An open position at a mid-sized company draws 20 to 80 applications on average; in recruiting agencies and staffing-provider pools, several hundred quickly pile up per wave. Manual CV review realistically takes 30 to 60 seconds per profile. At 500 applications, that is 4 to 8 hours just for the pre-sort. With a CV screening agent in Corporate LLM, you pre-sort the same 500 applicants in 5 minutes — with greater accuracy than manual review.
Overview: A CV screening agent derives a rubric from a stored scorecard. The scorecard fixes knockout criteria and weighted proxy criteria. The agent applies this rubric identically to every CV. The output is a sorted ranking with tier bands A, B, and C plus a separate knockout section. Your HR team saves up to 90 percent of the time spent on the screening process.
| Who it's for | Staffing providers, recruiting agencies, and any SME with more than 50 applications per month |
| Use cases | Initial screening, working through the backlog, re-scoring after a scorecard update |
| Outcome | A sorted list of all applicants in 5 minutes, with tier bands A/B/C and a knockout section |
How long does manual CV screening of 500 applications take in a mid-sized company?
The person responsible for recruiting reads CV after CV, usually on the side between interview slots and day-to-day HR work. At 500 applications, that adds up to 4 to 8 hours just for the pre-sort — before the first phone call is made.
How much time actually goes into each profile is disputed: recruiters estimate themselves at 2 to 3 minutes per CV, while a 2018 eye-tracking study by Ladders measured an initial scan of just 7.4 seconds. Realistically, effective processing time lands at 30 to 60 seconds per profile, depending on profile quality, the reviewer's form that day, and how closely the CV matches the role. What is not disputed: the pre-sort devours valuable HR time before a single interview has taken place.
The actual job: a fair, evidence-backed pre-selection of the top profiles.
The reality: the evaluation grid applied to CV #1 is no longer the same by CV #500 (fatigue), and knockout justifications are often gut feeling rather than facts.
Is ChatGPT suitable for CV screening? Three GDPR and consistency traps
The obvious reflex in recruiting: open ChatGPT, paste in the CV, type "evaluate this candidate against our job posting". Three problems immediately disqualify this setup for mid-sized B2B companies.
- The GDPR trap. Applicant data ends up on US servers — no DPA, no EU hosting, no audit trail. That is not "somewhat problematic"; it is a violation of Art. 28 GDPR plus a third-country transfer.
- The consistency trap. Without a stored scorecard, the LLM invents the evaluation criteria from the first CV and applies them inconsistently to the rest. Whoever sits second in the stack is judged by different standards than the first.
- The bias trap. Name, gender, age, photo, nationality, address, and marital status all travel into the prompt. The unconscious bias of manual review isn't eliminated — it is automated and reproduced.
So what mid-sized companies lack is not AI licenses. It is a system of platform, knowledge management, and scorecard that structurally rules out these three traps.
How does a CV screening agent with a scorecard and an LLM work?
The agent in our production setup runs on Claude Opus 4.8 at temperature 0.1. Deliberately deterministic: the task demands comparability, not creativity. Exactly one file is stored as knowledge: the scorecard for the role, as Markdown. The system prompt enforces two phases. Phase 1 derives the rubric before the first CV. Phase 2 scores every CV against the fixed rubric.
A scorecard defines three things:
- Knockout criteria (binary, hard): formal minimum requirements such as education, minimum years of experience, hands-on software experience.
- Proxy criteria with weights (summing to 100 percent): the role-critical skills that are visible in a CV. Skill stack, project scope, tenure, industry routine.
- Interview-only competencies (not assessable from a CV): communication, resilience, cultural fit. The agent deliberately ignores these instead of guessing.
Build your own CV screening agent: 4 steps to your own AI recruiting tool
- Write the scorecard. The job posting is not enough. You define the mission of the role and 5 outcomes after 6 to 9 months. Then the competencies with behavioral anchors (1 / 3 / 5) and weights. Separate knockouts from flags: flags are pointers for the manual review, not exclusions.
- Upload the knowledge file. You upload the scorecard as Markdown to Corporate LLM's knowledge management.
- Create the agent. Custom agent, Claude Opus 4.8, temperature 0.1. System prompt with phase 1 (derive the rubric) and phase 2 (score the CVs). Make the bias rules and the output schema (Markdown table with rank, score, tier, justification) explicit. Link the scorecard as the only knowledge item.
