Two things have to work together for an AI candidate screening system to do anything useful. One is the part that produces a number about a candidate. The other is the part that decides what to do with that number. Neither one is the product. The relationship between them is.
If you've ever wondered why one AI candidate screening system feels intelligent while another feels like a spreadsheet with extra steps, the difference almost always lives in that relationship.
Here's what the two parts actually look like inside a working AI candidate screening system, where they tend to fall out of sync, and what to look for when you're picking which one to bring in-house.
Two Engines, One System
Every AI candidate screening system is built around two engines.
The scoring engine takes a candidate's resume, interview transcripts, application answers, and public signals, and produces a structured evaluation. A score. A recommendation. A summary. A list of strengths. A list of concerns. And, in good systems, a per-criterion breakdown that explains which evaluation dimensions the candidate hit and which they missed.
The workflow engine takes that evaluation and decides what happens next. Auto-reject. Route to AI interview. Send to HR review. Schedule with a hiring manager. Pass to a take-home assignment. Each role has its own rules, and the workflow engine fires those rules against the scoring output to move the candidate forward.
Most platforms market themselves on one of these two engines. The good ones treat both as primary, and design the contract between them deliberately.
Inside the Scoring Engine
Strip the marketing language away and the scoring engine has a small set of jobs. Parse the input. Apply the criteria for the specific role. Produce per-criterion scores. Roll them up into an overall match score using a weighting model. Generate an explainable summary. Flag which Must Have criteria the candidate did or didn't meet.
The interesting design decisions live in two places: the weighting model and the explainability layer.
The weighting model. Different roles weight criteria differently. A staff infrastructure engineer's rubric isn't a junior support rep's. Good scoring engines let you assign weights per criterion and tag each one as Must Have or Nice to Have. The Must Have tags act as gates. A candidate can score 80 overall and still get rejected because they failed on a Must Have. A candidate can score 65 overall and still advance if they nailed every Must Have and the rest is workable. Without that distinction, you're ranking candidates by a single number that can't represent "this person is great except for one thing the role can't compromise on."
The explainability layer. The score isn't the output. The reasoning behind the score is. A real scoring engine produces a written summary explaining why the candidate scored what they scored, a list of specific strengths the system noted, a list of specific concerns, and a per-criterion breakdown showing which dimensions were strong and which were weak. Without this, the score is unauditable. With it, the score becomes a hiring conversation a manager can actually have.
Careerswift Hire's scoring engine surfaces this stack natively. Roles typically configure 8 to 18 weighted criteria, with the platform supporting hundreds of dimensions for roles that call for it. Each criterion gets a Must Have or Nice to Have tag. Each evaluation produces the overall weighted score, the AI recommendation, a written summary, a strengths list, a concerns list, and the per-criterion breakdown. You can start from a ready-made template for common roles, build your own criteria from scratch, or import a proprietary scoring model entirely.
Inside the Workflow Engine
The workflow engine is the part that reads what the scoring engine produced and decides what happens to the candidate next.
The fundamental unit is a routing rule. A routing rule has a condition and a destination. The condition reads from the candidate's evaluation. The destination is the next node in the workflow.
Routing rules come in a few flavors, and the kind of rules an engine supports tells you a lot about how it was built.
Threshold gates. The simplest form. If the overall score is below 60, route to rejection. If it's above 80, route to hiring manager review. Everything else, route to AI interview. Cheap to implement, easy to abuse. A pure threshold gate ignores everything the scoring engine produced except the headline number.
Multi-condition branches. The interesting form. If the overall score is above 70 and no Must Have criterion failed, route to AI interview. If the overall score is above 70 but a Must Have failed, route to recruiter review. If the overall score sits between 50 and 70, route to a take-home assignment. The conditions reference the structured output of the scoring engine, not just the number.
Fallback paths. The forgotten form. What happens when the score is ambiguous, the integrity flags fire, or the candidate's profile doesn't match any existing rule. Robust workflows always have a default route. Fragile ones drop candidates into limbo on the edge cases nobody designed for.
Parallel evaluation. The advanced form. Multiple evaluators (resume scoring, AI interview, identity verification, technical assessment) run concurrently. A join node aggregates the results before the next routing decision. This is the difference between a 14-day time-to-decision and a 3-day one.
Careerswift Hire's workflow engine supports all four through Logic and Routing nodes on the visual canvas. You can define named branches that read both the overall score and the per-criterion structure, with patterns like "Invite to HR Interview" routing to Approve, "AI Interview" routing to a next step, and "Otherwise" routing to Reject. Rules are editable per role without engineering involvement. The branches are visible on the canvas instead of buried inside modal tooltips.
Where the Two Engines Fall Out of Sync
Most of the gap between an AI candidate screening system that actually works and one that just runs is in the contract between scoring and workflow.
The workflow only reads the headline number
The scoring engine produces 14 fields of structured evaluation. The workflow engine reads one of them: the overall score. The other 13 are stored but never consulted in a routing decision. Candidates who failed a Must Have but scored high overall sail through. Candidates who passed every Must Have but had a low headline get rejected. The signal is there. The workflow just doesn't use it.
