Knock-out criteria are binary pass/fail decisions applied to job applicants before any human judgement enters the process. Unlike ranked scoring, where a lower score still allows a candidate to progress, a knock-out removes them entirely. That finality is precisely what makes them high risk. A poorly designed cut-off can exclude large numbers of qualified candidates in a way that constitutes indirect discrimination under the Equality Act 2010: a neutral provision, criterion or practice that puts people with a protected characteristic at a particular disadvantage, unless it can be objectively justified as a proportionate means of achieving a legitimate aim.
This guide takes you through the complete workflow: choosing the right metrics, setting defensible thresholds, running an adverse impact screen, building an audit trail, and monitoring performance over time. By the end, you'll have a reviewed knock-out rule set with documented evidence and a monitoring plan rather than a set of filters that quietly exclude the wrong people.
Use a knock-out only when a criterion is genuinely job-relevant, verifiable from objective evidence, and narrowly tied to a real capability requirement. The classic examples are: a mandatory qualification required by law (e.g. an SIA licence for a security role), a right-to-work eligibility check, or a language proficiency requirement that is genuinely essential to the role duties.
If you're uncertain whether a requirement is essential, or if it could reasonably be assessed at a later stage through structured tests or interviews, use progressive screening instead: broad application sift, then structured assessment, then interview. Progressive screening preserves more candidates for evaluation and reduces the risk of inadvertently removing people who would have performed well.
Do not use knock-outs for:
Before you define a threshold, you need a baseline dataset. Collect applicant counts across the full funnel (application, sift, knock-out stage, advance/pass) over a defined time window. For high-volume roles, 90 days is a reasonable starting point. For lower-volume specialist hiring, you may need 12 months or more to accumulate usable numbers.
You also need to be able to calculate selection rates by relevant demographic groups to run an adverse impact screen. Handle protected attribute data carefully and lawfully under UK GDPR: collect it via an anonymised equal opportunities form rather than through the scored assessment itself, and keep it separate from scoring records.
Collect downstream outcome data too: hiring manager ratings at 90 days, training completion rates, time-to-productivity scores. These become your validation anchor. Without them, you cannot demonstrate that the knock-out actually predicts job success.
Run data quality checks before you start: identify missing responses, duplicate applications, and inconsistent scoring labels across shifts or locations.
Step 1. Translate job requirements into job-relevant measures. Start from a current job analysis or task inventory, not from a historic job description or a recruiter's intuition. Map each requirement to observable, measurable evidence.
Step 2. Decide the trigger type. An eligibility knock-out answers a binary factual question (does this person hold the required licence?). A scoring-based knock-out applies a numeric threshold to an assessment result. Both are legitimate, but scoring-based knock-outs carry higher statistical risk and require validation evidence.
Step 3. For scoring-based knock-outs, choose between a hard cut-off and a score band. A hard cut-off (e.g. reject everyone below 60 out of 100) is simpler to administer but more likely to produce adverse impact. A score band (e.g. flag everyone between 50 and 65 for a secondary assessment, auto-advance those above 65, and only reject those below 50) preserves borderline candidates for further review. Score bands are often a proportionate alternative to hard cut-offs.
Step 4. Build in an equivalency pathway: a documented route for a candidate to demonstrate they meet the underlying requirement through alternative evidence. This is not an ad hoc discretion door; it should be a defined process with clear criteria, applied consistently.
Step 5. If any part of your knock-out depends on assessor judgement (e.g. a scored telephone screen), run a calibration exercise before go-live. Have two or three scorers independently assess the same set of practice responses and reconcile discrepancies.
Step 6. Define your appeal and review pathway. Any candidate removed by an automated knock-out must have a route to request human review, particularly where a disability adjustment may be relevant. Under UK GDPR Article 22C, solely automated decisions with legal or similarly significant effects on individuals attract additional safeguards, including the right to obtain human review and to contest the decision. The ICO's guidance on fairness in the AI lifecycle (published March 2023) sets out how organisations should provide meaningful information about the logic involved in automated decisions.
The four-fifths rule (80% rule) is a practical screening benchmark, not a legal test in its own right. According to the EEOC's Uniform Guidelines on Employee Selection Procedures, a selection rate for a protected group that is less than 80% of the rate for the group with the highest selection rate generally indicates adverse impact and warrants investigation.
Worked example:
Suppose you're running a numeric skills assessment with a cut-off of 60 out of 100 for a warehouse operative role. In a recent cohort of 400 applicants:
| Group | Applied | Passed cut-off | Selection rate |
|---|---|---|---|
| Group A | 250 | 175 | 70.0% |
| Group B | 150 | 75 | 50.0% |
Group A has the highest selection rate at 70%. The four-fifths threshold is 80% of 70% = 56%. Group B's rate of 50% falls below 56%. This is a trigger for further review.
What to do when the four-fifths threshold is breached:
Sample size guidance: Small sample sizes can produce misleading results in both directions. A ratio calculated from 15 candidates per group is statistically unreliable. Your organisation should define a minimum N policy before go-live. A commonly applied internal policy is to require at least 30 applicants per group before running an adverse impact calculation, and to flag any result based on fewer than 50 per group as provisional pending further data. Set your own thresholds based on your hiring volumes and risk appetite, and document them.
