If your must-haves only show up in a skills section, the screeners launching at HR Tech 2026 will probably mark them "unproven." The question they ask about your resume used to be "does it contain the keyword?" Now it's "does it prove each requirement a human approved?"
Several vendors are launching tools built this way at the show in Las Vegas (Oct 20 to 22). This is how they read a resume, and how to fix yours before you hit apply.
What do the new HR Tech 2026 screeners actually check?
They check each requirement in a posting against evidence in your resume, then sort it as demonstrated, unproven, or needing human review. Whether the keyword appears is no longer the whole test.
Three launches at Mandalay Bay show the shift:
- Defensible Zone, DZ Verify (Booth 2211). The company says it "evaluates job requirements, résumé evidence and candidate responses to show what has been demonstrated, what remains unproven and where human review is needed."
- HiringNode (Booth SU81). It calls the product "governed AI candidate screening." It runs on top of an employer's existing ATS. A human has to approve locked job criteria before any AI screening starts, and every decision goes into a tamper-evident record the employer owns.
- InLeika. It uses OpenAI-powered "Talent Intelligence" to put numbers on skills and score candidates against job requirements.
Criteria-locked screening means a human signs off on a fixed list of requirements before the AI looks at any candidate, and the AI screens against that list. Requirement-to-evidence matching means the tool judges each requirement by whether specific text in your resume supports it, not by how often the word shows up.
Greenhouse CEO Daniel Chait put the old approach bluntly: the first wave of ATS AI tooling "saw the job seekers as the problem." This new wave treats the employer's decision as the thing that needs defending, and that changes what your resume has to do.
Why doesn't a skills section count as proof anymore?
Because almost everyone's looks the same. A skills list tells a screener you know a word. It doesn't say where you used it, what you did with it, or what changed because of it.
In Refolk's index of professional profiles, 15,074 U.S. professionals titled Data Analyst list SQL, Python and Tableau, all three. Titles don't separate people much either. In a 25-profile sample, 19 (76%) held exactly the title "Data Analyst."
From Refolk's index. When everyone lists the same three skills, the list stops telling anyone apart.
Think about how a tool like DZ Verify is described. If it compares each requirement to evidence, a skills section that copies the posting gives it almost nothing to work with. What moves a requirement from "unproven" to "demonstrated" is probably a bullet under a real role that names the skill, what you did with it and the result.
To be clear, that's an inference from how the vendors describe their products. None of them has published its scoring internals. The logic still holds: if the output is labeled "demonstrated," something on the page has to do the demonstrating.
How does a criteria-locked screener read your resume?
Literally, one criterion at a time, looking for text it can point to. With HiringNode, every decision is written to an auditable record, so a pass has to hold up if someone checks it later.
That changes which wording works. Loose phrasing that sort of matches a criterion ("data-driven decision maker" against a requirement for "forecasting models in Python") is hard to justify once a human has locked the criteria and the decision is on record. The safe bet for the employer is a line that answers the criterion in its own words. And the person who approved those criteria knows the system: in Refolk's index, 6,971 U.S. recruiters and talent acquisition professionals list Applicant Tracking Systems as a skill.
A line that counts as evidence usually has four parts:
- The skill or tool, in the posting's words. If it says "Tableau," write Tableau, not "BI tools."
- A specific role and dates around it. Evidence floats if it isn't attached to a job.
- What you did. Built, migrated, owned, cut, shipped.
- Result or scope. Who used it, how big it was, what changed.
A skills list says you know the word. One bullet under a dated role says you used it and it mattered.
Why are employers switching to evidence screening now?
Two reasons: application volume and the law. Keyword matching is easy to game with AI. Evidence checks are harder to fake and easier to defend in an audit.
Volume first. Applications per recruiter on Greenhouse's platform are up 412% since 2023. Greenhouse's Oct 15 research says four in five job seekers use AI in their search, and 86% of U.S. recruiters say at least half the applications they get are AI-written. Once every applicant can copy a posting's keywords into a resume in seconds, keyword overlap means very little. Evidence screening is the employer's answer, and mass-tailoring keywords is exactly the habit these tools are built to beat.
That flood is why keyword matching broke and screeners moved to checking evidence.
Watch out for the funnel numbers, because they don't measure the same thing:
| Source | What it measures | Figure |
|---|---|---|
| Greenhouse (Oct 2026) | Applicants who ever reach an interview | Under 7% |
| Ashby (via secondary roundup) | Applications that reach an interview | 3.6 to 4.7%, about half the 2021 rate |
| Gem (via secondary roundup) | Applicants who pass initial screening | 8% |
One analyst points out that the Greenhouse figure is vendor-sponsored and doesn't define the role, time period or applicant cohort it covers. The Ashby and Gem numbers come from a secondary roundup and measure different stages. The fair reading is a range: somewhere between a few percent and under a tenth of applicants get past the early stages. Either way, most applications die early.
