Integrating Agentic AI with ERP Systems for Project-Based Recruitment - A Sociotechnical Framework for Fair Candidate Screening System Evaluated Through Empirical Audit of Gender Bias in Embedding and LLM
Authors
Dey, Piyal ; Jensen, Camilla Louise
Term
4. semester
Publication year
2026
Submitted on
2026-05-31
Pages
79
Abstract
Plant Supervision, a Danish project-based recruitment firm, currently searches its internal CV database with keyword filters and then shortlists via recruiter relationships. This is slow in a competitive market and structurally disadvantages candidates without prior ties. To address this, the company built HireX, an agentic AI screener that combines embedding-based retrieval (numerical representations of text used to find semantically similar CVs) with a LangGraph-orchestrated language model screener. The key question was whether HireX produces fairer recruitment than the legacy workflow, in line with EU responsible-AI obligations for high-risk systems. This thesis audits HireX using Sociotechnical Systems Theory across eight research questions to identify a jointly optimized configuration. We tested 485 paired counterfactual male/female CVs at three writing-style intensity levels, sweeping over embedding strategy and model, input format, prompting strategy, language model, and the presence of a bias-aware audit agent. Bias was measured with standard metrics: Statistical Parity Difference, Disparate Impact, Counterfactual Fairness Gap, and an exposure-weighted variant. Findings show that gender can be inferred from anonymized resumes; a section-wise embedding strategy yields the most biased retrieval; using full job descriptions as queries reduces retrieval bias versus title-only; and writing-style divergence amplifies bias even under fairer settings. The recommended deployment uses full-text MPNet retrieval on anonymized CVs, full job descriptions as queries, the baseline screening prompt on GPT-4.1-nano with anonymized labels, no separate audit agent, a recruiter-in-the-loop advisory agentic AI framework, and API-based integration with the Maconomy ERP system. Overall, jointly optimizing technical and social components supports a responsible, fairer deployment.
Plant Supervision, en dansk projektbaseret rekrutteringsvirksomhed, bruger i dag nøgleordsfiltre i en intern CV-database efterfulgt af udvælgelse baseret på rekruttørrelationer. Den proces er langsom i et konkurrencepræget marked og systematisk ufordelagtig for kandidater uden tidligere relationer. For at afhjælpe dette blev HireX udviklet: en agent-baseret (agentisk) AI-screener, der kombinerer embedding-baseret søgning (numeriske tekstrepræsentationer, som finder semantisk lignende CV’er) med en LangGraph-orkestreret sprogmodellscreener. Spørgsmålet var, om HireX reelt giver mere fair rekruttering end den tidligere arbejdsgang, set i lyset af EU-krav til ansvarlig AI for højrisikosystemer. Denne afhandling auditerer HireX ud fra socioteknisk systemteori på tværs af otte forskningsspørgsmål for at finde en samlet, optimeret konfiguration. Vi tester 485 parrede, kontrafaktiske mandlige/kvindelige CV’er ved tre niveauer af skrivestilsintensitet og varierer embedding-strategi og -model, inputformat, promptstrategi, sprogmodel samt tilstedeværelsen af en bias-bevidst audit-agent. Bias måles med standardmål som Statistical Parity Difference, Disparate Impact, Counterfactual Fairness Gap og en eksponeringsvægtet variant. Resultaterne viser, at køn kan udledes fra anonymiserede CV’er; en afsnitsvis embedding-strategi giver den mest biased hentning; fulde jobbeskrivelser som forespørgsler reducerer hentningsbias sammenlignet med kun jobtitel; og større skrivestilsforskelle forstærker bias, selv i mere fair opsætninger. Den anbefalede implementering er fuldtekst MPNet-hentning på anonymiserede CV’er, brug af fulde jobbeskrivelser som forespørgsler, baseline-screeningsprompt på GPT-4.1-nano med anonymiserede labels, ingen separat audit-agent, en rådgivende recruiter-in-the-loop-ramme for agentisk AI samt API-baseret integration til Maconomy-ERP. Samlet set viser en fælles optimering af tekniske og sociale komponenter en ansvarlig vej til at forbedre både effektivitet og fairness.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords