ATS With AI Screening in 2026: Native vs Plug-In, and What Each Shows the Recruiter
ATS with AI screening in two tiers, native vs plug-in: what is scored, the explanation shown, who decides, public bias audits and EU AI Act status for 13 tools.
An ATS with AI screening comes in two shapes, and most "best of" lists mix them up. Native screening runs inside the system on every application (Workable, Ashby, Greenhouse, Manatal, Breezy, Personio, Pickr); a plug-in layer runs beside it and scores something of its own, an interview, a chat, an assessment or a match, then writes the result back (HireVue, Paradox, CodeSignal, Peoplebox, hireEZ, Eightfold). The questions that decide which one you can defend are the same for both: what is scored, what the recruiter sees, whether a human still decides, whether the bias documentation is public, and whether the vendor acknowledges the EU AI Act. The two tables below answer those five questions for thirteen tools, as of 14 October 2026, one source per row.
I built Pickr, so Pickr is one row in the native tier; discount my conclusions accordingly. I also run ScalingPPL, a recruitment agency, and most of what follows comes from sitting through vendor demos where the score was shown and the reasoning was not.
Why do "best ATS with AI screening" lists mix two different products?
Because the query is commercial and the lists are written by vendors. A standalone screening tool that connects to your ATS wants to be in the same article as the ATS that already screens, so both end up in one ranked list with columns for "best for" and "starting price". None of the lists I found carried a column for explainability, human decision, public bias audit or AI Act status, and several repeated speed claims without a source.
The distinction matters in practice. A native screener adds no vendor, no second data flow and no second data processing agreement; its weakness is that you get the ATS vendor's screening whether or not it is good. A plug-in layer can be the stronger choice for one thing (a structured video interview, a coding assessment, a conversational knockout) at the cost of a second system of record for the candidate's data. From our work at ScalingPPL, the second system is where the audit trail breaks: the ATS says "rejected", the plug-in holds the reason, and a year later nobody can join the two.
What is actually scored, and by which tier?
Native screening scores the application against criteria the recruiter sets. Workable's Agent measures each candidate against an "Ideal Candidate Profile" and returns a score out of 100 with a status per criterion (Workable help centre, accessed 2026-10-14). Ashby returns no composite score: each criterion gets "meets, does not meet, or the AI was undecided", and a percentage of criteria met can sort the queue (Ashby docs, accessed 2026-10-14). Greenhouse's Talent Matching sorts applications "into tiers based on risk and relevance" (Greenhouse bias audit statement, accessed 2026-10-14). Breezy's Applicant Insights gives "every candidate a 0-10 score with a written explanation" against up to five categories you choose (Breezy Intelligence, accessed 2026-10-14).
Plug-in layers score something the ATS never sees. HireVue scores structured interviews, game-based assessments and job-relevant skills tests (HireVue, accessed 2026-10-14). Paradox's Olivia screens "each candidate against job requirements" in a text conversation (Paradox screening, accessed 2026-10-14). CodeSignal scores work in a real IDE and returns "a report showing proficiency by skill area" (CodeSignal skills assessments, accessed 2026-10-14). Eightfold's Matching Model produces a 0-to-5-star match (Eightfold bias audit results, accessed 2026-10-14). The useful question in a demo is not "do you use AI" but "which artefact goes in, and which number comes out".
The two-tier comparison table, as of 14 October 2026
Every cell comes from the vendor's own page opened on 14 October 2026. "Not stated" means the pages I opened do not say it, not that the vendor has no answer. Prices are deliberately absent; the column says whether a price is published.
