Technology · July 18, 2026 · 8 min read

AI Sandbox Assessment: Test How Candidates Work With AI

An AI Sandbox assessment measures the real 2026 skill: working with AI well. See how our live approach evaluates prompt craft, verification and judgement.

← Part of AI-generated assessments: the complete 2026 guide

If you are hiring in 2026, the most important thing about a candidate is no longer whether they can work without AI — it is how well they work with it. Yet most assessments still lock people in a proctored box and measure a skill the job no longer rewards. This is the case for the AI Sandbox assessment: a live, capability-first way to watch candidates use AI tools on a realistic slice of the role, so you evaluate the collaboration that will actually happen at their desk. For hiring managers and talent leaders, the stakes are direct — get this wrong and you filter for nostalgia while your best applicants prove nothing you can use.

Why banning AI tests the wrong thing

The single biggest shift in knowledge work since 2023 is that AI tools are now part of the daily toolkit. Engineers pair with code assistants, analysts lean on generation for first drafts, support teams triage with models in the loop. When an assessment blocks those tools, it measures whether a candidate can work without the very things they will use every day. That is not signal — it is a museum piece. You end up rewarding the person who memorised syntax over the person who ships good work with the tools in front of them.

There is a second problem. AI-banned tests have become an arms race. As models improve, the effort spent detecting and blocking them grows, and the candidate experience degrades into surveillance. The more honest response is to stop pretending the tools do not exist. If the job involves working with AI, the assessment should too — and it should reward doing it well. That reframing is the whole idea behind the AI Sandbox, and it sits inside a broader move towards AI-native hiring.

It helps to be concrete about what the old approach quietly optimises for. A candidate who spent last weekend rehearsing algorithm puzzles will out-score a stronger practitioner who solves real problems with a model at their side, because the test rewards the rehearsal, not the work. Worse, the gap between test conditions and job conditions is exactly the gap where bad hires hide: someone can pass a sterile, AI-free exam and then flounder the moment they are handed the actual toolkit — or the reverse, a genuinely effective operator who freezes under artificial constraints they will never face again. Neither outcome is what you are paying to find out.

What an AI Sandbox assessment measures

An AI Sandbox is a live environment where the candidate is given a realistic problem and the AI tools to tackle it, and the platform observes how they work. It is a work sample — for the general method, see work sample tests — but one built around the assumption that AI is present. Three things matter, and none of them is 'did it run'.

• Prompt craft — can they get a useful result from a model efficiently, and steer it when the first answer falls short?
• Verification instinct — do they notice where the AI is wrong, incomplete or confidently misleading?
• Judgement — do they know when to trust output and when to override it, and can they explain the call?

Those three map cleanly onto how people actually use AI at work. A candidate who fires off a vague prompt, accepts the first response and moves on is a different hire from one who interrogates the output, spots the flaw and corrects it — even if both produce something that superficially works. The Sandbox is designed to tell them apart.

Verification instinct is the one that most surprises hiring teams. In an AI-heavy workflow, the failure mode is rarely that the model produces nothing useful — it is that it produces something plausible and wrong, and the person accepts it. The candidates you actually want are the ones who treat model output as a draft to be checked rather than an answer to be shipped. An AI Sandbox surfaces that instinct because it puts a subtle flaw in front of the candidate and watches whether they catch it, rather than asking them in the abstract whether they would. That is the difference between a claimed competency and a demonstrated one.

How our AI Sandbox works in practice

The set-up is deliberately close to real work. We hand the candidate a broken or unfinished artefact — a flawed prompt, a piece of AI-generated output with a subtle bug, a draft that needs judgement to finish — and ask them to make it right. They have the editor, the AI tools and live runs against visible test cases. Using AI is the task, not the transgression.

Because the exercise is generated per job, the artefact reflects the role you are actually hiring for rather than a generic puzzle. A backend candidate debugs backend code; a data analyst validates an analysis. The session is standardised so every candidate for a role faces a comparable exercise, which keeps the decision defensible and the comparisons fair. Under the hood, the AI Sandbox is one part of the five-pillar framework — you can see how the pieces fit in the five pillars of hiring.