Each knock-out rule needs a documented record containing:
Keep a change log. If a threshold is adjusted, version-control the record so you can show what was in place during any given hiring campaign.
Banding is the most common first remedy. Instead of a single cut-off at 60, create three zones: auto-advance (above 65), review band (50 to 65), and reject (below 50). Candidates in the review band complete a secondary, lower-cost screen before a final decision is made. This approach is consistent with proportionality under the Equality Act 2010: if a less restrictive measure achieves the same legitimate aim, you should use it.
Multiple-input knock-outs reduce reliance on a single metric. Combining a short skills test with a structured work sample, for example, can actually reduce adverse impact signals compared to either measure alone, because different assessments surface different evidence of capability. Selection Lab's role-specific assessment bundles are designed precisely for this: combining cognitive, skills-based, and situational assessments into a validated funnel rather than relying on any single filter.
Before rolling out a revised rule, test it retrospectively against your historical data. Run the revised four-fifths calculation and compare it to the original. Only proceed if the adverse impact signal has reduced to an acceptable level.
Document all accommodations and alternative evidence routes centrally. These must not be handled case by case without a record.
Set up a structured review cadence from day one:
Define trigger-based alerts: for example, if the selection rate ratio for any group drops below 85% in a rolling 30-day window, escalate for immediate review before the next hiring campaign continues.
Produce two reporting outputs: a one-page summary for HR leadership showing selection rates, override rates, and any flagged anomalies; and a technical appendix for compliance stakeholders showing the full calculation, data quality notes, and the change log.
Platforms that integrate knock-out logic with ATS workflows, such as Selection Lab's SmartChat intake tool, can surface override and reversal rates directly within your existing reporting dashboard. That kind of visibility makes it far easier to spot drift early and demonstrate to stakeholders that human review pathways are being used, not bypassed.
If your knock-out rules are applied automatically by an ATS or AI screening tool with no human in the loop before a rejection is issued, UK GDPR Article 22C applies. Candidates have the right to obtain human review of the decision, to express their point of view, and to contest the outcome. You must also be able to provide meaningful information about the logic involved in the automated process.
In practice, this means:
The ICO's fairness framework for the AI lifecycle (March 2023) is a practical reference point for how to structure these disclosures.
A knock-out criterion is defensible when it is:
Getting this right takes more upfront effort than setting a quick filter. But a knock-out rule that can't survive scrutiny isn't saving you time: it's creating legal exposure and removing candidates who may have been among your best hires.
Knock-out criteria are binary pass/fail decisions applied to job applicants before any human judgement enters the process. Unlike ranked scoring, where a lower score still allows a candidate to progress, a knock-out removes them entirely. That finality is precisely what makes them high risk. A poorly designed cut-off can exclude large numbers of qualified candidates in a way that constitutes indirect discrimination under the Equality Act 2010: a neutral provision, criterion or practice that puts people with a protected characteristic at a particular disadvantage, unless it can be objectively justified as a proportionate means of achieving a legitimate aim.
This guide takes you through the complete workflow: choosing the right metrics, setting defensible thresholds, running an adverse impact screen, building an audit trail, and monitoring performance over time. By the end, you'll have a reviewed knock-out rule set with documented evidence and a monitoring plan rather than a set of filters that quietly exclude the wrong people.
Use a knock-out only when a criterion is genuinely job-relevant, verifiable from objective evidence, and narrowly tied to a real capability requirement. The classic examples are: a mandatory qualification required by law (e.g. an SIA licence for a security role), a right-to-work eligibility check, or a language proficiency requirement that is genuinely essential to the role duties.
If you're uncertain whether a requirement is essential, or if it could reasonably be assessed at a later stage through structured tests or interviews, use progressive screening instead: broad application sift, then structured assessment, then interview. Progressive screening preserves more candidates for evaluation and reduces the risk of inadvertently removing people who would have performed well.
Do not use knock-outs for:
Before you define a threshold, you need a baseline dataset. Collect applicant counts across the full funnel (application, sift, knock-out stage, advance/pass) over a defined time window. For high-volume roles, 90 days is a reasonable starting point. For lower-volume specialist hiring, you may need 12 months or more to accumulate usable numbers.
You also need to be able to calculate selection rates by relevant demographic groups to run an adverse impact screen. Handle protected attribute data carefully and lawfully under UK GDPR: collect it via an anonymised equal opportunities form rather than through the scored assessment itself, and keep it separate from scoring records.
Collect downstream outcome data too: hiring manager ratings at 90 days, training completion rates, time-to-productivity scores. These become your validation anchor. Without them, you cannot demonstrate that the knock-out actually predicts job success.
Run data quality checks before you start: identify missing responses, duplicate applications, and inconsistent scoring labels across shifts or locations.
Step 1. Translate job requirements into job-relevant measures. Start from a current job analysis or task inventory, not from a historic job description or a recruiter's intuition. Map each requirement to observable, measurable evidence.