Tier 1: screening native in the ATS
| Tool | What is scored | Explanation shown to the recruiter | Human decision required | Bias documentation public | EU AI Act high-risk acknowledged | Price published? | Source |
|---|---|---|---|---|---|---|---|
| Workable (Workable Agent) | Each applicant against an Ideal Candidate Profile: must-haves, nice-to-haves, dealbreakers; score out of 100; per criterion met, partially met, not met or unknown | Yes: "the reasoning for every ICP criterion"; every score and action logged | Yes: "every decision is yours to override"; auto-disqualify is off by default; knockout-question fails auto-disqualify | Not on the pages opened; states it "excludes protected characteristics by design" | Not stated | Yes, headcount bands; Agent runs on credits | Pricing, Help centre |
| Ashby (AI-Assisted Application Review) | Each application against criteria with a title and a prompt; a status per criterion (meets, does not meet, undecided) instead of a composite score; a sortable "criteria met" percentage can rank the queue | Yes: "Click on any criterion to see the reasoning" | Not stated as a rule: the AI "returns an evaluation status"; no automatic advance or reject is described; human ratings are calculated independently | Not stated; criteria prompts are scanned for "implicit or explicit EEO violations" | Not stated; offers an "Automated Processing Legal Notice" and a candidate opt-out | Foundations published, higher plans by quote; AI credits per plan | Pricing, Docs |
| Greenhouse (Talent Matching, Real Talent) | Applications sorted "into tiers based on risk and relevance" | Discrete categories with explanations rather than one composite score: "If AI can't explain itself, it doesn't belong in hiring" | Yes: "does not have the ability to automatically advance or reject a candidate"; candidates can request manual review | Yes: monthly independent audits by Warden AI across ten protected classes, results on a public dashboard; every release must pass before going live | Not stated; cites ISO 42001 | No, quote only | Pricing, Bias audit statement, AI principles |
| Manatal (AI Recommendations) | "Automatic AI based candidates scoring" against the job description; candidate fit and gap analysis; customisable weighting | Yes: "detailed AI summaries and see exactly how candidates match your requirements" | Not stated | Not stated; the AI Interviewer add-on claims "unbiased" assessments without a method | Not stated | Yes, per plan; Professional and above | Pricing |
| Breezy (Breezy Intelligence, add-on) | Applicant Insights: resume and application against the job description and up to five categories you choose; 0-10 score; Resume Audit flags AI-written or copied CVs | Yes: "a 0-10 score with a written explanation"; evidence behind every risk flag | Not stated on the pages opened | Yes: Warden AI audits Applicant Insights monthly for sex, race/ethnicity and intersectional bias; NYC Local Law 144 audit completed | Not stated | Yes, per plan; Intelligence runs on credits | Breezy Intelligence, Pricing |
| Personio (AI Talent Screening Agent) | "Evaluate every application against transparent criteria you set"; AI signals highlight best-fit candidates | Partial: transparent criteria and "AI signals"; depth of the explanation not stated on the page opened | Yes: "AI narrows the field and you make the call" | Not stated on the page opened | Not stated on the page opened | Not stated on the page opened | Personio recruiting |
| Pickr | Every candidate against the real requirements of the role on a professional skills graph, so related skills and other vocabulary count; verified qualifications shown separately from CV claims | Yes: plain-language reasoning per score, requirements met, partly met and missing, written so a client can read and challenge it | Yes: "You still make the call"; no hiring decision is fully automated | No public audit report named on the feature page | Yes: matching treated as a high-risk AI system; candidates informed before it is used; names and emails redacted from AI prompts by default; data stored in Europe | Pricing on request | /features/matching |
Tier 2: plug-in screening layers that run beside the ATS
| Tool | What is scored | Explanation shown to the recruiter | Human decision required | Bias documentation public | EU AI Act high-risk acknowledged | Price published? | Source |
|---|---|---|---|---|---|---|---|
| HireVue | Structured interviews, game-based and job-relevant skills assessments; an AI Interviewer runs voice-based interviews; the shortlist moves to live interviews with "smart summaries and scores" | Yes: "explainable recommendations" that support human decision-making; an AI Explainability Statement is published | Yes: "humans make the final hiring decision"; recruiters "remain responsible for evaluating candidates" | Partly: links an ORCAA algorithmic audit and an I-O psychology audit; no NYC Local Law 144 summary on the pages opened | Not stated on the pages opened | Not stated on the pages opened | HireVue, AI in hiring, Explainability statement |
