaim.ai
• Kapwing

None of those engineers are trawling job boards. They are employed, they are known to their peers, and they are recruited by name. The shape of a Rust sourcing map looks nothing like the shape of a Rust applicant pool, because there effectively is no Rust applicant pool at any meaningful scale.

This is the exact gap Refolk closes. Instead of hoping a Rust engineer applies to your Greenhouse posting and then hoping your screener ranks them correctly on the run that happens to matter, you describe the person in plain English and get a ranked shortlist across GitHub, LinkedIn, and the open web. The 1,088 people exist. They are indexed. They are addressable. What breaks is the assumption that they will find you.

Why the "AI doom loop" makes inbound signal worse every quarter

Inbound signal quality is degrading, not improving, because candidates are rationally responding to a broken funnel by flooding it harder. Fortune calls this the "AI doom loop": candidates mass-apply using AI tools, employers ghost most applicants, which teaches candidates not to invest in any single application, which makes them mass-apply harder.

The measurable consequences on the employer side:

• 257.6 applications per posting in 2026, up from 207.2 in 2024 (Employ Inc.).
• Roughly 40% of tech candidates believed to have meaningfully inflated their resumes (Deloitte 2026).
• Cold recruiting email reply rates down to 5-7% (Belkins, 2025).
• 76% of recruiters reporting they were ghosted by a candidate in the past year (Ghosting Index 2025).
• Average hiring journey now 68.5 days, with median time to first offer stretching from 57 days in Q1 to 83 days by Q4 (Matchcard).

The marginal application in 2026 carries less signal than the marginal application in 2022, even before your screener runs. The resume is more likely to be AI-written, more likely to be inflated, less likely to reflect a candidate who has actually researched your company. Then you feed that lower-signal input into a 14%-consistent ranker.

When your ranker is a coin flip and your follow-up is a coma, the pipeline is not a funnel. It is a lottery the candidates already know is rigged.

Watch out for survivorship bias in vendor metrics

Most vendor stats about hiring speed describe the survivors, not the majority who never got a decision. When a screening vendor says "positions fill in an average of 14 days" or "83% of candidates get accepted into hiring pipelines," ask what happened to the 72% of active candidates who sat in silence for 75 days.

The Pin data makes the survivorship problem quantifiable for the first time. The 2-day median from conversation to interview is real, and it is a genuinely fast number. But it applies only to the subset that ever got a second look. The base rate, across all active candidates on still-open jobs, is 72% stalled and 75 days silent. Any vendor number that does not disclose its denominator should be treated as marketing, not measurement.

The same critique applies to accuracy claims about AI resume screening. A vendor that reports "92% agreement with recruiter judgment" is almost always measuring agreement on the shortlist the screener already surfaced, not on the 200 resumes the screener buried.

What outbound sourcing actually changes

Outbound flips the problem from "review 257 applicants and hope" to "identify the 40 people who could actually do this job and reach out." When the top-of-funnel is 40 named humans instead of 257 anonymized resumes, the recruiter capacity math works, the screener becomes optional, and the follow-up silence collapses because there is a decision to make on every conversation.

The practical shift for a hiring team looks like this:

1. Start from the population, not the posting. Define the role as a search query over live professional data, not a JD stapled to a job board.
2. Rank by observable signal. GitHub contribution history, LinkedIn tenure at relevant employers, conference talks, and public writing beat resume keywords every time.
3. Reach out at volume that a human can actually follow up on. 40 well-targeted messages will out-hire 2,000 inbound applications, and every response gets a decision inside the 2-day window Pin measured.
4. Skip the screener. If your top-of-funnel is already narrow and signal-rich, the 14%-consistent ranker adds noise, not value.

The reason this works is not that outbound is new. It is that the alternative finally broke publicly, on the same week, in two different datasets. Refolk exists to make outbound cheap enough at the sourcing step that it stops being the specialist skill it was in 2018. You ask in plain English, you get the right people across GitHub, LinkedIn, and the open web, and you spend your recruiter hours on the part that still requires a human: the conversation.

FAQ

How accurate is AI resume screening in 2026?

The most rigorous public test in July 2026, run by industry commentator Greg Savage, found 14% shortlist overlap when the same AI screening tool was run twice on identical candidate data. That is a consistency floor, not a ceiling, and it means the tool is non-deterministic on top of whatever bias it also carries. Vendor accuracy claims typically measure agreement on already-surfaced candidates, not base-rate accuracy across the full applicant pool, so treat any number above 50% with skepticism until you see the denominator.

What is the actual candidate ghosting rate in 2026?

Pin's Employer Ghosting Index, drawing on 200,000+ recruiter-candidate conversations on still-open jobs, found 72% of active pipeline candidates stalled 30+ days with a median silence of 75 days. On the candidate side, Criteria Corp's 2026 research reported by Fortune found 53% of job seekers were ghosted by an employer in the past year, up from 38% in 2024. Both numbers are three-year highs and both are still climbing.

Why is outbound sourcing more reliable than inbound screening?

Outbound narrows the top of the funnel before applications arrive, which makes the recruiter-to-candidate ratio workable and eliminates the need for a non-deterministic screener. In Refolk's index, the US has 25.4 software engineers per technical recruiter, and inbound forces each recruiter to triage 257 applicants per posting; outbound lets you start with a shortlist of 40 named people ranked on observable signal like GitHub history and employer tenure. That is why the 2-day conversation-to-interview median that Pin measured applies only to the small subset that ever gets reviewed, which outbound recreates by default.

What recruiter follow-up time is actually normal?

When conversations reach a decision moment, the Pin index shows median time from positive conversation to interview is 2 days, with the 25th percentile at zero days and the 75th percentile at 9 days. Speed on active decisions is not the problem. The problem is that 72% of active conversations never reach a decision moment at all, and the 75-day silence figure is measuring the queue, not the recruiters.