- Smoke test (10 minutes). Upload three real CVs of different caliber. Check that the phase 1 rubric is plausible and that every score justification is backed by the CV text.
Time to value: one morning. The scorecard outlives the vacancy. With 12 recurring role profiles in mid-market recruiting, you build a scorecard set once. After that, the agent goes live for every new posting in under 30 minutes.
CV screening with AI: before/after comparison — 4 to 8 hours vs. 5 minutes
The tier bands are A (score 75 and up), B (55–74), and C (below 55). Tier A goes straight to an interview slot, so the best applicants get the fastest response. You review tiers B and C manually before the final cut; both tiers remain sorted in the applicant pool with a documented justification per candidate.
| Metric | Before (manual) | With the CV screening agent |
|---|---|---|
| Time for 500 CVs | 4 to 8 hours | 5 minutes |
| Evaluation grid, CV #1 vs. CV #50 | Degrades (fatigue) | Identical rubric from phase 1 |
| Knockout justification per rejection | Gut feeling, ad hoc | Named criterion from the scorecard |
| Bias from name, address, photo, age | Possible, unconscious | Blocked via system prompt |
| Audit trail for works council or AGG inquiries | Notes, or nothing | Ranking + justification as Markdown |
Is AI CV screening GDPR-compliant? EU hosting, works councils, and deletion duties
Applicant data is sensitive. Three points matter in an SME setup:
- EU hosting and a DPA. Corporate LLM runs on EU infrastructure, with model routing through an EU layer. Uploaded CVs never leave your account to enter model training. The DPA is in German — one your data protection officer can actually work with.
- Works council co-determination. Using the agent falls under Section 87(1) No. 6 of the German Works Constitution Act (BetrVG) (technical equipment for monitoring behavior and performance). The agent delivers exactly the material a works council agreement requires: documented criteria, documented scoring logic, an audit trail per candidate. In many mid-sized setups, a works council agreement on the CV screening agent is achievable within 2 to 4 sessions. With larger works councils that have their own data protection committee, expect correspondingly longer.
- AGG obligations and deletion. After a position is filled, applicant data must be retained for 2 months to cover potential damages claims under Section 15 of the German General Equal Treatment Act (AGG); in practice, 6 months are kept to cover the period for filing claims. After that, the deletion obligation under Art. 17 GDPR applies to all applicant data — including the LLM-generated ranking and the stored chat conversation in Corporate LLM. You schedule the deletion routine centrally in your HR software; Corporate LLM content can be removed selectively per chat or space.
Try the CV screening agent in Corporate LLM for free
Want to see what the CV screening agent in Corporate LLM looks like in practice? Create a free account, upload your scorecard, start your first run.
Background on the platform decision:
LLM platform for the mid-market: 4 routes compared.
Frequently asked questions
How does CV screening with a Corporate LLM agent work?
You store the scorecard for the role once as knowledge and upload the CVs per chat. The agent first derives a rubric with knockout criteria and weighted proxy criteria from the scorecard, then applies it identically to every CV. The output is a complete ranking with tier bands A, B, and C, plus a separate section for applicants who miss a knockout criterion.
Where does the applicant data end up, and is this GDPR- and works-council-compliant?
Corporate LLM runs on EU infrastructure, with model routing through an EU layer. Uploaded CVs never leave your account to enter model training, and the data processing agreement (DPA) is in German. For works council co-determination under Section 87 of the German Works Constitution Act (BetrVG), the scorecard and documented bias rules give you the material your works council agreement requires: traceable criteria, documented scoring logic, an audit trail per candidate.
Why does the agent need a scorecard rather than just the job posting?
Without a scorecard, the LLM invents criteria from the first CV and applies them inconsistently to the rest. The scorecard fixes — before the first applicant — what counts as a knockout, which skills are weighted, and which points are only assessed in the interview. That way, every CV is measured against the same yardstick, not the yardstick of random ordering.
Does the pattern work for every role, or only for accounting?
For every role you can write a scorecard for. We have built CV screening agents for financial accounting, software engineering, service technicians, and inside sales. The effort shifts from the tool side to the scorecard: spend 60 to 90 minutes cleanly defining what A-level performance means in the role, and the agent around it is up and running in an hour.
Can the agent avoid bias?
The agent is instructed via system prompt to ignore name, gender, age, photo, nationality, address, and marital status. Every score must be backed by concrete evidence from the CV. That doesn't eliminate every bias, but it removes the factors that most often creep in unconsciously during manual review.
