Work Samples and Trial Tasks: Testing the Real Job Fairly

Work samples — asking candidates to do a slice of the real job — are one of the most intuitive and best-supported ways to choose between people. They are also one of the easiest to get wrong. Badly designed, they test free time rather than skill, exploit candidates’ unpaid labour, expose the company to intellectual-property disputes and quietly screen out carers and disabled applicants.

This guide, part of the Finding the Right Person hub, shows how to design a fair task: what the evidence says, which format fits which role, time limits and when to pay, who owns the output, adjustments, scoring rubrics and how to handle AI in take-home work. It sits between the interview question sets by role and the final hiring decision.

What the evidence says, read carefully

For two decades the most quoted figures in selection came from Frank Schmidt and John Hunter’s 1998 review, which put work-sample tests at the top of the table with a validity of 0.54. In 2022, Paul Sackett and colleagues published a re-analysis in the Journal of Applied Psychology arguing that earlier corrections for “range restriction” had been applied too generously, inflating many estimates. Their revised figure for work samples is about 0.33; structured interviews came out highest at about 0.42, and job-knowledge tests at about 0.40.

Three cautions keep these numbers honest. They are averages across many jobs and studies, not a promise about your next hire. The work-sample estimate rests on a smaller and older base of studies than the interview figures, so it is less certain. And the methods combine: a short, well-designed task followed by a structured conversation about it captures much of the value of both, which is why many strong hiring teams use the task as the material for an interview rather than as a standalone pass-or-fail test.

Work samples have two further advantages the validity numbers do not capture. Candidates tend to see them as fair, because the link to the job is obvious; and they give a realistic preview of the work, so people who would hate the job often withdraw on their own. Both reduce early attrition, which is where much of the cost of a bad hire sits.

Choose the format that fits the role

“Work sample” covers several formats, and the right one depends on what you need to see. The table summarises the main options and where they fit.

A useful rule is to use the lightest format that shows you the skill. If a 45-minute live exercise reveals how an analyst structures a messy dataset, a weekend take-home adds little except cost to the candidate. Reserve longer tasks for the final two or three candidates, when the stakes justify the investment on both sides.

Designing a task that tests the job, not the candidate’s free time

Start from the role scorecard and pick the one or two outcomes that matter most and are hardest to judge from interviews. Then design a task that produces evidence on exactly those outcomes, using material that resembles the real work but is not live client or company work. A finance task might use a simplified, anonymised trial balance with planted errors; an engineering task, a small self-contained feature on a sample repository; a marketing task, a brief for a fictional product in your category.

Time limits, pay and the law on unpaid trials

Candidates’ time is the hidden cost of work samples. Many hiring teams cap take-home tasks at two to four hours and make the cap real by asking for a short output (“no more than six slides”, “one page”, “a pull request under 200 lines”). Give at least several days’ notice and a window that includes a weekend, because employed candidates and carers cannot always clear a weekday evening.

Paying for longer tasks is increasingly common and sends a strong signal about how you treat people. A flat fee or a day rate for anything beyond about four hours removes the suspicion of free consulting and widens your pool. In some places it is also the law. HMRC’s guidance in the UK treats an unpaid trial as legitimate only where it genuinely tests suitability for a specific vacancy, and suggests that a trial lasting more than a day is likely to attract minimum wage “in all but very exceptional circumstances”. In Australia, the Fair Work Ombudsman says an unpaid trial is lawful only if it is necessary to demonstrate skills for the job, lasts no longer than needed, involves only tasks relevant to the role and is directly supervised.

Elsewhere, assume that a candidate who does real work on your premises, or produces work you use, is likely to be treated as working, with pay, insurance and possibly social-security consequences. In countries where you hire through an employer of record or engage contractors, a paid trial may need a short contract of its own. When in doubt, keep tasks simulated, short and clearly separate from production work.

Who owns the work, and why you must say so

By default in most legal systems, the person who creates a piece of work owns the copyright in it unless they are an employee creating it in the course of employment or have signed it over. A candidate is neither. If you use a candidate’s campaign idea, code or analysis without agreement, you risk a copyright or confidentiality claim and, more likely, very public reputational damage.

The fix is simple. Use fictional or historical material, so nothing the candidate produces is directly usable. State in the task brief that the candidate retains ownership of their work, that you will use it only to assess their application and that you will delete it after the process. If you genuinely want to use a submission, offer to buy it on agreed terms. If you need candidates to see confidential information to do the task properly, use a short, mutual confidentiality note — not a one-sided agreement that leaves them carrying all the risk.

Accessibility and adjustments

Timed tasks, live exercises and complex instructions can all disadvantage disabled candidates, and in many countries employers must make reasonable adjustments to the process. Under the UK’s Equality Act 2010 that duty applies to recruitment, and similar duties exist under the Americans with Disabilities Act, EU equality law and Australian anti-discrimination law. Common adjustments are extra time, a different format (written instead of live, or the reverse), assistive technology, a quiet room, rest breaks and plain-language instructions.

Ask every candidate whether they need adjustments when you send the task, not only those who disclose a condition, and make the request easy — a single line in the invitation email. Keep the request separate from the people scoring the task. The wider principles are in our guides to candidate experience and reducing bias in hiring.

Scoring rubrics that make submissions comparable

A work sample without a rubric is an unstructured interview on paper: the scorer reacts to the overall impression and calls it judgment. Write the rubric before you see any submission. Use three to five criteria drawn from the scorecard, a 1–5 scale with written anchors for 1, 3 and 5, and a weight for each criterion. Two people score independently, ideally with names removed, and only then compare.

Follow every take-home task with a 20–30 minute conversation in which the candidate walks you through their choices: what they tried, what they rejected, what they would do with another day. This is where real understanding shows, and it neutralises most concerns about outside help. If you use assessment platforms for the task itself, our comparison of pre-employment assessment software covers the main options.

AI in take-home tasks: set the policy, then test understanding

In 2026, assume that any unsupervised task can be done with AI help. Pretending otherwise rewards the candidates who are least honest about it. The better approach is to decide what you actually want to assess and write a policy to match. If the job uses AI tools every day — most software, marketing and analytics roles now do — allow them and assess how well the candidate directs, checks and improves the output. If the job requires the candidate’s own unaided judgment or writing, say so, and shift that part of the assessment into a live session.

Be consistent with your own practices: if you use AI to score submissions or summarise interviews, the same transparency rules that apply to AI-assisted CV screening apply here, including the EU AI Act’s high-risk obligations for selection tools from December 2027.

Closing the loop with candidates

Candidates who invest hours in a task deserve a response that reflects that. Send a decision within a week, with two or three specific lines of feedback drawn from the rubric. It costs little and is one of the clearest signals of a well-run process. Then feed the results into the panel’s decision: the task score belongs on the scorecard alongside interview scores, and the debrief method in making the final hiring decision explains how to weigh them.