"AI-powered hiring" is one of the emptiest phrases in recruiting right now. Almost every tool says it. Far fewer can tell you what the AI actually touches, where it runs, and what it's allowed to decide. That gap is the whole problem — and it's exactly where OnSkillDemand is built differently.12
"AI-Powered" Means Nothing Until You Pin Down the Role
Look at how the market gets compared. Buyer guides now rank candidate-matching tools on seven dimensions at once — matching accuracy, database quality, integration ease, bias audits, pricing transparency, time-to-value, and agentic AI capability. The differentiators that actually matter are skills-based scoring, semantic job-description-to-resume fit, and false-negative rates: how many good people the AI quietly drops.2
Notice what none of those criteria assume: that you'd hand the AI a general-purpose chatbot and call it a hiring system. Some recruiting roundups deliberately leave ChatGPT off the list entirely, on the logic that you already know it and may already be using it at work. Knowing a model exists isn't the same as structuring its role. The role is the product.3
Start from the pain it's supposed to remove. A pile of resumes comes in, and someone has to pull out names, job titles, skills, and work history before any real evaluation can even begin. That's the job AI resume parsing was built to automate — extract the data, sync clean candidate profiles, and stop burning hours on retyping. Useful. But extraction is the easy half. The hard half is reasoning about fit — and being able to defend it.1
The Edge-Orchestrator Pattern
OnSkillDemand splits the work along a hard line. Ingestion happens at the edge. Reasoning happens on the backend. There's a strict separation between the two, and no unsupervised chatter between agents in the middle. One side gets your CVs in and structured. The other side thinks about them. Neither reaches across the line on its own.4
That boundary isn't decoration. When agents are allowed to free-associate with each other, you lose the thread of why an output happened. Keep ingestion and reasoning in separate lanes and every decision has a clean, replayable path from raw document to final score.4
Two Agents, One Job Each
The Ingestion Agent
The Ingestion Agent runs at the edge and does one thing well. It uploads every CV, extracts the content, chunks it into passages, and embeds those passages into a vector store. That's it. No scoring, no opinions, no ranking. It turns documents into searchable, addressable pieces — names, titles, skills, work history, all the structured signal — and stops there.1
The Reasoning Agent
The Reasoning Agent runs a RAG pipeline on the backend. It retrieves the most relevant CV passages for a given vacancy, scores candidate-to-vacancy fit, and writes a natural-language justification for the score. This mirrors the arbiter pattern at the end of retrieval: a single LLM call ranks the candidates with reasons and returns one typed object. The point isn't to sound smart — it's to produce an output your auditor can defend.4
Why This Is an Auditability Pattern
Here's the part that matters when someone asks you to justify a shortlist. Because the Reasoning Agent only ever scores passages that the Ingestion Agent retrieved and tagged, every fit score traces back to the exact CV chunk that produced it. You don't get a black-box number. You get a number, the reason, and the line of the resume the reason came from. Click the score, see the source.4
One LLM call ranks the candidates with reasons. The output is one typed object your auditor can defend.
Towards Data Science, on the arbiter RAG pattern
RAG-Fusion Across CVs, Job Descriptions, and Rubrics
Matching isn't a keyword search. OnSkillDemand runs RAG-Fusion across three things at once — CV chunks, the job description, and the scoring rubric — so fit is judged on meaning, not just word overlap. That's the semantic JD-to-resume match the market keeps pointing to as the real differentiator, and it's what lets the system surface strong candidates out of a large applicant pool in minutes rather than days.25
Semantic technology pulling top candidates from hundreds of applications with one click is the promise the category sells. The OnSkillDemand difference is that the click comes with a trail — rubric, retrieved passages, and a written rationale you can stand behind.5
The Hard Boundary: AI Suggests, You Decide
This is the line we won't cross. The AI parses, matches, transcribes, and suggests. It never makes a hiring decision and never rejects a candidate on its own. Every value it produces is a draft you can override. The AI suggests. The recruiter decides. Full stop.4
That isn't just a values statement — it's where regulation is heading. As of March 2026, if your business uses AI to screen, rank, or match candidates, the EU regulates those tools as high-risk systems under the AI Act, with staffing-specific guidance published on 17 March 2026. "High-risk" means traceability and human oversight aren't optional extras. An architecture where every score points back to a CV chunk and a human signs off on every outcome isn't a nice-to-have. It's the shape the rules now expect.6
- What the AI does: upload, extract, chunk, embed, retrieve, score, justify, transcribe, suggest.
- What the AI never does: reject a candidate, finalize a hire, or lock a value you can't change.
- What you do: review the rationale, check the source chunk, override anything, decide.
Recruiters are already measured on cost per hire, days from post to offer, hiring ROI, and the cost of attrition. AI should move those numbers — without making you defend a decision you can't explain. That's the whole design: structure the AI's role, keep ingestion and reasoning apart, tie every score to its source, and leave the final call where it belongs. With you.7