Set candidate AI rules by first deciding what the work sample should demonstrate. If the role requires producing and checking work with an approved assistant, an exercise can examine that collaboration. If the employer needs evidence of an ability without generated answers, it needs an explicitly defined assessment condition for that ability. A vague instruction to “use AI responsibly” leaves both candidates and assessors guessing.
For HR and hiring managers, the practical deliverable is a short assessment brief: purpose, permitted assistance, supplied tools, expected evidence and scoring criteria. Agree it before inviting candidates. Keep the decision with the people responsible for selection and the organization’s assessment standards.
This guide proposes such a brief and a fictional purchasing exercise. It is not a validated selection instrument, an assessment of real candidates or evidence that we have tested an AI hiring product.
Define the ability before choosing the tool rule
A work sample is useful only if the employer can explain what the observed work means. Consider a candidate who produces an accurate supplier comparison with AI. That may provide evidence of preparing an assisted comparison under the exercise conditions. It does not by itself show that the candidate could perform every calculation unaided, or that they checked the assistant’s reasoning.
Dunlop and Lievens’ March 2026 commentary argues that candidate AI use can change what an assessment score captures: AI literacy may enter a score intended to measure something else. It also identifies unequal tool access as a concern and discusses redesigning assessments for human–AI collaboration. These are research-informed arguments and design proposals, not experimental validation of a particular employer’s work sample.
Write down the target capability in ordinary language. “Prepare a reliable comparison from supplied quotes, identify missing information and explain the recommendation” is more assessable than “be good with AI.” Decide whether tool selection, prompting or independent calculation is actually part of the job requirement. Avoid adding them to the score merely because the technology makes them interesting to observe.
Our guide to justifying job requirements covers that earlier decision. The question here is how assistance changes the evidence an assessment can supply.
Treat preparation and assessment as separate stages
A candidate might use AI to understand the role, practise a generic scenario, check grammar or generate the assessed answer. Those are different activities. Permission for one does not communicate permission for all of them.
The University of Warwick’s employer toolkit, revised in May 2026, recommends explicit guidance on acceptable candidate AI use, with examples. Its own examples distinguish preparation and writing checks from generating assessment responses. That is practical guidance in a particular recruitment context; it is not a universal employer policy or proof that one boundary is appropriate for every role.
State the rules for each stage in the invitation. If practice with AI is allowed but live answer generation is not, say so directly. If an assessed phase includes an assistant, identify the supplied environment and the permitted uses. Explain what happens during a follow-up discussion, including whether candidates can consult their submitted work, notes or tools.
Provide a contact route for questions and adjustments before the exercise. Do not let an assessor improvise a new restriction after seeing a strong submission. Changes in conditions can change the interpretation of the evidence as well as the candidate experience.
Use warnings as communication, not proof of compliance
Robie and colleagues’ study, published in March 2026, examined warning statements and self-reported AI use during a standardized pre-hire assessment in two applicant samples collected in 2024 and 2025. Warning conditions were associated with lower reported use. However, the study relied on self-report and one assessment platform. It did not establish that a warning eliminates actual assistance or preserves the validity of a different work sample. Two coauthors were affiliated with the assessment provider, Talogy.
For an employer, the useful distinction is between stating a rule and knowing that the assessment measures the intended ability. A declaration can clarify expectations and invite discussion, but it is not independent evidence of how every answer was produced.
Avoid building the process around confident guesses from prose style. When an answer leaves uncertainty about a required capability, use the established assessment process to obtain relevant evidence. Apply the same planned follow-up opportunities and criteria consistently, rather than inventing a private test for someone whose writing seems unusually polished.
Work through a fictional purchasing exercise
Suppose a fictional employer is hiring an operations analyst. In the role, staff may use an approved assistant to organize supplier information, but remain responsible for checking numbers and identifying terms that need clarification. The employer wants to assess that workflow.
The exercise supplies three fictional quotes with different quantity assumptions and delivery terms. No real supplier documents, customer data or productive business assignment are involved. Candidates prepare a comparison, identify unresolved questions and discuss the basis for a recommendation. The following brief is an editorial example to adapt and review, not a ready-to-use validated test.
| Brief component | Proposed condition | Evidence the assessor seeks |
|---|---|---|
| Purpose | Compare the supplied quotes and explain a qualified recommendation | Relevant reasoning using the stated purchasing need |
| AI access | Use the employer-supplied assistant during the preparation phase | Work produced under a documented tool condition |
| Data boundary | Use only the fictional exercise pack in that environment | Compliance with the exercise’s information limits |
| Submission | Provide the comparison, assumptions and unresolved questions | Traceable numbers and clearly bounded conclusions |
| Discussion | Explain selected figures and respond to a changed quantity assumption under stated tool rules | Understanding of the submitted work and its limits |
The scoring criteria should distinguish numerical accuracy, treatment of missing information and reasoning from presentation quality. A visually polished comparison should not earn credit for resolving a delivery condition that remains unknown. Conversely, an accurate, clearly qualified recommendation should not lose credit solely for ordinary phrasing when literary style is irrelevant to the role.
Ask for a concise account of material AI assistance: what it helped produce and what the candidate checked or changed. Do not equate a longer prompt history with greater competence. The important evidence is how the candidate handled the work, not the volume of text exchanged with a model.
Make access and conditions visible
If AI-assisted work is the target, avoid quietly comparing one candidate’s personal paid assistant with another’s restricted free tool. Document the supplied tool, available functions, exercise materials and time conditions. Provide a short orientation that explains the mechanics without revealing the substantive answer.
Comparable access does not make candidates’ interactions identical, and it does not establish fairness by itself. Tool failures, accessibility needs and unfamiliar interfaces still require a planned response. Decide who can pause or reschedule an exercise and how interrupted sessions will be recorded.
CIPD’s September 2026 selection-methods factsheet recommends telling candidates what to expect, considering adjustments and designing skill-based tasks to resemble relevant work. Those principles support a clear brief. They do not establish that supplying the same AI account makes a new assessment reliable or predictive of job performance.
Have the responsible HR and assessment specialists review the proposed conditions against the role and applicable organizational requirements. This is particularly important when introducing an AI-assisted score into an established process: the meaning of the score may have changed even if the numerical scale has not.
Rehearse interpretation before using the exercise
Before assessing candidates, ask the assessors to work through fictional submissions under the published rules. Include an accurate but poorly explained comparison, a fluent answer with an unsupported assumption, and a submission that identifies a material ambiguity. Check whether the criteria distinguish the intended capabilities.
Our interview-guide rehearsal article explains how to identify unclear rating instructions. A rehearsal can expose disagreements and broken exercise materials; it does not validate the instrument. Further evaluation may be needed before the organization relies on a new method for consequential decisions.
Retain the approved brief and the tool conditions with the exercise version. Review them when the job, assessment design or assistant changes. Start by making one existing work sample explicit about the ability sought and the assistance allowed. Candidates should know how to demonstrate the relevant work, and assessors should know what conclusions that evidence can reasonably support.