| Paradox (Olivia) | A text conversation that screens "each candidate against job requirements", then schedules qualified candidates or dispositions others | Not stated; no scoring described | Partial: AI "may be used to support human decision-making"; Olivia can schedule qualified candidates automatically and can disposition others | Not stated; "regularly conducts bias evaluations", no results on the pages opened; points to Workday resources | Not stated; cites the NIST AI Risk Management Framework | Not stated on the pages opened | Screening, Ethical AI |
| CodeSignal | Coding, debugging and building in a real IDE; business, agentic and behavioural assessments; "scored automatically" | Yes: "a report showing proficiency by skill area" with qualitative and quantitative insights | Not stated explicitly; the team "gets back a skills report" to evaluate against the same criteria | Not stated; assessments "validated by I-O Psychologists"; the compliance page covers security only | Not stated on the pages opened | Not stated on the pages opened | Skills assessments, Compliance and trust |
| Peoplebox (Nova) | Resumes ranked against role requirements: skills, experience, certifications, career progression; profile enrichment from outside sources | Yes on the home page: "shows you why each candidate passed or failed"; the product page describes tagging, not reasoning | Yes: "Recruiters still make the final hiring decisions and can review any candidate manually" | Not stated; claims "consistent scoring" reduces bias, no audit named | Not stated on the pages opened | No: priced by employee count and modules, quote on request | Peoplebox, Resume screening |
| hireEZ (Applicant Review, EZ Agent) | Applicants sorted into Best, Good, Partial or Not a Match against an AI-generated Ideal Persona, on data the applicant submitted | Yes: "transparent insights into how candidates are matched and why"; "Every match comes with the why" | Yes: "Every decision comes to you"; recruiters can advance or reject a candidate at any point | Yes: Warden AI assurance page, monthly audits across 15 bias categories, latest 22 September 2026, sample 76,860 | Yes: the Warden page lists an "EU AI Act Bias Audit" under Regulation 2024/1689 | No, contact sales | Applicant Review, hireEZ, Warden AI assurance |
| Eightfold | Matching Model: skills and experience against a role, 0-to-5-star match score; AI Interviewer summarises each interview | Yes: each match score comes with reasoning in everyday language; each AI Interviewer insight "links back to the transcript and response that supports it" | Yes: "The AI recommends. People decide."; the software "does not autonomously reject a candidate" | Yes: BABL AI audit of the Matching Model dated 3 March 2026, impact ratios for gender, race/ethnicity and intersectional groups, all above 0.80; AI Interviewer audit also published | Yes: lists the EU AI Act as "Phasing in through 2027" with a readiness programme; ISO/IEC 42001 certified | No, demo request only | Responsible AI, Bias audit results, Eightfold |
Vendor pages change without notice. A statement on a marketing page is not a contractual commitment; put the explanation format, the override right and the audit cadence into the order form.
What does the recruiter actually see, and who decides?
The explanation column separates the tools more than the scoring column does. Three patterns appear in the tables:
- Reasoning per criterion. Workable, Ashby, Breezy and Pickr show why each criterion was met, partly met or missed. This is the format that survives a candidate's question or a client's challenge, because the recruiter can point at the sentence that produced the status.
- Categories instead of a composite. Greenhouse argues that "single scores obscure the reasoning behind a recommendation" and shows discrete categories with explanations (Greenhouse AI principles, accessed 2026-10-14). hireEZ's four match bands work the same way.
- A number with a summary. Manatal, Eightfold and HireVue attach summaries to a score; CodeSignal returns a skills report. Useful, as long as you can see the evidence under the summary.
Paradox is the outlier: its screening page describes the conversation and the outcome, not what the recruiter sees in between (Paradox screening, accessed 2026-10-14). That is not a flaw for a high-volume hourly role with yes-or-no requirements; it is a problem the moment somebody asks why a candidate was dispositioned.
On the decision itself, the pattern is consistent and worth reading carefully. Greenhouse states Talent Matching "does not have the ability to automatically advance or reject a candidate" (bias audit statement, accessed 2026-10-14). Workable ships auto-disqualify off by default and recommends leaving it off for a first run (help centre, accessed 2026-10-14). Eightfold, hireEZ, HireVue, Peoplebox and Personio say the recruiter decides. The gap is Paradox, whose assistant can schedule automatically and can disposition candidates, and Manatal and CodeSignal, which do not state it. A switch that is off by default is still a switch; the deployer who turns it on owns the result. The longer argument about what gets lost when a filter decides is in the match quality gap.
Which vendors publish bias documentation, and what does it prove?
