Define Must-Have Screening Criteria
An agent can only screen fairly on criteria that are written down and validated. Learn to turn a job into must-have rules the agent can apply.
Screening rules must be explicit and job-related
Experienced recruiters screen on instinct built from hundreds of placements. An agent has no instinct. It can only apply the rules you write down, so before any screening agent goes live, you have to turn the fuzzy sense of a good candidate into concrete, validated criteria. This is also where bias and compliance risk live, so it deserves more care than any other document in your agent build. Get this wrong and you do not just lose good candidates, you build a machine that loses them consistently, at volume, with a paper trail.
Separate must-haves from nice-to-haves
For any role, split requirements into two lists with very different powers. Must-haves are criteria that genuinely disqualify a candidate if missing, and even then the agent only flags, it never rejects. Nice-to-haves make a candidate stronger but must never gate anyone. A working split for a warehouse shift lead role:
| Criterion | Type | How the agent checks it | Defensible because |
|---|---|---|---|
| Legal right to work in location | Must-have | Direct screening question | Legal requirement of the job |
| Valid forklift certification | Must-have | Question plus document on file | Required to operate equipment |
| Available for night shifts | Must-have | Direct screening question | Core schedule of the role |
| 2+ years supervising a team | Nice-to-have | Scored, never gated | Trainable; gating it excludes strong internal-promote profiles |
| Local to the site | Nice-to-have | Noted for recruiter | Commute is the candidate's call, not a gate |
Validate every rule before it gates anyone
A rule nobody checked can quietly filter out good candidates and create legal exposure. For each proposed must-have, run three tests:
- The job test: is this truly required to do the job on day one, or is it a habit inherited from old job specs? "Degree required" fails this test for most roles.
- The regulator test: would you defend this filter to a client or a regulator as job-related and consistently applied? This is general good practice in the spirit of equal-opportunity rules such as EEOC guidance in the US, not legal advice; check your own jurisdiction with counsel.
- The back-test: run the rule against your last 20 placed candidates in similar roles. If your best placements would have been flagged out, the rule is wrong, not the candidates.
Write rules the agent can actually detect
Vague criteria produce vague screening. Each must-have becomes a checkable statement with a defined unknown path. The unknown path is the most important line: when the agent cannot verify a criterion, it must flag for a human, never assume a fail. A rule sheet the agent can consume:
ROLE: Warehouse Shift Lead - Site NW2
MUST-HAVES (flag if not met or unknown; NEVER auto-reject):
1. right_to_work = yes source: screening answer
2. forklift_cert = valid source: answer + document
3. night_shift_ok = yes source: screening answer
UNKNOWN HANDLING: any criterion unanswered or ambiguous
-> status "unknown", route to recruiter queue with the
candidate's exact words attached.
PROHIBITED AS INPUTS: age, gender, nationality, name-derived
inferences, address-derived proxies, employment gaps as a
standalone negative, school names as prestige signals.
NICE-TO-HAVES (score 0-2 each, never gate):
- supervision experience, WMS familiarity, local to site
The prohibited-inputs block matters. Models will happily infer from names, addresses, and gaps unless you explicitly forbid it, and those inferences are exactly where discrimination risk concentrates.
The human-review gate is not optional
Even a fully validated must-have list does not earn the agent the right to reject anyone. The agent sorts candidates into "meets all must-haves", "misses one or more", and "unknown", and a recruiter reviews the misses before any rejection goes out. Candidates misread their own situations, mis-tick forms, and phrase things oddly. The recruiter review of flagged candidates is cheap; a wrongful automated rejection is not.
Decision: gate, score, or note?
| The criterion is... | Treat it as |
|---|---|
| Legally or physically required to do the job | Must-have: agent flags misses for human review |
| Strong predictor but trainable or substitutable | Nice-to-have: scored, visible, never a gate |
| Subjective, like "strong communicator" | Recruiter note only; the agent cannot judge it reliably |
| A proxy for a protected characteristic | Prohibited input; remove it entirely |
| Unverifiable from available data | Unknown path: route to recruiter, never assume |
Criteria are living documents, not launch documents
The must-have list you validate today will drift out of date. Clients change what they actually need, markets change what is available, and your own placement data keeps accumulating evidence about which criteria predict success. Put a review rhythm on the rule sheet: after every five placements against a role family, or quarterly at minimum, re-run the back-test with the new placements included. Two patterns to look for. First, a must-have that flagged many candidates whom recruiters then advanced anyway on review - that gate is miscalibrated, and the recruiters' override behavior is telling you so. Second, a nice-to-have that turns out to separate successful placements from failed ones far better than expected - it may deserve more scoring weight, though be slow to promote anything to a gate. Version the rule sheet with dates and reasons, because when a candidate or a client asks why a criterion changed, "here is version 1.2, here is the placement evidence that drove it" is a much better answer than a shrug.
