For teams that run take-home interviews

Trust your take-home interview again.

AI broke screening. Aptlab brings the signal back: candidates work with AI in a real environment, and you see how they did it, not just what they turned in.

Run your take-home on Aptlab See an annotated submission

Bring your existing take-home or start from a template. No credit card required to get started.

Maya Bennett · Checkout experiment

Hold the rollout

The overall result does not show a clear conversion gain, and the drop among existing users is worth investigating before we make further changes.

How I measured conversion

Existing users account for the concern

Recommendation

Overview. Animation complete.

The problem

Did you interview the candidate — or their AI?

AI broke screening.

93% more applications than 2021. Most are AI-polished and hard to tell apart.
~40 interviews per hire, up a third since 2021. All for the few who get through.
29% of candidates already use AI to write their assessment answers. Take-homes and coding tests are easy to game.
~50% of take-homes are AI slop. The finished artifact stopped telling you who did the work.

Teams spend more time and trust the signal less.

Sources: Gem 2026 Recruiting Benchmarks; Gartner (2025).

How it works

Set up your interview in 5 minutes, free.

STEP 1

Bring your case

Upload the take-home you already use, or start from a template.

STEP 2

Send a link

The candidate works it with Claude Code / Codex. You pick the model, what it can touch, and how long the clock runs.

STEP 3

Read the annotated submission

What they produced, how it came together, and where the judgment calls were.

The path

You see the reasoning, not just the answer.

You see the candidate’s whole path: the questions they pose, the draft they take or rewrite, the mistake they catch. See how they use AI.

Take-home trace· Alex Morgan

Codex

Here’s the checkout breakdown.

Device	Users	Orders	Conversion
Mobile	28,800	605	2.10%
Desktop	19,200	727	3.79%

Who did what

See who directed, and who executed.

Every take-home comes back as an annotated submission showing who directed, who executed, and whether the candidate verified the work. Click any mark to jump to that moment in the session.

Decide on what happened, not on what was handed in.

Recommendation

Do not ship this version.No direction

AI led

AI chose to hold the rollout.

Maya asked for a recommendation; AI made the call.

“Don't ship, and hold further iteration until you understand why existing users are converting worse.”
AI response · excerpt

Open annotated submission

THE APTLAB ENVIRONMENT

Same conditions. Every candidate.

Consistent setup. The same AI, files, and time limit—with an enforced deadline.
Your case stays protected. Candidates work with your data inside the Aptlab environment.
AI on your tab. You cover usage, so candidates don’t have to ration tokens.

Two ways to run it

Your candidate’s environment, or Aptlab’s.

Feature	Candidate’s environmentFree	Aptlab environmentBest signal$7/session + AI usage
Best for	Candidates who live in their own tooling, free to you	The clearest hiring signal, with consistent conditions for every candidate
Both include	Annotated submission, your case data stays remote, time on case, and every deliverable in one place *ATS integrations forthcoming
Where the candidate works	Their own AI tool, such as Claude Code, Codex, ChatGPT, or Claude Desktop	Claude Code / Codex, right in the browser
Candidate setup	Connect their AI tool via MCP	Just open a link
Who pays for the AI	The candidate	You cover usage — billed at cost
Same AI for every candidate	Candidate’s preferred model	Same model. Same permissions. You choose both.
What you see	Partial prompt trace	Complete prompt trace

Case library

See what a case looks like.

Open-ended problems across analytics, product and data engineering. Start from one, or upload your own.

Data ScienceThe example on this page

An A/B test comes back flat. Ship it anyway?

Pull the real funnel data, decide whether a shrug-worthy test result can be trusted, and turn it into a clear ship / iterate / hold call — no answer key, no trick question.

Supplied materials: Three raw tables: assignments, orders, and users, plus a data dictionary.
Expected deliverable: A written recommendation to Priya with your findings, supporting evidence, and a ship / iterate / hold decision.

Data Engineering

Revenue jumped 13%. Can finance take that to the board?

Build a reusable revenue rollup from raw orders data and stress-test a headline number before it goes in front of the board — a data engineering case, not just another SQL puzzle.

Supplied materials: Five raw tables: users, items, orders, order_items, and order_status, plus a data dictionary.
Expected deliverable: A reusable revenue table or query and a written summary for Naomi explaining the trend, definitions, and data caveats.

Data Science

"Usage is up." Is it, really?

A product exec is tired of trusting a single smoothed line on a slide. Define "adoption" from raw event data yourself, and see whether it's sturdy enough to build a real dashboard on.

Supplied materials: Three raw tables: events, users, and plans, plus a data dictionary.
Expected deliverable: A written summary for Theo defining usage and adoption, explaining the trend, and flagging caveats for a recurring dashboard.

Browse the case library

Your take-home. A clearer hiring signal.

Bring the case you already use, or start from our library.

Run your take-home on Aptlab