AI for HR & Recruitment
Recruitment is the vertical where AI is most useful and most regulated at the same time. Everything we build here assumes an auditor will eventually ask how a score was produced, because in this industry one eventually will.
Keyword matching is not assessment
The default HRTech architecture parses a resume, extracts keywords and ranks against a job description. It is fast, it is cheap, and it measures how well someone writes a resume rather than whether they can do the work. The failure modes are well known: career changers with genuinely transferable skills score badly because their vocabulary is from the wrong industry, candidates returning from a break are penalised for a gap that says nothing about capability, and anyone who has learned to write for the parser outranks anyone who has not. Bolting a large language model onto the same architecture makes it faster and more confident without making it more accurate, and in a high-risk regulatory category, confident and wrong is the expensive outcome.
What AI actually does in hr & recruitment.
The four workloads that carry the value in this industry, in the order they usually pay off.
Conversational assessment
Instead of quizzing definitions, interview for evidence: the project someone is proudest of, a hard problem they solved, a mistake they learned from, how they handled conflicting priorities. This validates authentic experience while revealing how a person reasons and communicates. It is the core of the TalX platform we built.
Scenario-based scoring
Put a candidate in a real situation from the target role, a stakeholder demanding an off-roadmap feature or an escalated unhappy customer, and evaluate reasoning, prioritisation and empathy rather than hunting for one correct answer. Scenarios are far harder to game than question banks, and they produce evidence a hiring manager can actually read.
Skill-gap analysis and career pathing
The output that changes a decision is not a ranking, it is a gap. Which competencies are strong, which are missing for the target role, and what closes the distance. TalX turns a single assessment into a personalised report covering strengths, readiness, suggested paths and learning mapped to the gaps.
Structured resume and profile parsing
Parsing still has a job, it is just no longer the decision-maker. It seeds context for the assessment and normalises history so the conversational layer starts informed rather than cold.
You are building in a high-risk category. Design for it.
Annex III of the EU AI Act classifies AI systems used in employment, recruitment, screening, promotion and worker management as high-risk, and the obligations attaching to that classification became enforceable on 2 August 2026. Separately, New York City's Local Law 144 requires an annual independent bias audit of automated employment decision tools used for NYC-located roles, tested for disparate impact across race, ethnicity and sex including intersectional analysis, with the results published. Neither regime cares how sophisticated your model is; both care what it decides and whether you can evidence it. Practically that means score decomposition rather than a single opaque number, retained inputs and model versions so a past decision can be reconstructed, a documented human review step, and evaluation data held in a shape an independent auditor can actually work with. Building those in from the start costs a fraction of retrofitting them under deadline.
The practice areas this work draws on.
Published work in hr & recruitment.
HR & Recruitment questions, answered.
It depends on what the system decides, not on how advanced it is. A tool that screens, ranks or filters candidates for employment decisions falls under Annex III; a tool that gives an individual private career guidance with no employer in the loop generally sits differently. We classify the specific use case at scoping time and flag where a data protection impact assessment is required before build starts. Formal sign-off is your counsel's call.
You make it defensible by architecture, not by explanation after the fact. That means scoring against a defined competency rubric rather than free-form judgement, decomposing the result into per-competency evidence, versioning models and prompts so any past decision can be reconstructed, and retaining evaluation data in a form an independent auditor can test for disparate impact.
This is the case that motivates scenario-based assessment. Because the evidence comes from how someone reasons through a real situation in the target role rather than from the shape of their history, a gap or an unfamiliar industry vocabulary stops being a proxy for capability.
Yes, and the design should force it rather than hope for it. Full automation is the wrong goal in a high-stakes decision. The right pattern routes ambiguous cases to a reviewer with the evidence surfaced, which both catches what the model is unsure about and generates high-quality labelled data from expert decisions.