What problem does it solve?
Application volumes have outgrown the staff who read them. Virginia Tech received a record 57,622 first year applications for fall 2025, 10.2 percent more than the year before and the fifth straight year of record highs, and says it needs to keep pace while returning decisions sooner. The Zucker School of Medicine at Hofstra/Northwell gets more than 5,000 applications a year for a class of about 100, and its authors note that the number of medical school applicants has grown by 86% over two decades.
Reading every file twice is slow, and human readers vary: the same rubric is applied with different weight by different people and, over a long season, by the same person. AI can apply one rubric to every file at any hour, which is the reason the schools in the evidence give for adopting it.
The risks follow from the same design. A model trained on past human decisions can copy the bias in them, as the Zucker paper warns, and admission to an educational institution is a listed high risk use in the EU AI Act. Applicants also notice. The University of North Carolina at Chapel Hill faced strong negative feedback after its student newspaper reported that it used AI to evaluate the grammar and writing style of application essays. The schools that publish their design all keep the admit or deny decision with people.
- Virginia Tech reports a record 57,622 first year applications for fall 2025, a 10.2 percent increase over the previous year and the fifth consecutive year of record highs.FAQs: Updates to 2025-26 undergraduate admissions process (2025)
- The authors from the Zucker School of Medicine at Hofstra/Northwell write that the number of medical school applicants has been increasing 86% over the last 2 decades.Using artificial intelligence in medical school admissions screening to decrease inter- and intra-observer variability (2023)
How does it work?
- Staff define the job and the rubric. The admissions office decides which part of the file the AI reads (the essays or the structured fields such as grades and activities) and writes the scoring rubric with anchor examples.
- A model is built and tested on past cycles. Virginia Tech trained its reader on past applicant essays and the scoring rubric. NYU Grossman School of Medicine trained on more than 14,000 applications from 2013 to 2017 with the human screening outcomes, and the Zucker School used five consecutive years of applications.
- The AI scores or sorts the file. At Virginia Tech the AI replaces one of the two human readers of each essay. At the Zucker School it recommends who to invite to interview, who does not meet the baseline criteria and who needs further review, the largest group.
- A disagreement rule brings in a person. Under the old process at Virginia Tech, a gap of more than four points on a 12 point scale between two human readers brought in a third reader. Now a gap of more than two points between the human and the AI brings in a second human reader, so a large difference always gets another human read.
- People decide. Staff review the files the AI marks "yes" or "further review", and final admission decisions stay with admissions professionals. The Zucker model leaves out name, place of birth and photo.
- The team checks the model every cycle. It compares AI and human scores, tests results across applicant groups and decides whether to keep, retrain or switch off the model.
- Audience
- Back office
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | 88% | 1 | 1 organization |
Value drivers: Speed and cycle time, Employee productivity, Lower cost to serve.
Indicative value
A public university with 40,000 first year applications
USD 13,333 to USD 336,000
Reader time cost avoided on essay reading per year
How this is calculated
Formula: applications * essays * minutesPerRead / 60 * replacedShare * readerCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| First year applications per cycle applications, applications | 40,000 | 40,000 | The reference university. |
| Essays per application that the AI reads essays, essays per application | 1 | 4 | Editorial assumption. Virginia Tech's reader scores each of the application essays; a first deployment may start with one. |
| Minutes for one human read of an essay minutesPerRead, minutes per read | 2 | 4 | The VPM news report says a human reader averages two minutes per essay at Virginia Tech. The high end is an editorial assumption for a longer essay. Source |
| Share of second reads that the AI replaces replacedShare, fraction of second reads | 0.5 | 0.9 | Editorial assumption. A second human reader is still needed whenever the AI and the first reader differ by more than the set threshold, and no school in the evidence has published how often that happens. |
| Cost of one reader hour readerCost, USD per hour | 20 | 35 | Editorial assumption for a trained seasonal reader. Replace with your own fully loaded cost. |
What it leaves out: A proxy that prices reader hours. It leaves out the cost of building and validating the model, the fairness testing, staff time to handle disagreements and applicant questions, and the cost of a failed or withdrawn deployment. Virginia Tech's vice provost said at launch that the university is saving at least 8,000 hours, based on the previous year's application pool; that is an estimate, not an audited figure and not capacity measured after a full cycle. The value of earlier decisions to applicants is not counted.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Virginia Tech
United States · Education · 2025
For the 2025 to 2026 undergraduate cycle Virginia Tech replaced one of the two human readers of each application essay with an AI model built by its own researchers, after a record 57,622 first year applications for fall 2025. The model uses a large language model that the university says was trained, tested for accuracy and fairness and validated by its researchers. Each essay is scored by one admissions staff member and the AI reviewer.