One more discipline: when a client hands you a job spec with requirements baked in, do not import their list as your must-haves unchanged. Clients pad specs with wishes, and some client wishes fail the regulator test outright. Run their list through the same three tests, push back on the gates you cannot defend, and document that you did. Your agent screening on an indefensible client requirement is still your agent and your liability, and "the client asked for it" has never impressed anyone whose job is to check fairness.
Writing the screening questions themselves
A validated criterion still fails if the question asked about it is ambiguous, because the answer arrives ambiguous and the agent has to interpret - which is exactly the thing you were trying to avoid. Question design is the last mile and it is where most screening quality is actually won.
Four rules, with the same criterion written badly and then well:
CRITERION: available for night shifts
Bad: "Are you flexible with hours?"
-> "Yes" tells you nothing. Flexible how? Which nights?
The candidate and the agent mean different things.
Good: "This role is 10pm to 6am, Sunday to Thursday.
Can you work those hours? (Yes / No / Some of them)"
-> The hours are stated, so the answer is about the
actual job. The third option catches the honest
middle instead of forcing a wrong Yes.
CRITERION: valid forklift certification
Bad: "Do you have forklift experience?"
-> Experience is not certification. A candidate answers
truthfully and the record is wrong.
Good: "Do you hold a current, in-date forklift licence?
(Yes - I can upload it / Yes - expired / No)"
-> Names the artefact, asks about currency, and
separates "had one once" from "have one now".
- State the requirement inside the question. Candidates cannot answer accurately about a schedule or a site they have not been told about, and an inaccurate answer is worse than no answer.
- Offer a middle option. Forced binaries produce false positives from candidates who want the job and false negatives from candidates being careful. The middle option routes them to a person, which is where they belonged.
- Ask about the artefact, not the aura. "Do you hold a current licence" is verifiable. "Are you experienced" is not.
- Keep the list short. Every additional gating question is another chance to lose a good candidate to a misread, and completion falls as the form grows. If your must-have list has grown past a handful, most of what you added is a nice-to-have wearing a gate's clothing.
One more: ask the perishable questions every time, even for a candidate already in your database. Availability and salary expectation decay in weeks, as the previous lesson set out, and a screening agent re-collecting them is doing your data hygiene for you as a side effect.
Scoring the nice-to-haves without them becoming gates
Nice-to-haves need to influence order without influencing eligibility, and the way to keep that honest is to make the score visible and the reasoning attached. A worked sheet for the shift lead role:
NICE-TO-HAVES (0-2 each, 6 max) - AFFECTS ORDER ONLY
Supervision experience
0 = none stated
1 = informal (covered for a supervisor, led a shift)
2 = held a supervisory role for 12 months or more
WMS familiarity
0 = not stated
1 = used any warehouse management system
2 = used the system on this site, or a named equivalent
Commute realism
0 = over 60 min or not stated
1 = 30-60 min
2 = under 30 min or on a direct route
OUTPUT FORMAT (this exact shape, every candidate):
score: 4/6
reasons:
supervision: 2 - "shift supervisor at [employer], 2 years"
wms: 1 - "used Manhattan at a previous site"
commute: 1 - "38 minutes, stated own transport"
NOT a ranking of people. A ranking of evidence available today.
Three properties make this safe. The score is bounded and small, so it cannot swamp a recruiter's judgment. Every point comes with the candidate's own words attached, so a recruiter can disagree with the evidence rather than the number. And the sheet is explicit that a low score is not a rejection - it is a queue position, and queue position is reversible in a way a rejection is not.
Watch for the failure mode where a score quietly becomes a gate because nobody ever looks below a threshold. If your recruiters only ever open the top ten, you have built an auto-reject with extra steps. The fix is procedural, not technical: require that flagged and low-scoring candidates are reviewed too, and check that they are.
Auditing the agent's screening, monthly
Rules drift, models update, and job specs get edited by people who did not read this lesson. A short recurring audit catches all three. It takes an hour a month.
