Every list of work sample test examples I've read this year is technically correct and quietly out of date. Write a memo. Build the spreadsheet. Triage this inbox. Those were good examples back when the hard part was producing the artifact, and the hard part stopped being that. If you're a recruiter carrying open roles across sales, customer success, product and ops, the question worth your time isn't which work sample test to pick off a list. It's which of them still measure anything once the candidate has a chat window open on the second monitor. This piece gives you one test for that, plus examples by role family that pass it.
The classic examples, and why they stopped working
The old standbys are all the same shape underneath. The candidate gets a complete brief, works alone, and hands back a document. That shape is exactly what a general-purpose model is good at now.
| Classic example | Role it was used for | What happens when the candidate has AI |
|---|---|---|
| Write a short memo or blog post | Marketing, comms | A competent draft appears in under a minute. You're reading the model's prose and the candidate's taste in editing it, if they edited it |
| Build a spreadsheet model | Finance, ops | The formulas are right. You learn nothing about whether the candidate knew which numbers mattered |
| In-basket exercise: triage twenty emails | Admin, ops, management | Plausible triage, produced instantly. The ranking looks reasonable because reasonable is what these models output |
| Take-home coding exercise | Engineering | Passes the tests. This is the one the market noticed first, and it's why take-homes lost their reputation |
| Analyze this dataset, present findings | Data, analytics | Clean charts, fluent narrative, and no evidence the candidate would have asked whether the data was trustworthy |
None of these formats is broken as an idea. What broke is the assumption underneath them, which is that producing the artifact was the expensive part and therefore the informative part. When I raise this with talent leads the first instinct is usually proctoring, and I get why. I think it takes you somewhere unpleasant, which I've argued separately in why detecting AI cheating is the wrong question.
My bias, before I tell you what to do instead. Skillvee is a 60-minute "day at work" simulation that replaces the recruiter phone screen and the first interview round. The candidate solves a real challenge, talks to AI teammates to get the information they need, and defends their decisions on a call, while the screen is recorded. It runs for any role whose work happens on a computer, and it measures how someone works: communication, collaboration, agency, AI leverage, problem solving, quality of output and time management. So the problem in this article is the one I work on every day, and I'm not a neutral party about the answer.
The missing-information test
One question sorts work sample test examples into useful and decorative:
Can the candidate finish this without asking anyone for anything?
If yes, you're measuring drafting speed and tool access. If no, you're measuring the thing you actually hired for.
Think about what the job is on a normal Tuesday. Almost nobody in a white-collar role gets a complete brief. The brief is thin, the person who knows the real constraint is in another timezone, and most of the work is figuring out what you're missing and going to get it. That behavior separates the people who are good at these jobs from the people who aren't, and it's invisible in every self-contained task in the table above, because a self-contained task has deliberately removed it.
So build the sample the other way around. Decide what the candidate must not be told, put it inside a person they can reach, and see whether they go and get it. The artifact at the end still matters. It just stops being the only evidence you've got.
Work sample test examples by role family
These are the four I'd build first for a recruiter hiring across functions. Each one is roughly an hour, and each is defined by what's withheld rather than by what gets produced.
| Role family | The task | What is withheld | What you learn |
|---|---|---|---|
| Sales | Prep for and run a discovery call on a thin prospect brief, then send the follow-up email | The buyer's real problem, which only comes out if the candidate asks about the current process and what it costs | Whether they ask before they pitch, and whether the email matches the call they just had |
| Customer success | An account is going quiet before renewal. Produce a plan for the next two weeks | That the champion left six weeks ago, which the support tickets hint at and the account manager will confirm if asked | Whether they read the history, whether they ask a human, and whether they change the plan when the answer arrives |
| Product | Two stakeholders want opposite things for the next cycle. Write the call and defend it | Each stakeholder's real constraint, one commercial and one technical, neither volunteered | Whether they surface the constraints before deciding, and whether the write-up is honest about what it gives up |
| Ops | A weekly report has been wrong for a month. Fix it and say what to do about the past month | That two teams changed a definition on different dates, discoverable from a messy changelog plus one question | Whether they chase the cause or patch the symptom, and how they handle telling people the numbers were wrong |
What isn't in that table matters as much as what is. No scoring of politeness, no personality read. The evidence is behavioral: what they went looking for, who they went to, and whether the final artifact reflects what came back.
For the sales row I've written up the scoring in more detail in sales hiring assessment tools, including why the hiring manager playing the buyer is usually the part that breaks.
Behavioral interview vs work sample
This is the choice recruiters ask me about more than any other, and the two aren't really competing.
A behavioral interview asks a candidate to narrate work you can't see, at a company you can't call, filtered through how good they are at telling stories about themselves. Some people who are excellent at the job tell those stories badly. Some mediocre people tell them beautifully, because they've told them forty times.
A work sample gives you work that happened in front of you, under conditions you chose, identical for every candidate. It's worse than the interview at surfacing motivation, long-horizon judgment, and how someone behaves when a project goes bad for six months. It's much better at everything that happens on a Tuesday.
Run both if you can. If your process only has room to instrument one of them properly, instrument the sample, because consistency is cheap there and nearly impossible in the interview.
Job simulation hiring without buying a platform
You don't need software for any of this. You need about an afternoon.
- Pick one real thing from last quarter. A deal, an escalation, a broken report. Something that actually happened, with the names stripped.
- Write the thin brief. Include enough to start, and deliberately leave out the fact that decides the answer. Write that fact down somewhere separate so you remember what you hid.
- Put it inside a person. A colleague on Slack for an hour, playing the account manager or the stakeholder, with instructions on what to say only if asked directly. This is the load-bearing piece.
- Write the scoring before anyone takes it. Four or five observable behaviors, in plain language, agreed with the hiring manager. What the candidate asked for counts as evidence, and say so in writing.
- Give every candidate the same hour, the same brief, and the same person working from the same script.
Step five is where the homemade version falls over, and it's worth being direct about that. The moment your stakeholder is tired, or has read the CV, or likes this candidate, the conditions aren't identical any more and the comparison quietly stops meaning anything. That's the problem we ended up building a product around. It's also the one you can mostly solve with a script and some discipline.
Where this gets hard
Two caveats I'd want if I were reading this.
Cost is the first. A sample with a person inside it burns an hour of a colleague's time per candidate, and that doesn't scale to two hundred applicants. It scales fine to the six people you were going to phone screen anyway, which is where I'd put it, replacing the screen rather than stacking on top of it.
The second is that hiding information annoys some candidates, and a few will tell you so. The fix is saying what you're doing up front: this brief is incomplete on purpose, the people who know more are reachable, asking them is the job. Candidates who are good at this work tend to relax when they hear it. I'd also argue it's the fairer version, though that's my read rather than a finding, because the ones who lose under the old format are often the ones who would have asked.
If you want to see this with people inside the sample who never get tired and never read the CV, that's the thing worth putting on one of your open roles.