Under the old process, a gap of more than four points on a 12 point scale between two human readers brought in a third reader; now a gap of more than two points between the human and the AI brings in a second human reader. The university states that the AI only confirms the human score, that final admission decisions are made exclusively by admissions professionals, and that the aim is earlier decisions without raising application fees.
In a news report published by VPM in December 2025, in the first cycle of use, vice provost Juan Espinoza said that based on last year's application pool the university is saving at least 8,000 hours. That is an estimate made at launch from the previous year's volume, not capacity measured after a cycle, so it is not recorded as a metric.
No outcome disclosed.
Donald and Barbara Zucker School of Medicine at Hofstra/Northwell
United States · Education · 2023
The Zucker School of Medicine receives more than 5,000 applications a year for a class of about 100 students. It uses an AI system as the first reader of every application: the system recommends who should be invited to interview, who does not meet the baseline criteria, and a middle group (the largest) for further review, and human screeners on the admissions committee then review the "yes" and "further review" applications to decide which roughly 800 applicants get interview offers.
The associate dean for admissions told AAMC News that the school took the pool of 5,000 applications down to 1,500 to 2,000 for committee review. The model was trained on five years of applications and, per the school's JAMIA Open paper, reached 88% accuracy on its held out test set.
It works on the structured parts of the application, withholds name, place of birth and photo, and the paper warns that lower performance for Native American applicants means a human should review all of them.
- Accuracy: 88%, offline result on the held out test set of historical applications, before live use
"The algorithm had an accuracy of 95% on the training set, 88% on the validation set, and 88% on the test set."
Claimed by: organization
NYU Grossman School of Medicine
United States · Education · 2025
NYU Grossman School of Medicine built a machine learning model to screen medical school applications and now uses it for initial screening, according to AAMC News. The model was trained on a sample of more than 14,000 applications from 2013 to 2017, together with the outcomes of the human screening for those applications, so it could learn which elements most often led to an interview invitation.
A prospective trial in a later admission cycle compared its output with the human reviews while the AI assessments did not affect any decision. AAMC reports that the pilot showed the AI recommendations repeatedly replicated those of the human reviewers, after which the school moved to AI initial screening. The model assesses the structured parts of the application, and AAMC quotes a study of the NYU program estimating that manual screening "involves more than 6,000 hours of faculty time yearly".
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Several past cycles of applications with the human scores or screening outcomes, kept in a form the model can learn from
- The scoring rubric with anchor examples and a record of how readers were trained
- Applicant group data held apart from the model, used only to test results across groups
- A notice for applicants that says how AI is used in review
Systems to integrate
- Application portal or admissions CRM where essays and application forms arrive
- Reader workflow tool where human scores and notes are recorded
- Identity and access management for admissions staff
- Reporting for the fairness and agreement checks each cycle
Complexity: High
The build needs several past cycles of applications with the human scores attached, a research team able to test agreement and fairness across applicant groups, and sign off from the admissions office, legal counsel and the registrar before the model touches a live file. Virginia Tech reports three years of development before its essay reader went live, and the NYU model ran in a prospective trial where it did not affect decisions.
- 1
Pick the narrowest first job
Start with a second reader on essays, or a sort of structured data that staff review in full, not a decision on who is admitted. All three schools in the evidence keep the decision with staff.
- 2
Replay past cycles before any live use
Run the model on last year's applications with the human scores in hand and measure agreement per rubric trait. The Zucker paper reports 88% accuracy on its held out test set; NYU ran a prospective trial in which the AI assessments did not affect any decision.
- 3
Test results across applicant groups
Compare scores and error rates by group, and set a rule where the model underperforms. The Zucker authors found lower performance for Native American applicants and say a human should review all of those applications.
- 4
Set the disagreement rule and the staffing for it
Decide how large a gap between the AI and the human score sends a file to a second human, and staff for that volume. Virginia Tech uses a gap of more than two points on a 12 point scale.
- 5
Tell applicants and publish the design
Say on the admissions site what the AI reads, what it does not do and who decides. UNC faced a backlash after its student paper reported on AI essay analysis, and its website now states that every application is evaluated by trained human evaluators.
- 6
Monitor every cycle and keep a switch off
Recheck agreement and group results each cycle, log every score, and keep a documented way to turn the model off and return to two human readers.
