We stopped using resumes as a filter, and we let candidates use AI
This article was contributed by Alon Pearl, Founder and CEO of VA Masters. He founded the firm in 2021 after hiring a virtual assistant remotely to solve his own bandwidth problem, and now writes on hiring at volume, drawing on the 166,350 applications the firm processed between January 2025 and July 2026.
At a glance
- The resume has not become dishonest. It has become uninformative. When anyone can generate a flawless, keyword-optimised application in seconds, everyone looks exceptional on paper.
- Our own rejection data shows it. Two groups with almost identical CV ratings, 1.76 and 1.71, produced completely different results on a practical task, 3.51 against 2.40.
- We allow candidates to use AI during assessments. What we score is whether they catch where the output falls short.
- The cost of this approach lands on completion. Between January 2025 and July 2026 we invited 81,580 candidates to a skills test and 13,846 submitted one
Why the document stopped predicting the work, and what we replaced it with.
The problem is not cheating
Most of the hiring conversation about AI is about catching applicants out. Detection tools, policies, suspicion about who really wrote the cover letter.
I think that is the wrong problem. Banning applicants from using ChatGPT or Jobscan to tailor their materials makes little sense when employers set the standard by using AI to screen them out first.
Using a language model to align a resume with a job description is not dishonest. It is a modern response to an automated gatekeeper.
The real challenge for hiring managers is not that applicants are gaming the system. It is that traditional resumes have lost their signal.
When anyone can generate a flawless, keyword-optimised cover letter in seconds, everyone looks exceptional on paper. The resume no longer proves how well someone can do a job, only that they know how to prompt an AI.
Both sides of the table are now using AI, and most hiring processes are still written as though only one side is.
What our own data says about the document
We can put a number on this from our own rejection decisions. In an analysis of 16,308 recorded rejections between January 2025 and September 2026, two groups stand out when you put their paperwork next to their work.
| Candidates rejected for | Avg CV rating | Avg task score |
|---|---|---|
| Not enough relevant experience | 1.76 | 3.51 |
| Failing the practical task | 1.71 | 2.40 |
Internal five-point assessments, 16,308 recorded rejection decisions, January 2025 to September 2026.
Those CV ratings are effectively the same. The task scores are not. One group could do the work and simply did not have the background on paper. The other could not.
If you are screening a shortlist on background alone, that is the group you are throwing away. Nothing in the document told us which was which.
What we do instead
We no longer rely on resumes or cover letters as a primary filter. Instead we put candidates through practical, real-world task simulations built around the specific role a client has opened, and we actively encourage them to use AI during those evaluations.
What we measure is what matters: their judgment, their critical thinking, and their ability to catch where the AI output falls short. The skill in 2026 is not generating text with AI, it is knowing when the AI is wrong.
We use AI on our side of the table too. With hundreds of thousands of applicants, it helps us narrow a huge pool down to the specific niche a client’s role calls for, so our team can focus its human judgment where it counts.
AI does the first-pass sorting at scale. People make the actual decisions.
That split is the whole design. The machine handles volume, which it is good at. People handle the judgment call, which it is not.
The part nobody mentions: It costs you candidates
| Stage | Figure |
|---|---|
| Invited to a skills test | 81,580 |
| Submitted one | 13,846 |
| Returned it within 24 hours | 55% of those who submitted |
January 2025 to July 2026. The funnel is not linear; candidates enter and leave at different stages.
A real task asks for real work, and a large share of people do not do it. That is the honest cost of this approach and anyone considering it should budget for it.
The ones who do engage move fast. 55% return the test within 24 hours. And the process as a whole does not run slow: the median time from opening a role to a successful hire is 15 days, January 2025 to August 2026, which we cover in our analysis of what businesses actually outsource.
The other cost is internal. Reviewing a real piece of work takes longer than skimming a resume. The AI first pass is what made that workable. It took over the sorting so our people could spend their hours on the judging.
4 things I would tell anyone hiring right now
1. Tell candidates in advance whether AI is allowed.
In the application, in the assessment, or both. An unstated rule is not a standard, it is a trap, and the people most likely to fall into it are the ones taking your instructions literally.
2. If it is allowed, ask what they did with it.
What they accepted, what they changed, what they fact-checked. That question tells you more about someone’s judgment than the finished artefact does.
3. Judge the demonstration rather than the document.
A candidate may use AI to tidy up their wording and still be excellent at the job. Presenting yourself well has never been the same thing as doing the job well.
4. Keep the task close to the actual role.
A generic test measures test-taking. A task drawn from the work the person will actually do measures the work.
One thing this is not
None of this is a claim about AI taking over hiring, or about which jobs are exposed.
Our separate analysis of AI in virtual assistant job requirements found that 23.9% of 322 roles opened between February and August 2026 named AI, and that the categories asking for it most are the ones that already require judgment, at 44.0% for technical and analytics work against 5.3% for finance and accounting.
That measures where employers ask for AI skills, not where AI is capable of doing the work. Those are different questions, and the data answers only the first.
What I am describing is narrower and more practical. The filter moved. It used to be a document and now it is a demonstration, and the reason is that the document stopped telling us anything.
About this data
The figures above describe what passes through a single recruitment pipeline, not hiring as a whole. Every dataset of this kind reflects the business that produced it, and ours is a recruitment pipeline built around a practical skills test, so a breakdown from a process built differently would look different.
Each figure carries its own window because our fields reach complete coverage at different points, and we would rather label them than force them onto one period. Full methodology sits alongside our Filipino virtual assistant workforce dataset.
Independent
