Sorted shortlist with reasoning in minutes
What it does
The AI agent replaces the initial screening of incoming CVs — the recruiter's manual work on each application. Candidates arrive in HRIS already tagged with a relevance mark, a category, and a detailed comment. The recruiter sees not a stream but a shortlist — and spends their attention on people, not on filtering.
The basic process works as follows:
- A new resume lands in a monitored file storage folder (an attachment from a career landing page, a forwarded email, an export from a job board) or directly into HRIS via a built-in submission channel.
- The AI agent reads the document — PDF, DOCX, or a scan processed by OCR — and extracts structured fields: total experience, relevant experience, skills, education, location, language level, expected salary if stated.
- The agent checks the extracted data against a requirements rubric for the specific vacancy. The rubric is a pre-defined set of must-have and nice-to-have criteria, formal filters (visa, remote work, time zone), and role context.
- Each candidate is assigned a category (shortlist, reserve, rejection) and a rationale of 2–4 sentences is generated: what matched, what did not, and which gaps can be closed through training.
- The result is written to the candidate's card in HRIS: status, comment, tags, vacancy link, processing timestamp.
- The recruiter is notified only about shortlisted candidates; everything else remains in the database for manual review, re-search against future vacancies, or talent pool.
What automation does not do
- It does not make the final hiring decision. The AI agent sorts by formal and semi-formal criteria; the final choice remains with a human.
- It does not conduct interviews or assess soft skills, cultural fit, or motivation. These stages require live interaction and remain with the recruiter and hiring manager.
- It does not replace legal data verification. Diploma verification, references, background check, and GDPR compliance for CV storage are handled by a separate process.
How it works
Technically, resume screening is a bundle of four layers: ingest, extract, match, write-back. Each layer does one task and passes the result to the next, so a failure at one point does not break the entire pipeline.
Architecture
Layer | Purpose | Typical component |
|---|---|---|
Ingest | Receiving a new resume | File storage watcher or HRIS webhook |
Extract | Parsing the document into a structure | LLM with JSON-output + OCR for scans |
Match | Matching against the job rubric | AI agent on a language model |
Write-back | Writing to HRIS | REST API HRIS or middleware (low-code platform, Zapier) |
Implementation sequence
- Collecting baseline data. The team exports 50-150 recent resumes and the recruiter's decisions on them — who made the shortlist, who did not, and why. This becomes the calibration set for validating model quality.
- Describing the rubric for 1-3 positions. HR and the hiring manager jointly document must-have, nice-to-have, formal filters, and stop-factors. The rubric is stored in a separate document (Notion, internal wiki, a field in HRIS) and is versioned.
- Setting up the ingest channel. A source is selected: an incoming folder in file storage, an email alias with auto-forward, a webhook from HRIS or a job board. For a weekend implementation, one channel is sufficient — expansion is added later.
- Prompt engineering for the AI agent. A system prompt is written with the rubric, JSON-output format, and good/bad evaluation examples. The agent runs on the calibration set from step 1; results are compared against the recruiter's decisions. Discrepancies are discussed — either the prompt is adjusted or the rubric is refined.
- Write-back to HRIS. API integration is configured: creating or updating the candidate card, filling in the status, adding the agent's comment to the field tagged "AI assessment". If HRIS does not support a direct API, an intermediate layer is used — a workflow engine, Zapier, or middleware.
- Pilot. 1-2 weeks of shadow operation: the agent processes each new CV but does not send notifications to the recruiter. HR reviews the results at the end of the day. Discrepancies are logged.
- Go-live. After 2-4 iterations, when shortlist accuracy becomes acceptable for the team, the process goes into production. Notifications go to Slack, email, or directly to HRIS.
- Monitoring. During the first month, the recruiter manually reviews 10-20% of results. Discrepancies are a signal to adjust the rubric or the prompt.
What is stored where
The rubric — in a single document accessible to HR and the hiring manager. The prompt — in git or a config file of the workflow engine or middleware. Agent results — in HRIS in dedicated fields, to avoid mixing with the recruiter's manual notes. The calibration set is updated once a quarter or when the job profile changes.
Prerequisites
For a basic resume screening implementation, you need:
- An incoming resume channel. A folder in file storage (Google Drive, Dropbox, SharePoint, S3) with read access for the service account, or an HRIS with a webhook or candidate submission API.
- HRIS with an available API. BambooHR, Greenhouse, Workable, Hurma, Peopleforce, or equivalent. Fields for candidate status, comment, and an "AI assessment" tag are required.
- Documented rubrics for key positions. At least 1–3 roles: junior/middle/senior in one direction. Without a rubric, the AI agent does not know what to look for.
- Calibration set. 50–150 recent resumes with a recruiter decision (shortlist / reserve / reject) and a brief rationale.
- Responsible person within the team. One recruiter or HR manager is ready to invest 4–8 hours per week for the first 2–3 weeks on calibration and prompt adjustment.
Team readiness
The recruiter accepts AI agent results as a first filter, not as a final decision. The hiring manager participates in describing the rubric and is ready to return 2–3 weeks after launch to refine the criteria.
Timeline
Weekend complexity implies 2–7 days of setup: 1 day for collecting the calibration set and rubric, 1–2 days for the prompt and HRIS integration, 2–4 days for the pilot and iterations. A ready solution — by the end of the first working week.
Pain points
- Review — bottleneck
- Repetitive Routine Tasks
FAQ
How long does the launch take?
A basic launch fits within 2-7 days. Day one goes to collecting a calibration set of 50-150 resumes and writing the rubric. Days two and three — configuring the AI agent prompt and integrating with HRIS. The remaining days — a shadow pilot and 2-4 iterations on discrepancies. By the end of the first working week, the process is production-ready.
What if we don't have an HRIS?
Launching without an HRIS is realistic for small teams. A spreadsheet takes on the role of a candidate database — Google Sheets, Airtable, or Notion — with fields for status, comments, and tags. The AI agent writes results to the spreadsheet, and the recruiter works from it. This is an interim solution: when the flow grows to 50+ resumes per day, moving to a full HRIS becomes justified.
What can break?
Three main risk points. The first — a poorly written rubric: the agent filters out relevant candidates due to overly rigid formal filters. The second — quality drift when the job profile changes without updating the rubric. The third — unstructured resumes (scans without OCR, exotic formats). All three are controlled by manually checking 10-20% of results in the first month and a quarterly review of the calibration set.
Does this work in our industry?
Resume screening is universal across industries — IT, retail, manufacturing, logistics, services, nonprofit sector. Only the rubric differs: for IT, the stack and level matter; for manufacturing — certifications and experience with specific equipment; for retail — location and schedule. The AI agent is the same; industry-specific tuning goes into the rubric and calibration set.
How to reduce the risk of AI bias?
Bias is reduced by three practices. The rubric captures only professional criteria — no age, gender, photo, or name. The calibration set is checked for even distribution across demographic groups. The AI agent's results undergo a selective manual audit once a month. Complete elimination of bias is impossible — neither in AI nor in humans — but a controlled process noticeably reduces the risk of discrimination.
How to handle GDPR and candidate data storage?
The AI agent works with resumes under the same rules as a recruiter: access via a service account, logging of every document access, retention period matches the HRIS policy. For EU candidates, consent for automated processing is added to the application form. Rejected candidates' resumes are deleted on a schedule — after 6-12 months, in accordance with local legislation.
Want this in your business?
Book a free audit — we'll show how this automation will work for you.