Guardrails
- The AI never makes the admit or deny decision, and applicants can be told who did
- Direct identifiers such as name, place of birth and photo are kept out of what the model sees
- A second human read whenever the AI and the human score differ by more than the set threshold
- A fairness and agreement report for each cycle before the model is used on live files
- No inference of emotions, personality or health from essays or interviews
KPIs to instrument
- Agreement between AI and human scores per rubric trait, and per applicant group
- Share of files sent to a second human reader, and the reasons
- Reader hours spent per application, against the previous cycle
- Days from application to decision
- Applicant complaints and queries that mention AI
Human in the loop
Admissions professionals own the rubric, the threshold and every decision. They read the files the AI marks "yes" or "further review", handle all disagreements, and review a sample of AI scores each week of the season for scores that look fluent but wrong. A named officer signs off the model before each cycle and can switch it off.
Common failure modes
- Past bias in the training data
- A model that learns from past human decisions can repeat their bias. Test across groups, keep identifiers out and treat a model that merely matches old decisions as a risk, not a success.
- Scoring style instead of substance
- Grammar and writing style can stand in for quality and disadvantage applicants who write in a second language. Check scores against content, use anchor essays and read disputed files by hand.
- Applicant backlash
- Applicants and parents react badly to hidden AI use, as UNC found. Publish what the tool does and who decides.
- Drift between cycles
- Prompts, applicant pools and the way applicants write change from year to year, including with AI help. Retest each cycle and retrain only with approval.
What are the risks and rules?
EU AI Act
High risk
Annex III point 3(a): AI systems intended to be used to determine access or admission or to assign natural persons to educational and vocational training institutions at all levels are high risk. Even where the AI only recommends or scores one part of the file, the system is intended to be used in the admission process. Scoring an applicant's essay is profiling, so the Article 6(3) exemption for narrow procedural or preparatory tasks does not apply. Providers then need risk management, data governance, logging and human oversight, and the university as deployer has the duties in Article 26.
Guidance
- Annex III, high risk AI systems (point 3, education and vocational training) (European Union, Europe). Point 3(a) lists AI used to determine access or admission or to assign people to educational institutions as high risk.
- Article 14, human oversight (European Union, Europe). High risk systems must be designed so that people can effectively oversee them while they are in use.
- Article 26, obligations of deployers of high risk AI systems (European Union, Europe). Sets out what a university that deploys a high risk system must do, including human oversight by trained staff and keeping logs.
- Article 22 GDPR, automated individual decision making, including profiling (European Union, Europe). Gives people the right not to be subject to a decision based solely on automated processing that significantly affects them, which an admission decision does.
- Guidance on AI and data protection (Information Commissioner's Office, Europe). The UK regulator's guidance on fairness, transparency and accountability when AI processes personal data; it is under review because of changes made by the Data (Use and Access) Act.
Controls to put in place
- A conformity and fairness file for the model, with validation results from past cycles
- Applicant notice and a route to ask for a human review of the file
- Immutable log of every AI score, threshold and routing decision
- Named accountable owner and a documented switch off procedure
- A data protection impact assessment covering applicant data and any special categories
When it went wrong elsewhere
- UNC Chapel Hill faces criticism over AI analysis of application essays. After the student newspaper, The Daily Tar Heel, reported in January that the university used AI to evaluate the grammar and writing style of application essays, applicants, parents and students gave strong negative feedback. The university's website says it uses AI programs to provide data points and that every application is evaluated by trained human evaluators.
Frequently asked questions
- Does AI decide who is admitted?
Not at the schools in the evidence. Virginia Tech states that its AI only confirms the score of a human reviewer, that it does not make decisions and that final decisions are made exclusively by admissions professionals. At the Zucker School of Medicine and NYU Grossman School of Medicine the AI screens or recommends and staff review the files.
- How accurate are AI admissions screeners?
The Zucker School of Medicine at Hofstra/Northwell has published figures in JAMIA Open. Its model "had an accuracy of 95% on the training set, 88% on the validation set, and 88% on the test set", measured against who the human process invited to interview.
In a simulated real world trial on the 2020 to 2021 pool of 6,092 applicants, the paper says "When the results of the algorithm and human evaluations were combined, the entire process had an accuracy of 96%." These figures show how closely the model and the combined process match past decisions, which saves time but does not show the decisions were fair.
NYU Grossman School of Medicine has published a validation study of its own model that this page has not reviewed, and Virginia Tech has not published agreement or fairness figures that we found.
- Is AI in admissions high risk under the EU AI Act?
Yes. Annex III point 3(a) lists AI intended to determine access or admission to educational institutions as high risk, so the system needs risk management, human oversight and logging, and the university needs a data protection assessment. The university has the deployer duties in Article 26, and it is also the provider when it builds the system in house, as all three schools in the evidence did.
How to cite this page
Blits.ai AI Use Case Library, "AI support for university admissions application review", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/admissions-application-review-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published