By Reginal Campbell · August 15, 2026 · 8-minute read

The Four Sentences Nobody Reads

Every applicant tracking system can hold the reason behind a recommendation. Most organizations have never decided who is responsible for making sure it gets there.

Picture an engineer who spent three years sitting next to someone she would hire again tomorrow. She sends the job posting, writes four sentences in the notes field about what that person is actually good at, and submits the referral through the internal portal. Submitting it costs her something. She is spending a little of her own credibility inside the company on a person nobody else there has met.

A recruiter opens the record eleven days later. The notes field is collapsed. The line that shows without clicking reads: Source: Employee Referral.

Something strange happened to the employee referral. We kept the program. We kept the bonus. We kept the field in the applicant tracking system. We kept the quarterly report showing referral volume trending in whatever direction the slide needed. Somewhere in there we stopped keeping the referral itself.

The channel still works, which is what makes the loss expensive. In Ashby’s dataset of roughly 38 million applications across about 93,000 jobs, referred candidates were about one percent of applications and converted to interview at roughly 40 percent, against roughly 3 percent for inbound applicants.[1] That is vendor data from companies sophisticated enough to buy modern recruiting software, so read the levels with caution and the gap with less. The peer-reviewed work explains the mechanism: referrals carry information about a candidate that the application itself does not contain.[2]

Referrals still work, then. But notice what the 40 percent measures. It is the tag doing the work: a referred candidate gets routed out of the inbound pile and in front of a human, which is most of the battle when the pile is three hundred deep. That happens whether or not anyone reads the four sentences. And in the same dataset, the share of referred candidates receiving an offer fell from 12.1 percent to 7.3 percent.[1] The door still opens. What arrives on the other side of it carries less than it used to.

Three things we collapsed into one word

Ask a talent organization how its referral program is performing and you will get a number. That number is aggregating at least three behaviors with almost nothing in common.

The introduction. An employee knows someone who might fit and passes along a name. Useful, low-cost, low-information.

The endorsement. An employee worked with this person, watched them do the job, and is willing to put their own credibility behind a specific claim about a specific capability. This is the version the research is about, and the version carrying real signal.

The attribution. Someone applied through a shared link, or checked a box naming an employee, or got tagged to a source code somewhere between the career site and the interview panel. The employee may not know the candidate. In the referral-link case, the employee may not know the application happened. Ashby’s own documentation distinguishes direct referrals from referral links, and a referral link can be posted publicly.[3]

All three land in the same bucket and roll up into the same metric.

So the metric is doing something worse than leaving information out. It reports several unrelated behaviors under one label and presents the total as a trend.

“Referral performance improved 18 percent” can mean colleagues are vouching for colleagues more often. It can also mean employees got better at sharing links on LinkedIn. The report cannot tell you which, and most executives have never been given a reason to ask. The metric looks precise because it has a percentage attached to it. That does not make the category coherent.

The platforms allow for better. Greenhouse provides default fields for relationship, work history, referrer rating, and referral notes. Oracle permits configurable additional referral information. Ashby and SAP SuccessFactors support structured capture.[4] Those are statements about what the software can do. How many organizations require those fields, and whether anyone downstream ever opens them, is a separate question, and every company running the same platform answers it differently.

The architecture AI inherited

By the time ChatGPT arrived in November 2022, we had already spent nearly two decades removing friction from applying and adding automation to screening.[5] The Department of Labor issued its Internet Applicant rule in October 2005, narrowing who legally counted as an applicant, because online recruiting had made the older definition unworkable at the volumes employers were by then receiving.[6] LinkedIn shipped “Apply with LinkedIn” in 2011 and removed most of what friction remained.[7]

AI did not invent this architecture. It inherited it.

What it inherited was a system already running at a volume nobody had staffed for. Greenhouse, across more than 6,000 companies and more than 640 million applications, shows applications per job rising from roughly 115 in 2022 to 244 in 2025, while the average recruiting team shrank by about 56 percent, to roughly five recruiters per organization.[8] Ashby reports applications per hire roughly tripling between 2021 and 2024 and holding above 300 through 2025.[9]

The per-candidate attention budget collapsed. Fewer recruiters, more applications each, and no additional minute available to read four sentences about why this person is worth twenty.

AI sits on top of that, and the load it adds is not marginal. Ashby finds candidates today are roughly half as likely to receive an interview as they were five years ago, with teams reporting a rise in automated and fraudulent applications that makes genuine candidates harder to identify.[9] That is the difference between a review capacity that was strained and one that has stopped functioning as review. I am not going to claim AI doubled application volume, because the causal decomposition is unresolved and anyone asserting a clean number is guessing. Plenty of recruiting organizations do preserve referral context well. But the defensible claim is sharper than it first sounds: the architecture that made the human handoff optional was built two decades ago, and AI is what made skipping it easy to normalize.