- Pull twenty screened candidates at random from the last month, not the ones anyone flagged. Random is the whole point - you are looking for what nobody noticed.
- Re-screen them yourself against the rule sheet without looking at the agent's verdict first. Then compare. Any disagreement is either a rule that is unclear or a question that is ambiguous, and both are fixable.
- Read the flagged pile. How many flags did recruiters override on review? A high override rate on one criterion means that gate is miscalibrated - the recruiters are telling you so with their behaviour, which is more reliable than asking them.
- Check the unknown path is being used. If almost nothing is landing in "unknown", the agent is probably resolving ambiguity by guessing rather than routing it. That is the most dangerous quiet failure in the whole system, because it looks like efficiency.
- Look at outcomes, where you have them. Of the candidates who were flagged out and then advanced by a recruiter anyway, how many were placed and stayed? Every one of those is evidence that a gate is wrong.
Write the result down each month, even when it is boring. A dated audit trail showing that you check your own screening is worth a great deal the first time a client or a candidate asks how the process works - and it is not something you can produce retrospectively.
What to tell candidates
Candidates increasingly ask whether AI is involved, and some jurisdictions have begun regulating automated employment decision-making directly - New York City and the EU AI Act are both live examples, and the details vary by location and change, so check what applies to you with someone qualified rather than relying on a course page.
Regardless of what is required where you operate, the defensible position is the same and it is not hard: be able to say what the system does, and make sure the sentence is true. If your setup follows this lesson, the true sentence is a good one - an agent collects structured answers and checks them against written, job-related requirements, and a recruiter reviews every case where a requirement is not met before any decision is made. Nobody is rejected by software.
That sentence only stays true if the human review is real. The moment a recruiter starts clicking through the flagged queue without reading it, you no longer have the process you are describing, and the audit above is how you find that out before somebody else does.
The back-test, worked with the arithmetic shown
The back-test is the most important of the three validation tests and the one most often skipped, because "run it against your last twenty placements" sounds like a project. It is about ninety minutes with a spreadsheet. Here it is done, on the warehouse shift lead role, against a proposed must-have that a client asked for: "2+ years in a supervisory role". The dataset is illustrative; your own placements are the only ones that count.
Step 1: list the placements, not the applicants. Twenty people you actually placed into comparable roles, oldest first. Applicants tell you about your funnel; placements tell you about your judgment, and it is your judgment you are trying to encode.
Step 2: for each, record only what was knowable at screening time. This is the discipline that makes the test honest. Not what you learned at interview - what was on the application or in the screening answers, because that is all the agent will ever have.
Step 3: apply the proposed rule and count.
| Group | Count | What it tells you |
|---|---|---|
| Placed, and would have passed the rule | 11 | The rule is not useless - it does describe most of your hires |
| Placed, but would have been flagged out | 6 | The finding. Nearly a third of your successful placements fail this gate |
| Placed, and the rule was unknowable at screening | 3 | These would have landed in the unknown queue - a cost, not a failure |
Step 4: read the six. This is where the actual learning is, and it is qualitative. In this worked case the six break down as: four were internal promotes with informal team-lead responsibility that never appeared as a job title, one had two years of supervision under a title that said "senior operative", and one had eighteen months and was placed anyway because the client liked them. Five of the six are the same problem - the rule tests for a job title, not for the capability the client actually wants.
Step 5: rewrite rather than abandon. The conclusion is not "drop the requirement". It is that supervision belongs in the scored column with a question that reaches the capability directly - "have you been responsible for directing the work of others, formally or informally? Tell us about it" - rather than in the gate column keyed off a title. That single change recovers five of the six.
Two rules of thumb for reading your own numbers. If more than roughly one in ten of your past successful placements would have been flagged out, the gate is wrong and you should not ship it. And if a gate flags nobody at all across twenty placements, it is not protecting you from anything - it is decoration that will one day catch someone unexpectedly, so either remove it or find out why it never fires.
Pushing back on a client's must-haves, in the words you actually send
The lesson says do not import a client's list unchanged. That is easy to agree with and awkward to do at 4pm on a Friday with a live vacancy, so here is the message, written out. Adapt the specifics; keep the structure, because the structure is what makes it land as service rather than obstruction.
Subject: Shift Lead spec - two things to confirm before we screen
Hi Dan,
Spec received, we're briefing candidates tomorrow. Two of the
requirements I want to confirm with you before they go in as
hard filters, because we apply them consistently and I don't
want to cut you off from people you'd want to see.