Step 2. Decide the trigger type. An eligibility knock-out answers a binary factual question (does this person hold the required licence?). A scoring-based knock-out applies a numeric threshold to an assessment result. Both are legitimate, but scoring-based knock-outs carry higher statistical risk and require validation evidence.
Step 3. For scoring-based knock-outs, choose between a hard cut-off and a score band. A hard cut-off (e.g. reject everyone below 60 out of 100) is simpler to administer but more likely to produce adverse impact. A score band (e.g. flag everyone between 50 and 65 for a secondary assessment, auto-advance those above 65, and only reject those below 50) preserves borderline candidates for further review. Score bands are often a proportionate alternative to hard cut-offs.
Step 4. Build in an equivalency pathway: a documented route for a candidate to demonstrate they meet the underlying requirement through alternative evidence. This is not an ad hoc discretion door; it should be a defined process with clear criteria, applied consistently.
Step 5. If any part of your knock-out depends on assessor judgement (e.g. a scored telephone screen), run a calibration exercise before go-live. Have two or three scorers independently assess the same set of practice responses and reconcile discrepancies.
Step 6. Define your appeal and review pathway. Any candidate removed by an automated knock-out must have a route to request human review, particularly where a disability adjustment may be relevant. Under UK GDPR Article 22C, solely automated decisions with legal or similarly significant effects on individuals attract additional safeguards, including the right to obtain human review and to contest the decision. The ICO's guidance on fairness in the AI lifecycle (published March 2023) sets out how organisations should provide meaningful information about the logic involved in automated decisions.
The four-fifths rule (80% rule) is a practical screening benchmark, not a legal test in its own right. According to the EEOC's Uniform Guidelines on Employee Selection Procedures, a selection rate for a protected group that is less than 80% of the rate for the group with the highest selection rate generally indicates adverse impact and warrants investigation.
Worked example:
Suppose you're running a numeric skills assessment with a cut-off of 60 out of 100 for a warehouse operative role. In a recent cohort of 400 applicants:
| Group | Applied | Passed cut-off | Selection rate |
|---|---|---|---|
| Group A | 250 | 175 | 70.0% |
| Group B | 150 | 75 | 50.0% |
Group A has the highest selection rate at 70%. The four-fifths threshold is 80% of 70% = 56%. Group B's rate of 50% falls below 56%. This is a trigger for further review.
What to do when the four-fifths threshold is breached:
Sample size guidance: Small sample sizes can produce misleading results in both directions. A ratio calculated from 15 candidates per group is statistically unreliable. Your organisation should define a minimum N policy before go-live. A commonly applied internal policy is to require at least 30 applicants per group before running an adverse impact calculation, and to flag any result based on fewer than 50 per group as provisional pending further data. Set your own thresholds based on your hiring volumes and risk appetite, and document them.
Each knock-out rule needs a documented record containing:
Keep a change log. If a threshold is adjusted, version-control the record so you can show what was in place during any given hiring campaign.
Banding is the most common first remedy. Instead of a single cut-off at 60, create three zones: auto-advance (above 65), review band (50 to 65), and reject (below 50). Candidates in the review band complete a secondary, lower-cost screen before a final decision is made. This approach is consistent with proportionality under the Equality Act 2010: if a less restrictive measure achieves the same legitimate aim, you should use it.
Multiple-input knock-outs reduce reliance on a single metric. Combining a short skills test with a structured work sample, for example, can actually reduce adverse impact signals compared to either measure alone, because different assessments surface different evidence of capability. Selection Lab's role-specific assessment bundles are designed precisely for this: combining cognitive, skills-based, and situational assessments into a validated funnel rather than relying on any single filter.
Before rolling out a revised rule, test it retrospectively against your historical data. Run the revised four-fifths calculation and compare it to the original. Only proceed if the adverse impact signal has reduced to an acceptable level.
Document all accommodations and alternative evidence routes centrally. These must not be handled case by case without a record.
Set up a structured review cadence from day one:
Define trigger-based alerts: for example, if the selection rate ratio for any group drops below 85% in a rolling 30-day window, escalate for immediate review before the next hiring campaign continues.
Produce two reporting outputs: a one-page summary for HR leadership showing selection rates, override rates, and any flagged anomalies; and a technical appendix for compliance stakeholders showing the full calculation, data quality notes, and the change log.
Platforms that integrate knock-out logic with ATS workflows, such as Selection Lab's SmartChat intake tool, can surface override and reversal rates directly within your existing reporting dashboard. That kind of visibility makes it far easier to spot drift early and demonstrate to stakeholders that human review pathways are being used, not bypassed.
If your knock-out rules are applied automatically by an ATS or AI screening tool with no human in the loop before a rejection is issued, UK GDPR Article 22C applies. Candidates have the right to obtain human review of the decision, to express their point of view, and to contest the outcome. You must also be able to provide meaningful information about the logic involved in the automated process.
In practice, this means:
The ICO's fairness framework for the AI lifecycle (March 2023) is a practical reference point for how to structure these disclosures.
A knock-out criterion is defensible when it is:
Getting this right takes more upfront effort than setting a quick filter. But a knock-out rule that can't survive scrutiny isn't saving you time: it's creating legal exposure and removing candidates who may have been among your best hires.