Routing rules can't reference structured output
You want a rule that says "advance if the candidate passed every Must Have criterion." The workflow engine only supports rules of the form "advance if score is greater than X." You end up writing convoluted threshold rules that approximate the logic you actually want. Six months later, nobody on your team can read your own workflow.
Ambiguous cases get force-routed
The candidate's score is 68. Your threshold is 70. The workflow auto-rejects. There's no path for "this is borderline, send to a human." Every borderline case is treated as a clean rejection, and the system loses every candidate who didn't quite hit the magic number for reasons that have nothing to do with their fit.
The AI interview is blind to the screening output
The scoring engine identified that the candidate has strong system design experience but weak performance optimization experience. The AI interview asks generic questions about both, instead of probing the gap the scoring engine already flagged. Hours of interview signal go uncollected because the two engines don't share context.
Manual overrides skip the workflow entirely
The recruiter sees a candidate they like and clicks "advance" outside the workflow. The override is recorded as a status change, not as a routed decision with an audit trail. The next time someone asks "why did this candidate get an interview," there's no record of who decided, on what data, or why.
When the Two Engines Were Built as One
In a system where the two engines were designed together rather than glued together later, a few specific things are true.
The scoring engine produces structured output, not just a number. Per-criterion scores. Must Have versus Nice to Have. A written reasoning trail. Integrity flags from the authenticity layer.
The workflow engine reads the structured output, not just the number. Routing rules can reference specific criterion scores, Must Have status, integrity flags, and the overall recommendation.
Borderline cases get human review nodes built into the workflow, not auto-routed past a hard threshold.
The AI interview layer adapts in real time, generating contextual follow-up questions based on the candidate's actual responses instead of reading from a static script.
Fraud detection runs in parallel with the screening workflow, with real-time alerts when something looks off across the candidate's session.
Every routing decision is recorded with the evaluator inputs, the rule that fired, and the outcome. Manual overrides are logged with the recruiter, the reason, and the candidate state at the time.
Careerswift Hire is built along these lines. The Scoring Configuration nodes produce structured evaluations with the Must Have and Nice to Have framework. The Logic and Routing nodes consume that structured output directly on the workflow canvas. The AI interview layer is adaptive, generating contextual follow-up questions based on candidate responses, and covers both HR pre-screening (behavioral and cultural fit) and technical interviews (knowledge validation and problem-solving), running unlimited interviews in parallel. The Fraud Detection layer (AI Answer Detection, Profile Cross-Check against LinkedIn and GitHub, Identity Consistency across stages, Behavioral Anomaly flags, and Browser Focus Monitoring during interviews) runs in parallel with the screening workflow and surfaces real-time alerts. GDPR-compliant and privacy-first by default. Usage-based pricing (roughly €200 per 1,000 candidates screened and €10 per 50-minute AI interview) instead of seat licenses that punish you for screening more candidates.
How to Tell If the Two Engines Were Designed Together
Most vendor demos walk you through the dashboard. They don't walk you through the contract between scoring and workflow. These five questions do, and they take about ten minutes.
"Show me a routing rule that references a per-criterion score, not just the overall." A vendor with a connected system has examples ready. A vendor with a glued system pulls up a threshold gate and tells you "you can build that with our scripting feature."
"What happens to a candidate whose overall score is 70 but who failed a Must Have criterion?" Connected systems route them to recruiter review or reject them with a specific reason tied to the failed criterion. Glued systems either advance them (because the number is high) or hide the Must Have failure inside an export field nobody looks at.
"Show me a real evaluation, with the overall score, the recommendation, the summary, the strengths, the concerns, and the per-criterion breakdown all in one view." Connected systems show you the whole evaluation natively. Glued systems show you the number and tell you the reasoning is "available on request."
"Show me the audit trail for a candidate the recruiter manually advanced past the workflow." Connected systems log the override with the operator, the reason, and the state diff. Glued systems show you a status change with no surrounding context.
"How does the system handle a candidate whose integrity flags fired mid-interview?" Connected systems re-route in real time, surface the flag to a reviewer, and capture the context. Glued systems flag the candidate in a separate dashboard nobody opens.
You can usually tell within five minutes of those questions which kind of vendor you're talking to.
The scoring engine is one product. The workflow engine is another product. The contract between them is the actual platform. Most marketing pages emphasize the first two. Most of the operational value lives in the third.
When you're evaluating an AI candidate screening system, the work isn't to compare scoring features against scoring features and routing features against routing features. The work is to ask whether the two engines were designed together or grafted onto each other in version 2.4. That difference is what determines whether the platform feels like a system that works or a collection of features that happen to ship under the same login.
Platforms that get this right (Careerswift Hire being one current example) make the contract between the engines visible on the workflow canvas, so the routing logic and the scoring logic stay in conversation instead of drifting apart. That's the part worth checking before you buy anything else.