Four vendors in the tables publish something you can open without a sales call. Eightfold publishes a BABL AI audit of its Matching Model dated 3 March 2026, with impact ratios by gender, race/ethnicity and intersectional group, all above the 0.80 four-fifths threshold (Eightfold bias audit results, accessed 2026-10-14). hireEZ and Greenhouse use Warden AI, which publishes a monthly dashboard; hireEZ's shows a latest audit on 22 September 2026 across 15 bias categories (Warden AI, hireEZ, accessed 2026-10-14). Breezy states a monthly Warden AI audit and a completed New York City Local Law 144 audit (Breezy Intelligence, accessed 2026-10-14). HireVue links an ORCAA algorithmic audit and an I-O psychology audit from its AI hub (HireVue AI in hiring, accessed 2026-10-14).
What a published audit proves is narrower than the badge suggests. A Local Law 144 audit reports selection rates by sex and race/ethnicity on the vendor's data, not on your applicants, and Eightfold's own page says the audit "does not make any determination whether the model is, in fact, an AEDT under this law". It tells you the vendor measured something and was willing to print it. It does not tell you how the model behaves on a German engineering pipeline with 40 applicants.
SHRM's 2022 survey is a useful reminder of the baseline: among organisations that get their tools from a vendor, "only 2 in 5" said the vendor was very transparent about bias safeguards, and 46% wanted more resources on identifying bias in these tools (SHRM, 2022, accessed 2026-10-14). Four public audits out of thirteen tools, four years later, is progress against that number and still a minority.
Does the EU AI Act change which tier you can use?
Not which tier; it changes what you must be able to show for either. Annex III of Regulation (EU) 2024/1689 lists AI intended to analyse and filter job applications and to evaluate candidates, which covers both tables. The AI Omnibus, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026, entered into force on 27 July 2026, and moved the application date for stand-alone Annex III high-risk systems from 2 August 2026 to 2 December 2027 (Cooley, accessed 2026-10-14; Future of Privacy Forum, accessed 2026-10-14). This is not legal advice; check the dates and your deployer duties with counsel.
What the tables show is how few vendors say so out loud. Eightfold lists the EU AI Act as "Phasing in through 2027" and names a readiness programme; hireEZ's assurance page lists an EU AI Act bias audit under Regulation 2024/1689; Pickr treats matching as high-risk and informs candidates before it runs. Ashby offers an "Automated Processing Legal Notice" and skips the evaluation for candidates who opt out, which is a GDPR Article 22 posture more than an AI Act one. The other nine do not mention the Act on the pages I opened. That silence is not non-compliance, but it is the first question for the vendor call, and the Article 6(3) assessment is the second. The deployer obligations, human oversight, candidate information and record-keeping, are in the EU AI Act and recruiting.
How many teams actually screen with AI, and what goes wrong?
Fewer than the category's marketing implies, and the gap between agencies and employers is wide.
- Germany, employers. Bitkom's representative survey of 852 companies with three or more employees, fielded in autumn 2024 and published 10 March 2025, found 1% screening applications with AI and 21% able to imagine it; 3% use AI-based competence analysis; 1% run AI-led interviews. The digital tools that are mainstream are video interviews (63%) and online tests or digital assessment centres (47%) (Bitkom, 2025, accessed 2026-10-14).
- Germany, agencies. The GVP's Barometer Personalvermittlung 2025, run by index Research on 354 placement firms, 370 companies and 1,000 candidates in June and July 2025, found 64.8% of placement firms using AI in recruiting, up 27 percentage points since 2023; 19.4% use matching tools and 17.8% let AI analyse incoming application documents. Among companies, 75.3% still use no AI tools in recruiting (GVP, Barometer Personalvermittlung 2025, accessed 2026-10-14).
- United States. SHRM's survey of 1,688 members, fielded 1 to 17 February 2022, found nearly one in four organisations using automation or AI in HR, 16% among employers with fewer than 100 workers against 42% at 5,000 or more. Of those using the tools, 64% said they automatically filter out applicants deemed unqualified, and 19% had seen qualified applicants excluded by accident (SHRM, 2022, accessed 2026-10-14).