The hidden labor behind a good referral program

Preserving the endorsement takes work that most organizations have never budgeted.

Somebody reads the referrer’s four sentences and decides whether they contain anything job-relevant. Somebody routes them to the person screening, in a form that person will actually open. Somebody tells the referrer what happened. Somebody audits whether referred candidates are advancing on evidence or on familiarity.

GitLab shows the failure is avoidable. Its published process commits to reviewing referrals within five business days, generates weekly reports on outstanding referrals, and applies the same interview bar to referred and non-referred candidates.[10] One employer describing its own process, so treat it as an existence proof rather than a benchmark. What matters is that every commitment in it is specific, assigned, and measured.

Programs that skip this labor keep producing referral volume, which is exactly why the failure is so hard to see. The number holds steady while what it measures quietly empties out.

The software was purchased. The workflow was never funded.

Nobody owns the handoff

The recruiter owns screening. HR owns policy. The hiring manager owns the decision. The applicant tracking system owns the data record. The employee owns the recommendation, right up until they submit it.

Who owns making sure the information survives the trip between them?

In most organizations, nobody. The handoff has no role attached to it, no metric, no line item. It falls in the gap between four functions that each did their job correctly, which is the kind of failure that persists indefinitely because no performance review ever catches it.

Here is the test. If an employee recommends someone this afternoon, can anyone in your organization show you where the reason for that recommendation goes?

If the answer requires three people and a system administrator, you already have your answer.

The engineer finds out too. She spent a piece of her own standing on that candidate, and when nothing comes back, she has learned something about whether her judgment counts here. She will price that into the next request. The research reads as intuitive once you see it that way: referral rejection produces negative reactions in the referrer, perceived fairness buffers the effect, and referrers show meaningfully lower turnover when the person they vouched for gets hired and stays.[11] Closing the loop is the cheapest fix on this list, and the one most often skipped.

Four moves

Separate the referral from the attribution. Someone clicking a shared link and someone vouching for a former colleague are different events with different value. Split the category in the system and measure the subtypes separately, or keep reporting an average that describes nothing. Executive referrals get their own bucket, since endorsements arriving with power attached carry fairness risk the aggregate hides.[12]

Capture the actual recommendation. Two required questions at submission: how do you know this candidate, and what job-relevant capability can you personally attest to. That field is where an endorsement becomes distinguishable from a lead.

Guarantee review without preference. The candidate gets meaningful consideration within a stated service level. The bar stays where it is. A five-person team facing 244 applications per job cannot promise that to everyone tagged as a referral, which is precisely why the first move comes first: once link attributions are out of the bucket, the population owed a guaranteed read is a fraction of what the metric currently reports, and the promise becomes affordable. This distinction is load-bearing, because referral networks reproduce their own demographics and social ties can create evaluative advantages unrelated to performance.[13] A referral is information, not proof.

Close the loop and audit the outcomes. Tell the referrer that review happened. Track who refers, whom they refer, and how those candidates advance. Measure time to review, conversion, and whether employees still trust the process enough to keep participating. Bonuses paid into a system with no review capacity buy volume and little else.[14]

The information exists

Referral tooling is running well ahead of the operating models meant to use it. The relationship, the rationale, and the specific thing one person is willing to vouch for can all be captured, stored, and routed. The software has supported that for years.

Somewhere in your organization this week, someone will spend a little of their credibility writing four sentences about a person they believe in.

The system will reduce those four sentences to a source code.

We decided the human context was optional, then built a system efficient enough to prove it.