1. "2+ years in a supervisory role"
We looked at the last twenty people we placed into shift lead
roles. Six of them wouldn't have met this at application -
mostly internal promotes who'd been running a team without
the title. Two of those six are still in post after eighteen
months.
Suggestion: keep it as a strong preference, and we ask every
candidate to describe where they've directed others' work,
formally or informally. You still see the experienced ones
first; you also see the four or five good ones this would
have hidden.
2. "Must live within 30 minutes of the site"
We'll flag commute distance on every profile so you can see
it. We won't use it as a filter - how far someone will travel
is their decision, and screening on where people live tends
to correlate with things we're not allowed to screen on.
You'll have the information; you make the call.
Everything else in the spec goes in as written - right to work,
valid forklift certification, and night shift availability are
all hard requirements and we'll confirm each one directly with
every candidate.
Happy to talk it through if you'd rather. Otherwise I'll brief
as above and you'll have the first profiles Thursday.
Priya
Four things that message does deliberately. It opens with the work continuing, so it does not read as a delay. It brings evidence rather than an opinion - the back-test is doing the persuading, which is the real reason to have run it. It offers a substitute rather than a refusal on the requirement that is merely inherited, and holds firm on the one that is a genuine risk, with a one-line reason that is about fairness rather than about the client. And it ends with a default: silence means it proceeds. Requests for approval stall; statements of intent with an opt-out do not.
Send it in writing, always, and keep it. If a screening decision is ever questioned, the file containing "here is what the client asked for, here is what we advised, here is what we did and why" is worth more than any amount of recollection.
Troubleshooting a rule sheet that is not working
| Symptom | Cause | Fix |
|---|---|---|
| Almost everyone is flagged | Too many must-haves, or one gate keyed off something rarely stated | Count flags per criterion. One criterion is usually responsible for most of them |
| Almost nobody is flagged | The questions let candidates self-select into a pass | Re-read the question wording. "Are you available for night shifts?" invites a yes; "Which shifts can you work?" does not |
| The unknown queue is empty | The agent is resolving ambiguity by guessing | The most dangerous quiet failure here. Test it deliberately with a deliberately vague answer and confirm it routes |
| Recruiters override the same gate repeatedly | The gate does not match how the desk actually judges | Move it from gate to score. Their behaviour is better evidence than their stated opinion |
| Scores cluster tightly and separate nobody | Every candidate scores the same because the criteria are near-universal | Replace with criteria that actually vary across your applicant pool |
| Nobody below the top ten is ever opened | The score has silently become a gate | Procedural fix, not a technical one: require the flagged and low-scoring piles to be reviewed, and check that they are |
Common mistakes
- Letting the agent reject candidates on rules nobody validated, then discovering it screened out qualified people on a filter that was never actually required.
- Copying must-haves from an old job ad instead of testing them against real placements.
- Leaving the unknown path undefined, so the agent silently treats missing data as a fail.
- Forgetting the prohibited-inputs list and letting the model infer from names and gaps.
What you should be able to do now
- Split any role into must-haves, nice-to-haves, and recruiter-note criteria.
- Run the job test, regulator test, and back-test on every proposed gate.
- Write a rule sheet with sources, unknown handling, and prohibited inputs.
- Explain why the agent flags misses for review and never auto-rejects.
- Defend every must-have as job-related in one sentence each.
- Write screening questions that state the requirement and offer an honest middle option.
- Run the monthly screening audit and act on a high override rate as evidence a gate is wrong.
Where this goes next
That is the end of the free part. Together with the ATS audit, you now have the two documents every recruiting agent build depends on: a readiness report for the data, and a validated rule sheet for the screening. Write the rule sheet for one live role this week and back-test it against your last twenty placements. If your best placements would have been flagged out, you have learned something worth more than the rest of the course.
The paid lessons build the desk around those documents. The inbound screening agent end to end. Writing back to Bullhorn, Recruit CRM, or Greenhouse without corrupting the reporting your clients see. Candidate messaging that sounds like your firm. Interview scheduling, and shortlist packs clients actually read. Then the governance: keeping a named human accountable for every rejection, auditing screening consistency at volume, and handing the system over so the desk owns it rather than depending on whoever built it. Every lesson ships the templates and workflow files.
If you would rather have it built, that is project work we take on - see AI automation services, or start with the free automation audit, which includes a straight assessment of whether your ATS and your rules are ready. AI automation cost has the pricing picture, and the human in the loop entry is the short version of why the review gate is not negotiable.