Put the three together and the picture is an agency market that adopted screening quickly and an employer market that mostly has not, with one in five adopters in the US sample already having watched the filter reject someone it should not have. The reason that happens is usually vocabulary: the candidate did the job under a different noun. That is the case for skills matching over keywords in skills-based hiring.
Native or plug-in: which should you pick?
The choice follows the artefact you need scored, not the vendor list.
Pick native screening when the thing to evaluate is the application itself. One system, one DPA, the reasoning next to the pipeline stage, and the rejection reason in the same record as the rejection. For an agency presenting a shortlist to a client, a per-criterion explanation that can be forwarded as the recommendation is the whole point; see /features/matching for how Pickr does that, and compare it against Workable's criterion statuses and Ashby's reasoning links before you believe either of us.
Add a plug-in layer when the application is not enough evidence. A coding assessment (CodeSignal), a structured asynchronous interview (HireVue), or a conversational knockout for hourly volume (Paradox) scores something no CV contains. The price is a second system: get the write-back in writing, including the reasoning, so the ATS record does not read "rejected by integration".
Do not stack two scorers on the same artefact. An ATS that scores the CV plus a plug-in that also scores the CV gives you two numbers, two audit trails and one argument. From our work at ScalingPPL, the second score never settles the first; it just moves the discussion from the candidate to the tooling. What an interview-stage layer should do instead, and how it differs from CV screening, is in interview intelligence.
Where is Pickr the wrong choice?
- You need a published third-party bias audit today. Eightfold, Greenhouse, hireEZ and Breezy publish one; Pickr's feature page names none.
- You need to score interviews or assessments, not applications. HireVue, CodeSignal and Eightfold's AI Interviewer score artefacts Pickr's matching does not touch.
- You want screening on a free plan or a published price list. Workable, Breezy, Manatal and Ashby's Foundations plan publish prices; Pickr's pricing is on request.
- Your volume is hourly and the requirements are yes-or-no. A conversational knockout like Paradox resolves that without reasoning anyone will read.
- You hire only in the US. Data storage in Europe and candidate information before matching runs are not reasons to choose anything.
Where Pickr fits is a team whose applications are the evidence, whose clients or hiring managers ask "why this person", and who need the answer in the same record as the decision. Test that claim the way you should test every row above: open one real job, ask the tool to show the reasoning behind one score, ask who can override it and where that override is logged, and ask for the bias audit and the Article 6(3) assessment by email. The answers you get in writing are the comparison table that matters.
Sources
- SHRM, "Fresh SHRM Research Explores Use of Automation and AI in HR" (12 April 2022), https://www.shrm.org/about/press-room/fresh-shrm-research-explores-use-automation-ai-hr (accessed 2026-10-14)
- Bitkom, "Bewerbung fast überall digital, meistens ohne KI" (10 March 2025), https://www.bitkom.org/Presse/Presseinformation/Bewerbung-fast-ueberall-digital-meistens-ohne-KI (accessed 2026-10-14)
- GVP / index Research, "Barometer Personalvermittlung 2025" (PDF), https://personaldienstleister.de/wp-content/uploads/2025/11/Barometer-Personalvermittlung-2025.pdf (accessed 2026-10-14)
- Cooley, "Digital AI Omnibus delays key deadlines, introduces new rules" (3 August 2026), https://cdp.cooley.com/digital-ai-omnibus-delays-key-deadlines-introduces-new-rules/ (accessed 2026-10-14)
- Future of Privacy Forum, "The AI Act Implementation Timeline: What Changes Under the AI Omnibus?" (28 July 2026, updated 5 August 2026), https://fpf.org/?p=250371 (accessed 2026-10-14)
- Workable pricing, https://www.workable.com/pricing (accessed 2026-10-14)
- Workable help centre, "Getting started with the Workable Agent", https://help.workable.com/hc/en-us/articles/41050763193623-Getting-started-with-the-Workable-Agent (accessed 2026-10-14)