Then the law. Colorado's SB 26-189 replaces the state's original AI Act and takes effect Jan 1, 2027. It requires:
- Notice before AI is used
- A 30-day notice after an adverse action
- Human review
- Three years of recordkeeping
That's why "defensible" keeps showing up. Candidates are pushing back too. In Greenhouse's May 2026 survey, 70% of U.S. candidates who were evaluated by AI said it wasn't clearly disclosed beforehand, and 38% had dropped out of a process because of an AI interview.
Is "unproven" the same as rejected?
No. An unproven requirement can send you to a human instead of an automatic no. DZ Verify explicitly flags "where human review is needed," and Colorado's new law requires human review.
That gives you room to deal with a gap honestly instead of hiding it. A screener that hits a requirement with no evidence and no explanation has nothing to pass along. A screener that finds adjacent evidence or a direct note has something a reviewer can weigh.
Ways to handle a real gap:
- One honest adjacent-evidence line. No Python forecasting, but you built forecasts in SQL that did the same job? Say exactly that.
- One sentence in the cover letter. Name the requirement, say where you stand, and say how you're closing the gap.
- Don't fake it. HiringNode writes every decision to a tamper-evident record. A made-up claim that passes the screen now sits in a file someone can open after a bad interview.
How do you map every must-have to one resume line before applying?
Copy the posting's must-haves word for word into a checklist, then write one line under a real role that proves each one. Wherever you can't, choose between adjacent evidence and an honest note about the gap. Do this before you apply.
This is how to match a resume to a job description the way an evidence screener will read it:
- Paste the must-haves verbatim. Split required from preferred. The required list is the most likely source of locked criteria.
- Run your own audit. Next to each requirement, paste the resume line that proves it. If you can't find one, mark it unproven. That's what the screener will see.
- Rewrite weak lines into evidence. Posting's words, plus action, plus result or scope, all under a dated role.
- Move proof up. Put the strongest evidence in the first two bullets of your most recent relevant role, where it's hardest to miss.
- Handle the gaps. One adjacent-evidence line or one cover-letter sentence per unproven item.
- Keep the skills section, but don't lean on it. It can still help you get found. It just doesn't count as proof.
- Check the ratio. One bullet can honestly prove two requirements. It can't prove five.
Here's what the rewrite looks like in practice:
| Requirement | Unproven version | Demonstrated version |
|---|---|---|
| Tableau dashboards for leadership | Skills: Tableau | Built the Tableau revenue dashboard the CFO's staff reviewed weekly, replacing manual spreadsheet reports |
| Python for forecasting | Data-driven problem solver | Wrote the Python demand forecast the ops team used to set monthly inventory orders |
| Stakeholder management | Strong communicator | Ran the weekly metrics review with sales and finance leads and owned the definitions doc both teams signed off on |
Steps two and three are the slow part, and they're where most people start cutting corners. It's the exact work Refolk takes off you: paste the posting, and I rewrite your own resume against its requirements, using only things you actually did, then score how well you really fit, requirement by requirement. You see the unproven items before an employer's screener does.
For the gaps, the cover letter is where a single honest sentence goes. Refolk drafts it from the same history, so the gap note matches your resume and doesn't contradict it, which matters when every decision is logged.
FAQ
Do AI resume screeners still use keywords in 2026?
Yes, but keywords are no longer enough. DZ Verify checks job requirements against résumé evidence, and HiringNode screens against criteria a human locked in advance. The keyword can still get you found, but a bullet under a dated role is what gets a requirement marked demonstrated. Use the posting's exact terms inside your experience bullets, not only in a skills list.
Should I delete my skills section?
No. Keep it short and accurate, because a recruiter may still search the ATS by keyword. Just don't count it as proof. In Refolk's index, 15,074 U.S. data analysts list the same SQL, Python and Tableau set, so the section only tells you apart once each of those skills also shows up in a bullet with a result.
What if I genuinely don't meet a must-have requirement?
Say so in one line instead of stretching another bullet to cover it. Write an adjacent-evidence line in the resume or one sentence in the cover letter. DZ Verify flags "where human review is needed," and Colorado's SB 26-189 requires human review from Jan 1, 2027, so a gap you've handled honestly can reach a person. A made-up claim sits in an auditable record.
Is it true that fewer than 7% of applicants get an interview?
It's one vendor's figure, from Greenhouse's Oct 15 research, and it doesn't define the role, period or cohort it covers. Ashby data puts applications reaching an interview at 3.6 to 4.7%, and Gem found 8% get past initial screening, but those measure different stages and come from a secondary source. Treat it as a range: most applications die early, which is why proving each requirement on the page matters.