# Halve candidate screening time: the AI build for recruitment firms Source: https://getstartedwithai.net/tutorials/ai-for-recruitment-firms-build Site: Get Started With AI Tagline: Build AI automations that actually run your business. Audience: solo professionals, freelancers, consultants and AI implementers in training Contact: hello@getstartedwithai.co.uk This Markdown companion is generated from the prerendered page for AI crawlers. # Halve candidate screening time: the AI build for recruitment firms Eight hours of consultant time a week disappears into the same triage: read the CV, hold it against the spec, decide if the candidate goes through. The build below moves the reading-and-sorting layer to a Custom GPT scored against a written rubric, with a one-paragraph reason per candidate and a personalised outreach line for the shortlist. Final shortlist decisions stay with the consultant. Built in 90 minutes; runs at £35 a month plus a fortnight of output review to get the bias and accuracy checks where they need to be. Disclosed to candidates via a one-paragraph privacy notice update. ## What you need - A ChatGPT Plus or Claude Pro account at £16 a month. - A Zapier paid plan and your existing ATS (Bullhorn, Vincere, JobAdder, Recruitly, Mercury or Salesforce-based). - The job spec template the firm uses today plus 30 anonymised CVs (15 strong, 15 borderline) for calibration. - A written privacy notice update disclosing AI use, ready for the firm legal sign-off. - 90 minutes uninterrupted, then 30 minutes per role on every shortlist after. ## Step 1: Write the rubric (20 minutes) A spec is not a rubric. The rubric is what makes the GPT score consistently. Skip this and the output reads as generic AI summary. - Hard requirements: must-haves the candidate cannot pass without (right to work, regulated qualification, named tech stack). - Strong indicators: 4 to 6 markers that the consultant looks for in a strong candidate (named employer, project type, years). - Weighting: hard requirements at 100 percent, strong indicators at 50 to 80 percent, soft fits at 20 percent. - Disqualifiers: explicit reasons to reject regardless of fit. ## Step 2: Build the screening Custom GPT (25 minutes) 1. Open ChatGPT, create a Custom GPT named CV Screening Rubric. 2. Upload the rubric, the spec template and the 30 calibration CVs with the consultant manual scores. 3. System prompt: you are screening UK CVs against a job rubric. UK English. Output: candidate ID, score out of 100, one-paragraph reason, suggested outreach line. Anonymise the candidate name and any contact detail; reference by ID only. 4. Add the rule: if a hard requirement is missing or unclear, mark the score at 50 or below. Never assume a qualification not stated on the CV. 5. Test against the 30 calibration CVs. Consultant agreement should land around 85 to 90 percent in the first run. ## Step 3: Wire the ATS into the GPT (20 minutes) 1. In Zapier, trigger on a new CV attached to a role in the ATS. 2. Action 1: anonymise the CV (strip name, address, photo). Action 2: send to the GPT with the rubric inserted. Action 3: write the score, reason and outreach line back to the candidate record. 3. Action 4: tag the role with the count of shortlisted candidates ready for the consultant to review. ## Step 4: The bias-audit fortnight (15 minutes setup) The fortnight of output review is what gets the build to production accuracy and where the bias mitigations land. 1. Add a daily Sheet log: every AI-suggested rejection with the rubric reason, the candidate ID and the fortnight reviewer. 2. The consultant reads every rejection across the fortnight. Where the AI rejected wrongly, log the reason and update the rubric. 3. After the fortnight, accuracy lands at around 95 percent and the bias audit log goes into the firm regulatory file. ## Step 5: First production roles Run the build on three live roles. Compare the AI shortlist against the consultant manual shortlist. The third role is where the build stops surprising the consultant and starts saving real time. ## What you ship A Custom GPT with the rubric, a Zap into the ATS, an anonymised review queue, a bias audit Sheet and a privacy notice paragraph for the candidate-facing disclosure. Cuts CV screening time by roughly half across the desk. ## What to build next Move into the talent acquisition build for the in-house side of the same workflow, layer the AI sales agent build on the BD side, or move into the WIRE Method to package the screening overlay across desks and offices. ## Related builds More tutorials in Industries. Each one runs in 60 to 90 minutes and ships a finished asset. ### Reclaim 8 hours a week as an accountant: the AI build for practice work 30 to 60 minutes. Solo Pro / Implementer. [Open tutorial](/tutorials/ai-for-accountants-practice-build) ### Cut 12 hours a week of legal admin: the AI build for solicitors and small firms 30 to 60 minutes. Solo Pro / Implementer. [Open tutorial](/tutorials/ai-for-lawyers-firm-build) ### Replace 6 hours a week of session admin: the AI build for coaches and therapists 60 to 90 minutes. Solo Pro. [Open tutorial](/tutorials/ai-for-coaches-therapists-build) ### Reclaim a day a week as a property agent: the AI build for letting and sales 60 to 90 minutes. Solo Pro / Implementer. [Open tutorial](/tutorials/ai-for-real-estate-agents-build) ## Questions people ask before they start ### Will AI replace recruiters? No. AI replaces the CV-triage and outreach-personalisation parts of the role, not the candidate conversation or the client management. Firms that adopt AI run more roles per consultant rather than reducing headcount. ### How accurate is AI candidate screening? Against a clear rubric, agreement with a consultant manual shortlist runs at 85 to 90 percent in the first month and rises to 95 percent after a fortnight of output review. The fortnight is what gets the build to production accuracy. ### What about bias in AI candidate screening? A real risk and the most important reason for the fortnight of output review. The mitigations: anonymise CVs on input, use objective criteria from the spec, log every AI-suggested rejection so a human can audit it. The build ships with the bias-audit prompt and the audit log template. ### Is AI for recruitment GDPR-safe in the UK? On Pro plans yes, with caveats. Inputs are not used to train models. The candidate has a right to know AI was used in the decision; the build ships with a one-paragraph disclosure that goes into the candidate-facing privacy notice. UK GDPR Article 22 limits fully-automated decisions; the build keeps a human on the final shortlist for that reason. ### Does this work with Bullhorn, Vincere or Salesforce-based ATS? Yes across the major UK recruitment platforms. Bullhorn and Vincere have Zapier connectors that fire on a new CV against a role. JobAdder and Mercury have webhook patterns the build picks up. Salesforce-based recruiting orgs have Make modules that fit the same flow. The Custom GPT layer is ATS-neutral.