- Ashby pricing, https://www.ashbyhq.com/pricing (accessed 2026-10-14)
- Ashby docs, "AI-Assisted Application Review", https://docs.ashbyhq.com/ai-assisted-application-review (accessed 2026-10-14)
- Greenhouse pricing, https://www.greenhouse.com/pricing (accessed 2026-10-14)
- Greenhouse, "Bias audit statement", https://www.greenhouse.com/bias-audit-statement (accessed 2026-10-14)
- Greenhouse, "AI principles", https://www.greenhouse.com/ai-principles (accessed 2026-10-14)
- Manatal pricing, https://www.manatal.com/pricing (accessed 2026-10-14)
- Breezy, "Breezy Intelligence", https://breezy.hr/p/breezy-intelligence (accessed 2026-10-14)
- Breezy pricing, https://breezy.hr/pricing (accessed 2026-10-14)
- Personio, "Recruitment Management Software", https://www.personio.com/product/recruiting/ (accessed 2026-10-14)
- HireVue, home page, https://www.hirevue.com/ (accessed 2026-10-14)
- HireVue, "AI in Hiring", https://www.hirevue.com/ai-in-hiring (accessed 2026-10-14)
- HireVue, "AI Explainability Statement", https://legal.hirevue.com/product-documentation/hv-4-explainability-statement (accessed 2026-10-14)
- Paradox, "Screening", https://www.paradox.ai/products/screening (accessed 2026-10-14)
- Paradox, "Ethical AI", https://www.paradox.ai/ethical-ai (accessed 2026-10-14)
- CodeSignal, "Skills Assessments", https://codesignal.com/skills-assessments/ (accessed 2026-10-14)
- CodeSignal, "Compliance and Trust", https://codesignal.com/compliance-trust/ (accessed 2026-10-14)
- Peoplebox, home page, https://www.peoplebox.ai/ (accessed 2026-10-14)
- Peoplebox, "AI Resume Screening Software", https://www.peoplebox.ai/ai-resume-screening-software/ (accessed 2026-10-14)
- hireEZ, "Applicant Review", https://hireez.com/applicant-review/ (accessed 2026-10-14)
- hireEZ, home page, https://hireez.com/ (accessed 2026-10-14)
- Warden AI, "hireEZ Candidate Sourcing" assurance page, https://trust.warden-ai.com/hireez/candidate-sourcing (accessed 2026-10-14)
- Eightfold, "Responsible AI", https://eightfold.ai/trust/responsible-ai/ (accessed 2026-10-14)
- Eightfold, "Bias Audit Results", https://eightfold.ai/trust/bias-audit-results/ (accessed 2026-10-14)
- Eightfold, home page, https://eightfold.ai/ (accessed 2026-10-14)
Frequently Asked Questions
What is the difference between native AI screening and a plug-in screening layer?
Native screening runs inside the ATS on every application, so the score sits next to the pipeline stage. A plug-in layer such as HireVue, Paradox or Eightfold runs beside the ATS, scores its own activity (interviews, chats, assessments or matches) and writes a result back. It adds a vendor, a data flow and a second DPA.
Which ATS shows the reasoning behind an AI screening score?
As of 14 October 2026, Workable, Ashby, Breezy and Pickr state on their own pages that each score or status comes with written reasoning per criterion; Manatal describes AI summaries. Greenhouse shows discrete categories with explanations instead of one composite score. Personio names transparent criteria and AI signals without further detail.
Does AI screening make the hiring decision?
On the vendor pages opened, no. Greenhouse states Talent Matching cannot advance or reject a candidate, Workable ships auto-disqualify off by default, and Eightfold, hireEZ, HireVue, Peoplebox, Personio and Pickr all state that the recruiter decides. A tool that can auto-reject with the switch on is still a tool a human switched on.
Is AI screening high-risk under the EU AI Act?
Annex III of Regulation (EU) 2024/1689 lists AI used to analyse and filter job applications and to evaluate candidates. After the AI Omnibus, Regulation (EU) 2026/1744, the Annex III obligations apply from 2 December 2027 instead of 2 August 2026. This is not legal advice; check the dates with counsel before setting internal deadlines.
How many companies actually use AI to screen applications?
Fewer than the marketing suggests. Bitkom's survey of 852 German companies, published 10 March 2025, found 1% screening applications with AI. SHRM's 2022 survey of 1,688 members found nearly one in four organisations using automation or AI in HR, with 64% of those saying their tools automatically filter out applicants.
Written by Andreas Amann, Founder of Pickr and of ScalingPPL
Read more
Add Pickr as a preferred source on Google (opens in a new tab)