The starting artefact matters more than it might seem. A blank editor tests whether someone can begin from nothing, which is not the shape of most modern work. Handing over an imperfect starting point — a prompt that nearly works, output that is ninety per cent right — mirrors the reality of a job where you inherit, refine and correct far more often than you build from scratch. It also raises the ceiling of the exercise: a strong candidate can go beyond the fix and improve the approach, while a weaker one may not even see what is wrong. The same task, in other words, reads signal across a wide band of ability without needing a different test for every seniority level.

AI Sandbox versus AI Fluency

These two get conflated, so it is worth being precise. AI Fluency is a pillar of the assessment framework — a competency: does a candidate understand responsible, effective AI use? The AI Sandbox is the environment where that competency is put to work and observed live. Fluency is the thing you want to know; the Sandbox is how you find out. If you are mapping the fluency side specifically, AI fluency as a hiring signal goes deeper on the competency itself.

Integrity in an open-tool exercise

The obvious objection to letting candidates use AI is that it opens the door to cheating. In practice, the reverse is true. When using AI is the point, there is nothing to smuggle in — the tools are meant to be there. What you still want to catch is someone pasting in a colleague's finished work or having a third party sit the exercise, and that is where the integrity engine comes in, watching for activity inconsistent with genuine work rather than policing tool access. It is a far cleaner posture than an AI-banned test that quietly loses the detection race. For the wider picture on keeping AI-era assessments honest, see preventing cheating in AI-generated tests.

This shift also improves the candidate experience, which is not a soft concern — it shows up in your completion rates and in how offers are received. Being watched through a locked-down proctored box is an adversarial way to start a relationship with someone you may soon employ. A live exercise where candidates work the way they normally work, with the tools they normally use, treats them as professionals rather than suspects. Strong applicants notice the difference, and in a competitive market the perceived quality of your process is part of what convinces them to say yes.

How to introduce an AI Sandbox into your process

You do not need to rebuild your pipeline. The most reliable way to adopt an AI Sandbox is to slot it in where a take-home or a whiteboard round sits today, and to keep it short and role-tuned. A focused session gathers stronger signal than a long unpaid assignment, and it respects candidates' time — which protects your completion rates and your funnel. Start with one or two high-volume roles, calibrate the exercise to the seniority you are hiring for, and compare outcomes against your existing method.

How you frame the exercise to candidates matters as much as the exercise itself. Tell them, in the invitation, that AI tools are available and that using them well is part of what you are assessing. Candidates who arrive expecting a trick relax when the rules are explicit, and the ones who would have hidden their AI use in a conventional test instead show you their genuine working style. Set a realistic time expectation, name the tools available inside the environment, and explain that the working process — not just the final artefact — informs the evaluation. That transparency is not a courtesy; it is what makes the signal honest, because everyone is playing the same game by the same rules.

• Pick a role where AI tools are already part of the day job, so the Sandbox reflects reality.
• Keep the session focused — enough to gather signal on prompt craft, verification and judgement, no more.
• Standardise the exercise per role so every candidate is comparable and the decision is defensible.
• Debrief the session in a structured interview rather than treating the score as the whole answer.

Common mistakes to avoid

Two errors undermine most first attempts. The first is treating the Sandbox as a pass-or-fail gate on whether the artefact runs, which throws away the very signal that makes it worth doing — a candidate who verifies carefully and lands just short of a clean run is often a better hire than one who gets a green tick by luck. The second is bolting it on without changing anything downstream: if your interviewers never look at what the candidate actually did in the session, you have gathered rich evidence and then ignored it. The Sandbox earns its place only when the working session feeds the human conversation that follows.

Done this way, the AI Sandbox stops being a novelty and becomes the part of your process that actually predicts how a person will perform once the offer is signed. If you want to see it end to end, you can book a demo or explore the AI Sandbox feature directly.

Stop testing whether candidates can work without AI. Give them the tools, give them real work, and watch how they use both.