Diversity and inclusion in hiring: what assessment can and cannot do

What assessment is actually being asked to do

A hiring process is a sequence of filters, and assessment is one of them. It sits after sourcing and before the offer, which places a hard boundary on its influence: it can only choose among the people who applied, and it can only reward the things the role was defined to need. If neither of those is right, a better-designed exercise in the middle changes the ranking without changing the outcome that matters.

This is worth saying plainly, because "we introduced structured assessment" is frequently offered as an answer to a question about representation. It is at best a partial answer. The honest version is narrower and more defensible: structured assessment reduces how much a decision depends on signals that correlate with someone's background rather than their capability, and it produces a record that makes the remaining reasoning inspectable.

That narrower claim is still a real one. Most of the discretion in a conventional process is concentrated in exactly the places where background does its work — CV screening and the unstructured conversation. Taking discretion out of those two places is not everything, but it is not nothing either.

Proxies are the thing to displace

A proxy is a signal used to stand in for capability because capability was not directly observed. University name, employer prestige, an uninterrupted career history, a referral from someone already inside, fluency in the interview genre — each is treated as evidence of skill, and each is distributed by things other than skill: where someone grew up, what their family could pay for, who they already knew, whether they had a period of caring or illness, whether English is their first language.

A work sample displaces proxies by making the thing itself observable. If the job involves defusing an angry customer, the candidate defuses an angry customer, and the score comes from what they wrote rather than from where they wrote it before. The candidate with the unusual path and the candidate with the expected one are answering the same prompt, and neither is being credited for a signal that was never about the work.

Public-sector guidance on assessment methods makes the same point from the other direction: methods differ in how directly they sample job content, and the more directly a method samples the work, the less it depends on inference from background. That is the whole mechanism. It is not sophisticated, and it is not a claim about any particular product.

• Name each proxy your current process rewards, in writing, before replacing it.
• Ask what the task would look like if a career-changer with no relevant employer could still do well on it.
• Strip identifying context from the material evaluators see when the task does not need it.
• Check that the exercise does not silently reward interview fluency in a role that does not require it.

Structure is the mechanism, not the tool

The active ingredient is structure: the same task, in the same form, scored against criteria that were written down before the first candidate applied, by evaluators who score criterion by criterion rather than forming an overall impression first. None of that requires software. Software makes it cheap to hold, and hard to quietly abandon on a busy week.

The reason ordering matters is that unwritten criteria adapt. When the rubric is decided afterwards, a strong impression finds its justification and a weak one finds its objection, and the reasons feel entirely genuine to the person giving them. Writing the criteria first does not make an evaluator neutral. It makes their judgement checkable — a specific claim about a specific piece of work that someone else can read and dispute.

The same applies to disagreement. When two evaluators split on a candidate, structure turns that into a usable question — which criterion, and why — rather than an average of two impressions that were never about the same thing. Where enough evaluators score enough shared candidates, agreement can be computed rather than assumed. AI can summarise a long response or point a reviewer at the part of a transcript worth reading; it does not produce a decision, and it should never be the thing that decides.

What it cannot do: a biased job definition stays biased

An assessment measures what the blueprint told it to measure. If the role description asks for five years in a named industry, a degree the work does not require, or availability that quietly excludes anyone with caring responsibilities, then a perfectly structured exercise will faithfully select for those things and produce clean, defensible-looking evidence for a narrowed pool.

This is the most common way a fairness project fails. The instrument is scrutinised carefully and the requirements it serves are not scrutinised at all, because they were written months earlier by someone who has since moved on. The requirements are where the leverage is: every criterion in a rubric should trace back to something the role genuinely needs, and any requirement that cannot be traced to the work should be argued for explicitly or removed.

Nor can assessment reach what happens on either side of it. It does not change who saw the advert, whether the advert's language deterred people, whether the pay band excluded them, or whether the people hired are supported and promoted afterwards. A process that assesses fairly and retains poorly has not solved anything; it has just moved the loss further down the line.

Accessibility is part of fairness, not an exception to it

An exercise that is fair in principle can exclude in practice. A timed task rewards typing speed in a role where nobody is timed. A design that fights a screen reader excludes candidates who use one. A requirement to be alone in a silent room for ninety minutes is easy for some people and impossible for others, and heavy monitoring turns that into a hard condition rather than a preference.

Treat accommodation as a normal setting rather than an exception granted on request. Offer extended time in the invitation, keep the interface usable without a mouse or a large screen, and remove speed pressure wherever speed is not part of the job. Most candidates who need an adjustment will not ask for one, particularly at the stage where they feel least secure.

Match integrity controls to what the stage actually decides, for the same reason. Every control is paid for by every honest candidate, and it is not paid evenly: the cost of being watched falls hardest on people who have reason to expect scrutiny. A short screening exercise rarely justifies recording or lockdown, and any monitoring signal belongs in front of a human reviewer as context, never as an automatic rejection.

Where the legal frameworks sit — and where they do not

Selection procedures are regulated, and the regulation is jurisdictional. In the United States, federal enforcement guidance treats a test or selection procedure as something that must be job-related and consistent with business necessity, and the Uniform Guidelines on Employee Selection Procedures set out how selection rates across groups are examined. In Great Britain, the Equality Act 2010 frames the same territory through protected characteristics and the concepts of direct and indirect discrimination. Other jurisdictions differ again, sometimes substantially, and several now have specific rules about automated tools used in hiring.

The practical consequence is that group differences in selection rates are a question you should expect to be asked about — an analysis in its own right, with its own thresholds, sample-size caveats and remedies, which we cover separately in our post on understanding adverse impact. Structure and documentation do not answer that question for you; they put you in a position to answer it, because you can show what was asked, why it was asked, how it was scored and by whom.

None of this is legal advice, and no assessment method or platform can be described as making a hiring process compliant. Compliance depends on the role, the jurisdiction, the data and the decisions taken — take qualified local advice, and be suspicious of any vendor claim that skips that sentence.

Measuring whether it actually worked

A change to a hiring process should be checked rather than assumed. The useful checks are boring, and most of them do not require special data: has the composition of the pool that reaches the assessment stage changed, or only the composition of the pool that applies? Are candidates from non-traditional backgrounds progressing at the assessment stage, or only entering it? Do the criteria that drive rejections trace back to real job requirements?

Two cautions. Small groups produce unstable numbers, and a difference across a handful of candidates should prompt a look at the instrument rather than a conclusion about it. And demographic data — which would sharpen the analysis most — may only be collected in some places, under specific conditions and for stated purposes. Do not start collecting it because it would be analytically convenient.

The strongest position to be in is unglamorous: a defined set of requirements traced to the work, tasks that sample it directly, rubrics written in advance, scores that can be re-read against the evidence that produced them, and a record showing who decided what. That does not make a process fair by itself. It makes an unfair one visible, which is the only reliable way anyone has found to fix it.

Sources & further reading

• Prohibited Employment Policies/Practices — U.S. Equal Employment Opportunity Commission
• 29 CFR Part 1607 — Uniform Guidelines on Employee Selection Procedures (1978), eCFR
• Equality Act 2010: guidance — GOV.UK
• Other Assessment Methods (work samples, situational judgement tests) — U.S. Office of Personnel Management