The formats at a glance
Four formats cover most technical interviews. Here is what each one shows, what it misses and what it costs the candidate and your team.
| Algorithm test | Take-home task | Live pair coding | Work sample | |
|---|---|---|---|---|
| What it shows | Speed with data structures and algorithms under a time limit | How someone builds a small project when they have time to think | How someone reasons out loud, takes a hint and explains code | A short version of the real job: finding the problems, building, checking and explaining |
| With AI available | An assistant can solve many puzzle questions, so the result means something only with AI blocked and the session watched | You cannot tell who or what did the work | You can allow AI and watch how they use it, for as long as the session lasts | Still works with AI on, if the task is built so the assistant cannot finish it alone |
| What it misses | Most of the daily work: messy data, unclear requirements and checking results | A fair comparison, because candidates spend different amounts of time and get different amounts of help | Anything that takes longer than the session, such as a pipeline on real data | How someone talks a problem through with a colleague, unless you follow up with a conversation |
| The candidate’s time | Usually an hour or two, online | Several hours at home, unpaid and hard to cap | About an hour, at a set time | One to two hours in one sitting, with a hard time limit |
| Your team’s time | Little, because the platform scores it | An engineer to review each submission by hand | One or two engineers for every session, plus notes | A realistic task and its checks take real effort to build; reading a scored result is quick |
| Fairness | The same questions and scoring for everyone, but practice with puzzles counts for a lot | Favors people with free time, and hard to compare | Depends on the interviewer and on nerves, unless the questions and notes are structured | The same task, time and checks for everyone, and close to the job |
| Fits best when | You screen many applicants for work where algorithms are the daily job | The role needs long, unhurried work and you can cap the time, or pay for it | You want to see how someone collaborates, in a later round | You judge people on real output, such as data pipelines or AI apps, and they will use AI at work |
Algorithm tests
An algorithm test gives every candidate the same timed puzzles, often on data structures, sorting or graphs, and scores the answers automatically. It is quick to run for many applicants and the same for everyone, which is why so many companies use it as the first screen.
Its weakness now is AI. Puzzles are the kind of task AI does best: at the 2025 ICPC World Finals, the leading university programming contest, Google DeepMind reported that its Gemini model solved 10 of the 12 problems within the five-hour limit. So an algorithm test means something only with AI blocked and the session watched, and even then it shows little of jobs built on messy data and judgment.
It fits best when you screen many applicants for roles where algorithms are the daily work.
Take-home tasks
A take-home task asks the candidate to build something small on their own time, usually over a few days. It is closer to the real job than a puzzle, and good candidates often like showing real work.
It has three problems. It costs the candidate hours of unpaid time, which favors people with free evenings. Results are hard to compare, because everyone spends a different amount of time. And you cannot tell who or what did the work, now that an AI assistant can write most of a small project.
If you use one, keep it short, cap the time, give everyone the same brief and review notes, and pay for the time if you can.
Live pair coding
In live pair coding, an engineer from your team and the candidate work on a problem together in a shared editor, usually for about an hour. You hear how the candidate thinks, how they take a hint and how they explain their code.
It is hard to fake, and you can allow AI and watch how they use it. But an hour is too short for work that needs real data or a real codebase, and the result depends on the interviewer. It is fairest when every candidate gets the same problem and every interviewer takes notes against the same written criteria.
Work samples
A work sample asks the candidate to do a short, realistic version of the job, then judges the result. For a data engineer, that means messy source data and tables to publish. For an engineer who builds AI apps, it means a client’s problem and a working app.
To hold up when candidates use AI, a work sample needs three things: problems planted in the work, so there is something to find; automatic checks or a source of truth, so the result is judged on facts; and a re-run on fresh data, so answers typed in by hand fail. The cost is building it, because a realistic task and its checks take real effort to write.
What the research says
In the largest recent review of hiring studies (Sackett and colleagues, Journal of Applied Psychology, 2022), structured interviews predicted job performance best (0.42). Work-sample tests (0.33) did better than general ability tests (0.31) and far better than unstructured interviews (0.19).
