The opaque score
“Trust the 87%” is not a defence.
Composite scores with hidden weights cannot be debugged. If the hiring manager disagrees, there is no productive conversation — only deference to the algorithm.
AI resume screener
An AI resume screener that cites its sources.
Most AI resume screeners hand you a fit score and ask you to trust it. ShortlistTable gives you per-criterion AI verdicts, the resume sentence behind each one, and a one-click recruiter override on every cell. Defensible to the hiring manager, auditable to legal, fast enough to ship before lunch.
Per-cell evidence · One-click override · No auto-reject · No model lock-in
shortlisttable.app / tables / ai-screener-q2Auditable AI
| Candidate | AI verdict | Source cited | Confidence | Override? | |
|---|---|---|---|---|---|
Alice Chen Ledger · 7y | Yes | p.1 | 0.91 | — | Call first |
Marco Silva CloudCo · 12y | Review | p.2 | 0.64 | Pending | Review fit |
Nina Patel Indie · 4y | No | — | 0.22 | — | Hold |
Citation · Alice Chen · “Built and operated Kafka streaming pipelines at Ledger from 2022-present” · resume p.1Source-backed
Per-cell
AI verdict + source citation
1-click
Recruiter override
0
Auto-rejections
Reproducible
Same input, same verdict
Why most AI screeners fail
An opaque 87% fit score is not a screening decision.
AI resume screening went mainstream in 2024 and a lot of recruiters got burned. The first generation of tools optimised for the wrong thing: a single composite score, no per-criterion breakdown, no source citation, often a hidden auto-reject below threshold. When the verdict is wrong, you have nothing to push back against. When the hiring manager disagrees, the conversation becomes about the model instead of about the candidate.
We took the opposite design path. The AI does the reading; you make the call. Every verdict carries the criterion, the source sentence, and a one-click override. The shortlist holds up because every cell in it can be traced back to a sentence in the resume.
The auto-reject button
Hidden screen-outs are a legal liability.
Many AI screeners silently auto-reject candidates below a confidence threshold. That’s how lawsuits get written. We never auto-reject — the bottom of the list is a hold queue with evidence intact.
The non-deterministic run
Same pile, different shortlist.
If you re-run the same resumes through most AI screeners you get a subtly different ranking. Ours is deterministic — the same input produces the same verdicts every time, so your shortlist is reproducible for audits.
How auditable AI screening works
Six properties of a defensible AI screener.
01 · Feature
Per-criterion AI verdicts
The AI judges each candidate column-by-column. The shortlist position is derived from the verdicts, not the other way around — and every verdict is independently auditable.
Yes · 5y seniorReview · domain fitNo · stack mismatch
02 · Feature
Source sentence per cell
Every AI verdict links to the resume sentence behind it. Click any cell to jump to the highlighted PDF. No verdict without evidence.
Resume p.1 · Experience“Led platform team of 6 engineers building event-driven services on AWS.”
03 · Feature
Per-column confidence threshold
Below your confidence threshold, verdicts land in a review queue instead of being auto-decided. The recruiter handles the ambiguous cases; the AI handles the obvious ones.
0.75
default confidence
threshold
04 · Feature
One-click recruiter override
Disagree with a verdict? Click, pick a different value, add a note. The original AI verdict stays in the audit history; your override becomes the active value.
AI: review · 0.62
Override: yes (manual)
05 · Feature
No auto-reject, ever
Low-confidence or low-fit candidates land in a hold queue with evidence intact — never silently deleted. Any rejection is reviewable, not invisible.
Hold queueRe-runnableAudit trailOverride-friendly
06 · Feature
Exportable audit trail
Every verdict, override, re-run, and column edit is timestamped and exportable. When legal asks why someone was screened out, you have a per-cell paper trail.
AI verdict · 09:14
Override · 11:02
Re-run after column edit · 08:30 (next day)
How it compares
Auditable AI screening vs the alternatives.
| Property | Black-box AI screener | ATS AI scoring | Spreadsheet | ShortlistTable |
|---|---|---|---|---|
| Per-criterion verdicts | ✕Composite score | ✕Single rank | ✓Manual | ✓Typed per column |
| Source sentence cited | ✕Rare | ✕No | –Manual notes | ✓Every cell |
| Recruiter override + audit | –Limited | –Limited | ✓Yes | ✓Full audit trail |
| Auto-reject below threshold | ✕Common | ✕Common | ✓Never | ✓Never — hold queue |
| Reproducible across runs | ✕Non-deterministic | –Varies | ✓Manual | ✓Deterministic |
| Exportable audit trail | ✕Rare | –ATS-shaped | ✕None | ✓Per-cell timestamps |
✓ Supported– Partial / manual✕ Not supported
FAQ
Common questions on AI resume screening.
Is this just another AI resume screener?+
No. The architectural difference is that every verdict surfaces its source sentence, recruiter override is first-class, and there’s no auto-reject anywhere in the system. Other AI screeners optimise for sorted output; we optimise for defensible decisions.
How does this comply with the EU AI Act / NYC AEDT / Colorado AI law?+
These regulations target “automated employment decision tools” — systems that make or substantially influence hiring decisions on their own. ShortlistTable is explicitly a human-in-the-loop tool: AI gathers evidence, the recruiter decides. Every verdict is overridable; the bottom of the list is a hold queue, not a delete bucket. We provide a per-workspace audit log for any required disclosures or bias audits.
What model do you use? Can it be locked in?+
The product is model-agnostic by design — we route different column types to different models based on what’s appropriate. We use evaluation-driven model selection rather than picking a single provider. There is no model lock-in for the customer.
Will the AI auto-reject candidates?+
No. There is no path in the system that removes a candidate without a recruiter action. Low-confidence and low-fit candidates land in a hold queue with full evidence intact, so any rejection is reviewable rather than silent.
How do you avoid systematic bias?+
Two design choices: (1) every verdict surfaces the source sentence, making systematic mismatches visible to the recruiter rather than hidden inside a model; (2) recruiter override is logged and analysable per workspace, so a team can audit where they’re consistently disagreeing with the AI and adjust columns or workflow.
Is the verdict reproducible if I re-run the same pile?+
Yes. Same inputs + same column definitions = same verdicts, every time. This matters for audit trails and for any retrospective on a hire that did or did not work out.
How does this compare to using ChatGPT directly?+
Direct LLM chat doesn’t give you per-cell citations, persistent column templates, confidence thresholds, override audit trails, or ATS-shaped exports. ShortlistTable is the orchestration layer around the model — built for the workflow, not the chat.
More from us
Keep exploring.
Related
Explainable AI resume screening
EU AI Act, NYC AEDT, Colorado AI law.
Related
The AI recruiting copilot
AI gathers, the recruiter decides.
Related
Source-backed screening
Citations on every cell.
Free tool
Try the free JD ↔ CV matcher
See evidence-backed matching live.
Read
Keyword vs semantic screening
What semantic AI changes.
AI you can defend
The AI resume screener that holds up in a debrief.
Try ShortlistTable on a 25-resume pile. Every verdict carries a source sentence. Every cell is overridable. Free, no credit card.