Put your email in below and both free lessons arrive with the rule sheet template, the screening question set, and the ATS audit worksheet as editable files you can run on a real role on Monday.
Key takeaways
What to remember from this lesson
- Get this wrong and you do not just lose good candidates - you build a machine that loses them consistently, at volume, with a paper trail.
- Must-haves flag, they never auto-reject. Nice-to-haves are scored and visible but never gate anyone.
- Every proposed gate faces three tests: the job test (required on day one?), the regulator test (defensible as job-related and consistently applied?), and the back-test (would it have flagged out your last 20 good placements?).
- The unknown path is the most important line on the rule sheet. Unanswered or ambiguous means 'unknown' and a recruiter queue - never a silent fail.
- Models infer from names, addresses, and employment gaps unless you explicitly forbid it. The prohibited-inputs block is where discrimination risk concentrates.
- A client job spec is not a must-have list. Run their requirements through the same three tests - 'the client asked for it' has never impressed anyone checking fairness.
Try it yourself
Screen a candidate against must-haves without rejecting anyone
This is the rule sheet from the lesson, as code. Three must-haves, one fixture application, and three possible states per criterion: met, flagged, or unknown. The one thing the function must never do is reject.
Try it yourself
Find the bias-prone criteria in a client's filter
None of the filters below name a protected characteristic, which is exactly why they get shipped. Write the checker that flags the ones standing in for one. This is a smoke alarm and not a legal audit: it catches wording, not outcomes, and it is not legal advice.
Check yourself
The three tests, applied under pressure
A client insists on '5 years minimum' for a role where your last three successful placements had three years each. Which test does the requirement fail?
A candidate does not answer the right-to-work question at all. What should the agent record?
Which of these belongs in the prohibited-inputs block of the screening prompt?
Answer every question to mark.
Exercise
Back-test a rule sheet against your last 20 placements
- 1Pick one role you fill repeatedly and split its requirements into three columns: must-have, nice-to-have, and recruiter-note-only.
- 2For every must-have, write one sentence defending it as required to do the job on day one. Any you cannot defend in one sentence moves to nice-to-have.
- 3Run the back-test: take your last 20 placed candidates in similar roles and apply each must-have to their records. Count how many of them the rule would have flagged.
- 4Any rule that would have flagged a strong placement is wrong - fix the rule, not the candidates. Write down which rule and which placement caught it.
- 5Write the rule sheet in the machine-readable form from the lesson: each must-have with its source, the explicit unknown-handling line, the prohibited-inputs block, and the scored nice-to-haves.
- 6Write the prohibited-inputs block explicitly, including employment gaps as a standalone negative and school names as prestige signals.
- 7Take a real client job spec for this role and mark every requirement in it that fails the job test or the regulator test. Draft the pushback sentence for each.
- 8Date and version the sheet, and write the review trigger on it - every five placements or quarterly, whichever comes first.
You are done when Your rule sheet carries a date and version, every must-have has a one-sentence job-related defence, the back-test against 20 real placements produced a count of false flags with each traced to a named rule, and the sheet contains both an explicit unknown-handling line and a prohibited-inputs block.
Companion files
Screening rule sheet and validation tests
The must-have/nice-to-have split table, the machine-readable rule sheet with unknown handling and prohibited inputs, and the gate-score-note decision table - ready to fill in for your own role.
The free-preview pack includes the worksheet for this lesson for Define Must-Have Screening Criteria, plus the other free lessons of this course. No email needed.
Free preview pack - 5 KB zipCheck yourself
Answer before you reveal
A candidate leaves a must-have question blank. What does the agent do?Show answer
Sets the status to 'unknown' and routes to the recruiter queue with the candidate's exact words attached. It never assumes a fail - candidates misread their own situations and mis-tick forms constantly, and a wrongful automated rejection is expensive while a recruiter glance is cheap.
Recruiters keep overriding one gate and advancing the candidates anyway. What is that telling you?Show answer
The gate is miscalibrated. The override behaviour is evidence - re-run the back-test with the new placements included and either loosen the rule or demote it to a scored nice-to-have.
The client's spec says 'degree required'. Do you screen on it?Show answer
Not without running it through the three tests, and it fails the job test for most roles. Push back, document that you did, and screen on what the job actually requires - your agent screening on an indefensible client requirement is still your liability.