Two lessons follow. Structure matters as much as format: the same questions and the same written criteria for every candidate. And the strongest loop pairs a work sample with a structured conversation about the work.
When candidates can use AI
Companies now take one of three positions on AI in technical interviews. Some block AI and watch the session, which keeps older tests meaningful but tests a way of working the job no longer uses. Some allow AI and judge how the candidate uses it. Some build tasks where AI help cannot decide the result, because the hard part is finding what is wrong and proving the answer.
Whichever you choose, tell candidates the rule before they start, and make the task match it. A puzzle test with AI allowed mostly measures the AI. A work sample with AI allowed measures whether the candidate checks it.
Tools that run each format
Many platforms now run more than one of these formats, and their products change often. As their own product pages described them on 5 November 2026:
- HackerRank describes online coding tests and certified assessments, real-world coding questions and projects, take-home assessments, and a way to set clear expectations for when candidates can and cannot use AI.
- Codility describes coding tests before the interview, with AI on or off for each task, and live coding interviews in a shared editor.
- CoderPad describes coding assessments built around real-world engineering tasks, and real-time, collaborative coding interviews where candidates can use an AI assistant and the team can review the prompts.
- CodeSignal describes skills assessments built as job-relevant simulations, and live technical interviews in a shared workspace.
If your main need is to screen many applicants on general coding skill, one of these is likely the better fit.
Where a BuildInterview Audition fits
A BuildInterview Audition is a timed work sample for AI and data roles. The candidate plans with AI off, builds with an AI assistant that never plans the work or writes the whole solution, and writes two short notes. Problems are planted in the work, code checks the result on fresh data, and an AI reviewer rates what code cannot settle. Your team gets one report with the evidence behind every finding and no hire or reject recommendation.
It fits when:
- You hire Forward-Deployed AI Engineers or Data Engineers, the roles live today.
- Candidates will use AI at work, and you want to see whether they check what it writes.
- You want every candidate to get the same task, the same time and the same checks.
Another format fits better when:
- You need to screen thousands of applicants on general coding skill. A coding test platform is built for that volume.
- The role is one we do not cover yet. A short, capped take-home or a structured live session built around your own work will show more.
- You mainly want to see how someone collaborates in real time. A structured live session shows that better.
Read how it works for each role in the Data Engineer interview guide and the FDE interview guide, or try the candidate’s workspace in the demo.
Common questions
What is the best technical interview format when candidates use AI?
A timed work sample of the real job, with AI allowed and the result checked by code, followed by a structured interview about the work. It shows whether the candidate can check AI-written work, which most technical jobs now need. Algorithm tests hold up only with AI blocked and the session watched.
Are algorithm tests still useful?
Yes, for roles where algorithms are the daily work, and for screening large numbers of applicants on general coding skill. They say less about jobs built on messy data, unclear requirements and judgment, and an AI assistant can solve many common puzzle questions.
Are take-home assignments fair to candidates?
They can be, if the time is short and capped, the brief is clear and every submission is reviewed against the same notes. Long, open-ended take-homes favor people with free time and are hard to compare, and you cannot tell how much of the work was the candidate’s own.
What is a work sample test?
A work sample test asks the candidate to do a small, realistic piece of the job and judges the result. In the largest recent review of hiring research (Sackett and colleagues, 2022), work samples predicted job performance better than general ability tests and far better than unstructured interviews.
Is live coding a good interview format?
It is good for seeing how someone reasons and communicates, especially when every candidate gets the same problem and every interviewer takes notes against the same criteria. It is too short for work that needs real data or a real codebase, and results vary with the interviewer.
How long should a technical interview take?
Keep the candidate’s total time to a few hours. A timed work sample of about two hours in one sitting, then a structured follow-up of under an hour, shows the job without asking for a day of unpaid work.
Should candidates be allowed to use AI in technical interviews?
If they will use AI at work, testing with it shows more of the real job, as long as the task is built so the AI cannot finish it alone and you judge how they checked the result. If the role must work without AI, block it and watch the session. Either way, tell candidates the rule before they start.
See a work sample on the real job
Every approved company gets up to 5 interviews free. Create a company account, pick an Audition for the level you are hiring, and send it to your first candidate.