AI CV screening against the job ad

The HR team receives CVs for open roles throughout the day, typically through email or directly into their HR SaaS system. For each CV, a member of the HR team reviews it against the job ad requirements, looking at skills, experience, qualifications, and overall fit. They then generate a score that reflects how well the candidate matches the role, document their reasoning for that score, and prepare a set of suggested interview questions tailored to the candidate's background and the role's needs. This scored package is then passed to the hiring manager for decision-making. The process runs multiple times daily and handles around 2,600 CVs per year, with each evaluation taking roughly 13 minutes of focused HR time.

The challenge is that the current manual approach leads to inconsistency and burnout. The HR manager and assistant spend over 500 hours per year on routine screening, and when volume spikes they resort to shallow first-pass reviews that sometimes cause strong candidates to be overlooked. Around 10 to 15 percent of CVs do not fit a standard pattern, such as candidates with non-traditional career paths or ambiguous qualifications, and these require more careful judgment. The rest are relatively straightforward comparisons of listed skills and experience against job requirements, work that is repetitive and time-consuming but does not require deep expertise once the criteria are clear.

The hybrid approach is the right fit for this process because it addresses both the efficiency problem and the regulatory reality. The client needs to free up HR time spent on routine screening, but they operate in an environment where GDPR, works council oversight, and the AI Act all require human accountability in recruitment decisions. A fully automated AI workflow or autonomous agent would deliver stronger time savings on paper, but it would create unacceptable compliance and discrimination risk. The AI Act explicitly classifies recruitment AI as high-risk and mandates human oversight. The works council consultation mentioned in the process record signals that this organization takes employee and candidate rights seriously, and pushing through a black-box AI solution would likely face internal resistance and external scrutiny.

The hybrid path threads this needle by using AI to handle the 85 to 90 percent of CVs that follow standard patterns, where the matching logic is clear and the risk of error is low, while routing the 10 to 15 percent of edge cases and any low-confidence scores to human review. This design keeps the HR team in control of final decisions and satisfies the regulatory requirement for human oversight, while still automating the bulk of repetitive work. The AI generates a score, explains its reasoning, and suggests interview questions, but a human must review and approve before the output reaches the hiring manager. This is not just a compliance checkbox. It is also good practice because the decision points identified in the assessment, such as how to score ambiguous qualifications or whether exceptions require special handling, genuinely require human judgment that current AI cannot reliably replicate.

The client has no stated technology stack preference, which gives flexibility in implementation. The solution can be built on any major cloud platform with access to a capable LLM, such as GPT-4 on Azure OpenAI, Claude on AWS Bedrock, or Gemini on Google Vertex, paired with workflow orchestration to handle routing and human review gates. The HR system already has an API on a higher subscription tier, so upgrading that tier and building the integration will eliminate the current reliance on email and spreadsheets and create a clean end-to-end flow. The tradeoff here is upfront investment in the HR system upgrade and API integration work, but that investment pays for itself quickly given the 563 hours per year currently spent on manual screening.

The alternative paths fall short for specific reasons. A fully automated AI workflow scores well on efficiency but fails on compliance and the client's explicit concern about audit and discrimination risk. An autonomous AI agent is even riskier and would almost certainly run afoul of AI Act requirements. RPA could move data around but cannot perform the cognitive work of evaluating fit or generating questions, so it only solves a small part of the problem. Building a traditional coded solution would require the one-person IT team to encode and maintain complex hiring rules, which is not realistic given their capacity and the variability of CVs and job requirements. Staying manual fails to address the core issue the client came to solve.

The hybrid approach is honest about what AI can and cannot do in this context. It automates the repetitive comparison work that consumes most of the 13 minutes per CV, but it does not pretend that AI can replace human judgment on edge cases or take legal accountability for hiring decisions. It delivers 70 percent time savings, meets regulatory requirements, and keeps the HR team in control, which is exactly what this client needs.

