Neha, 28: four hundred résumés, one role, and a screening tool she does not trust

A hiring manager who wants "someone like Arun", a job description written by a template, and a pile she cannot read in the time she has.

Planned for release phase 2. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.

In brief

The situation. Neha recruits for a Gurugram IT company. One role, four hundred applications, a hiring manager who describes the ideal candidate as "someone like Arun, basically", and a screening tool HR bought that ranks résumés in a way nobody can explain. She has read that such tools reproduce bias. She has also read the four hundred résumés herself, last time, and taken three weeks.

  1. Step 1 — Define the End State

    "A shortlist of eight in five days that the hiring manager interviews without complaint, chosen by criteria I can show to any candidate who asks why they were not chosen, with no criterion I could not defend in front of a labour court. The role description reads like the job, not like a template." Audiences: the hiring manager, the eight, the 392, and a lawyer.

  2. Step 2 — Map the Flow

    Her list: pin the manager down on what "like Arun" means, rewrite the JD, define screening criteria, screen, shortlist. AI review adds: a structured interview with the manager first (five questions, written down); a criteria document approved by the manager before any résumé is read — so the tool cannot be blamed or credited for judgement; and a bias check on the criteria themselves. It suggests letting the ranking tool do the first cut. She takes it in a smaller form: the tool may sort by verifiable facts only — years, skills named — never rank.

  3. Step 3 — Assign Tasks

    What the role actually needs: the manager's context, extracted by Neha. Writing the criteria: hers, approved by him. Extracting verifiable facts from four hundred résumés into a table: AI, verifiable, low cost — the ideal task. Judging fit against criteria: hers, on the sixty that pass the facts. The rejection note: she writes it once, by hand.

  4. Step 4 — Select the Tool

    A document-and-data tool to extract facts to a table (input documents, output structured); a conversational tool for the JD rewrite; the company's ATS for records. The bought ranking tool is switched to "sort", not "rank".

  5. Step 5 — Prepare and Direct

    Criteria for the extraction: every field must be a fact stated in the résumé, never inferred; anything missing is blank, not guessed. For the shortlist: each of the eight is tied to the written criteria; each rejection reason is one of the criteria; no criterion is a proxy for age, gender, college or address.

  6. Steps 6 and 7 — Execute and Evaluate

    The extraction table has forty rows with inferred fields — "probably 5+ years" from a graduation date. The criterion says blank, not guessed; she re-runs with the brief tightened and spot-checks fifty by hand. The JD draft is a template with the company name swapped; she rewrites it from the manager's five answers. Sixty résumés pass the facts; she reads sixty, not four hundred.

  7. Step 8 — Integrate and Stress-Test

    Shortlist of eight to the manager. Small fix: one candidate's notice period missed. Structural: the manager rejects three of the eight as "not like Arun" — the criteria document, which he approved, does not contain whatever he means. A Step 1 problem: his end state was never actually extracted. She goes back with Arun's résumé and asks him to point. The word he lands on is "owned a release end to end". It goes into the criteria; the next eight hold.

What changed

Five days. Eight interviews. Every rejection tied to a written criterion she can show. The ranking tool ranks nothing; it sorts. Neha keeps the criteria document as the template for the next role — the thing she reuses is the thinking, not the tool.

What to take from this case

Let tools extract facts and never infer them — "blank, not guessed" is the rule. Get the criteria approved before any résumé is read. "Someone like X" is an end state that has not been written down yet; make them point at the line.

A composite case, not a client engagement; the person is invented, the situation is not.

← Suresh, 48: a bank branch, new digital targets, and a team that has stopped listeningHarish, 34: a construction site, forty vendors, and a daily report written at 11 p.m. →