Sources

  1. Joel Westmark, “Are referred candidates more likely to get hired?”, Ashby Talent Trends Report, May 16, 2025. Analysis of over 38 million applications across 93,000 jobs, January 2021 to December 2024. Referrals were 1% of applications; 40% of referred candidates advanced from application to interview, against 3% of inbound applicants. The same report shows the referral share of applications falling from 2% in Q1 2021 to under 1% by Q1 2024, and referral offer rates falling from 12.1% to 7.3% between Q3 2021 and Q4 2024. Vendor dataset reflecting Ashby’s customer base, not a representative sample of employers. ↑
  2. Burks, Cowgill, Hoffman, and Housman, “The Value of Hiring through Referrals”, Quarterly Journal of Economics 130(2), 2015, covering nine large firms across call centers, trucking, and high tech; Pallais and Sands, “Why the Referential Treatment? Evidence from Field Experiments on Referrals”, Journal of Political Economy 124(6), 2016; Fernandez, Castilla, and Moore, “Social Capital at Work: Networks and Employment at a Phone Center,” American Journal of Sociology 105(5), 2000; Pinkston, “How Much Do Employers Learn from Referrals?,” Industrial Relations 51(2), 2012. ↑
  3. Ashby, “Referrals and Referral Links”, product documentation. Ashby distinguishes direct referrals from referral links, and states that a referral link can be posted to a network such as LinkedIn, with anyone who applies through it credited to the person who shared it. ↑
  4. Greenhouse, Oracle, Ashby, and SAP SuccessFactors product documentation on referral fields and context capture. Cited as evidence of platform capability, not of how widely those fields are required or used. Ashby’s own documentation notes that beyond name, email, and role, the remaining referral fields are optional and the form is admin-configurable. ↑
  5. OpenAI, “Introducing ChatGPT,” November 30, 2022. See also EEOC materials on online recruiting and screening; the EEOC Uniform Guidelines on Employee Selection Procedures; and EEOC and Department of Justice guidance on algorithms, artificial intelligence, and the Americans with Disabilities Act, 2022. ↑
  6. U.S. Department of Labor, Office of Federal Contract Compliance Programs, “Obligation To Solicit Race and Gender Data for Agency Enforcement Purposes”, 70 Fed. Reg. 58946, final rule published October 7, 2005, effective February 6, 2006. The rule defines “Internet Applicant” and revises recordkeeping requirements to address the use of the internet and electronic data technologies in contractor recruiting and hiring. ↑
  7. “Apply with LinkedIn” launched July 25, 2011 as an embeddable one-click application button. Contemporaneous coverage noted at launch that the button risked inundating recruiters with low-quality applications. ↑
  8. Greenhouse, “The Hire Standard”, benchmark report, March 2026, analyzing over 6,000 companies and over 640 million North American applications from 2022 to 2025. The report gives an 111% increase in applications per job, from roughly 115 in 2022 to 244 in 2025, and a 55.6% decrease in recruiters per organization over the same period, to approximately five in 2025. Vendor dataset reflecting Greenhouse’s customer base. ↑
  9. Ben Perry, “Recruiter Productivity”, Ashby 2026 Talent Trends Report, April 28, 2026. Analysis of over 109 million applications and 247,000 jobs, January 2021 through March 2026. Applications per hire tripled from 2021 to 2024 and remained above 300 throughout 2025, running at 291 in the most recent period against roughly 100 in early 2021. The same report finds the share of applications resulting in an interview fell from 7-8% in 2021 to between 3.6% and 4.7% by Q1 2026, and notes talent teams reporting a rise in AI-generated and fraudulent applications. ↑
  10. GitLab, “Referral Program and Process”, public handbook. Referrals are expected to be reviewed and actioned within five business days of submission; a Weekly Referral Report flags referrals outstanding five or more days; and referral interviews are explicitly held to the same standard as non-referral interviews. GitLab also separates a “referral” from an “endorsement,” excludes candidates who applied through a shared link from referral status, and requires referrers to describe how they know the candidate and why the candidate is qualified. One employer’s self-published process, not an industry benchmark. ↑
  11. Pieper, Trevor, Weller, and Duchon, “Referral Hire Presence Implications for Referrer Turnover and Job Performance”, Journal of Management 45(5), 2019, using data from 265 referrers in a U.S. call center: the presence of a referral hire was negatively related to referrer voluntary turnover and positively related to referrer job performance. On rejection and fairness, see Pieper et al. on referral rejection and referrer reactions; Hausknecht, Day, and Thomas, meta-analysis of applicant reactions to selection procedures, 2004 (86 samples, 48,750 participants); and Truxillo et al., meta-analysis on explanations and fairness perceptions, 2009 (26 samples). ↑
  12. Derfler-Rozin, Baker, and Gino, research on power, executive referrals, and perceived fairness, 2018. ↑
  13. Hederos, Sandberg, Kvissberg, and Polano, “Gender homophily in job referrals: Evidence from a field study among university students”, Labour Economics 92, 2025. In a field study of 453 students at a Swedish business school referring peers for real jobs, 71% of women referred a woman and 75% of men referred a man; 73% of participants overall referred someone of their own gender. Note that participants were students rather than employees inside a corporate referral program. See also Beaman, Keleher, and Magruder on referral gender composition in Malawi, 2018; Rubineau and Fernandez on network recruitment and segregation, 2015 and 2019; Shwed and Kalev on social ties and evaluative advantage, 2014; and Castilla, “Social Networks and Employee Performance in a Call Center,” American Journal of Sociology 110(5), 2005. ↑
  14. Friebel, Heinz, Hoffman, and Zubanov, “What Do Employee Referral Programs Do? Measuring the Direct and Overall Effects of a Management Practice”, Journal of Political Economy 131(3), 2023 (NBER Working Paper 25920, 2019). Employee referral programs were randomly introduced across a grocery chain. Larger referral bonuses increased referral quantity and decreased referral quality. Having a program reduced attrition by roughly 15%, driven mainly by indirect effects on non-referred workers, with the best-supported mechanism being that workers value being involved in hiring. ↑

About the author

Reginal Campbell writes about enterprise technology, AI governance, leadership, and the systems organizations build to make consequential decisions.

Read more articles · Connect with Reginal on LinkedIn