The top three paths for this process are Hybrid, AI Workflow, and RPA. The Hybrid approach scored highest at 9 because it delivers the core benefit of AI-powered screening while satisfying the legal and ethical constraints that make this decision complicated. The AI Act classifies recruitment systems as high-risk and mandates human oversight of automated decisions. The works council consultation requirement means any solution that removes humans from the hiring loop will face resistance and possibly rejection. The Hybrid path routes 85 to 90 percent of standard CVs through automated AI scoring and question generation, then pushes the 10 to 15 percent of exceptions and all final decisions to human reviewers. This keeps the HR team accountable while eliminating the repetitive work that currently consumes 563 hours per year.

AI Workflow scored 6 and remains viable if the client is willing to accept tighter regulatory scrutiny. A fully automated workflow with no mandatory human gates would save more time and cost less to build, but it creates compliance risk. The AI Act requires explainability and human oversight for high-risk use cases, and a pure automation approach makes it harder to demonstrate that a human made the final hiring decision. The 10 to 15 percent exception rate also means a fully automated system would produce poor outcomes for a meaningful minority of candidates, which undermines the goal of improving hiring quality. If the client can establish that the AI system only provides recommendations and humans always make the final call, this path becomes more attractive, but the legal advice and works council consultation would need to confirm that structure is acceptable.

RPA scored 5 because it solves a different problem. It can move data between email, spreadsheets, and the HR system, which reduces manual entry errors and saves a few minutes per CV. However, it cannot evaluate whether a candidate is a good fit for the role or generate relevant interview questions. The 13 minutes per CV is mostly evaluation time, not data entry time, so RPA would deliver only marginal savings. It makes sense as a tactical fix if the client wants to automate data handling while keeping the evaluation manual, but it does not address the core pain point the client described. The Hybrid path is better because it tackles the evaluation work directly while still preserving human judgment where it matters.

The recommended Hybrid solution is built on an LLM-powered workflow with human review gates and API integration to the HR system. When a new CV arrives for an open role, the system pulls the job ad and CV from the HR platform via API. An LLM such as GPT-4, Claude, or an equivalent model hosted on Azure OpenAI, AWS Bedrock, or Google Vertex parses both documents and extracts key information such as required skills, experience level, qualifications, and candidate background. The LLM then compares the CV against the job ad requirements and generates a fit score with written reasoning that explains why the candidate is strong, marginal, or weak. It also produces three to five suggested interview questions tailored to gaps or strengths in the candidate's profile. This entire process takes seconds and replaces the 13 minutes of manual review.

The workflow engine evaluates the confidence level of the AI output using rules defined during the build phase. If the CV fits a standard pattern, the required skills are clearly present or absent, and the LLM's reasoning is coherent, the system assigns a high confidence score and routes the result directly to the hiring manager for final decision. If the CV has an unusual career path, the qualifications are ambiguous, or the LLM's reasoning contains uncertainty markers, the system flags the case for human review. The HR assistant or manager reviews these exceptions, reads the AI reasoning, and either accepts, modifies, or overrides the score and questions. All final decisions, whether routine or exceptional, are recorded in the HR system with a timestamp and the name of the approving human. This audit trail satisfies GDPR and AI Act documentation requirements.

The integration layer is critical because the current process relies on email and spreadsheets, which create data gaps and manual entry errors. Upgrading the HR SaaS subscription to a tier with API access allows the system to pull CVs and job ads automatically and push scores and interview questions back into the platform where hiring managers already work. This eliminates the need for HR staff to copy and paste between systems or maintain parallel spreadsheets. The workflow orchestration can be built on a low-code platform such as Make, Zapier, or n8n, or on a custom service using Python and a task queue if the client prefers more control. The LLM API calls are stateless and fast, so the system can handle multiple CVs in parallel during high-volume hiring periods.

The pilot phase is essential because this is a high-stakes use case with regulatory and reputational risk. Running the Hybrid system in parallel with the manual process for one or two open roles allows the HR team to compare AI scores and questions against their own judgment and identify cases where the AI misunderstood requirements or missed nuance. This feedback loop refines the confidence thresholds and exception rules before full rollout. Monthly audits after deployment track AI decisions by protected characteristics such as age, gender, and ethnicity to detect bias patterns early. If the data shows that the AI systematically scores certain groups lower, the system can be retrained or the confidence thresholds adjusted to route more cases to human review until the bias is corrected.