How to prevent AI cheating on take-home case studies and modeling tests

Why take-homes broke in 2025

A take-home case study used to test two things at once: whether the candidate could do the analysis, and whether they would put in the hours. In 2025 both signals went away. A three-statement model, a DCF with a sensitivity table, a market-sizing memo or a LBO paper can be produced by a general-purpose AI model in under 20 minutes, formatted and footnoted. A candidate who could not build the model can submit one that looks better than the one your associate would build.

The same is true for written assessments in consulting, accounting, law, and for the "respond to this client email" exercises that search firms use. The deliverable no longer proves authorship.

Most advice on the web says: drop the take-home and move everything to live, adaptive interviews. For finance and professional-services recruiting that advice is wrong. The take-home is where you see how a candidate structures a problem over hours, not minutes, and it is the only stage that scales across a campus class of several hundred. The fix is to make authorship verifiable, not to delete the stage.

The three ways candidates cheat on a take-home

1. AI writes it. The candidate pastes the prompt into a chat model, iterates for an hour, and submits. Detectable in the submission itself (statistical fingerprints of machine-generated prose and code) and in what ran on the machine during the window.
2. Someone else does it. A friend, a tutor, a paid service. Often combined with remote-access tools (TeamViewer, AnyDesk, Chrome Remote Desktop) so the helper works directly on the candidate's laptop. Detectable at the operating-system level as a remote session, and in the follow-up interview.
3. The candidate does it with undisclosed AI help throughout. This is the grey zone. Many firms now allow AI for research and prohibit it for the analysis itself. The policy only works if you can see which one happened.

Control 1: Set the rules and say they will be checked

Write the AI policy into the case study instructions. Say what is allowed (reading sources, spell-check) and what is not (generating the model, the memo or the slides; having anyone else touch the work). Say that submissions are reviewed for AI-generated content and that the candidate may be asked to walk through any part of the work live.

Stating the rule and the review is itself a deterrent. It also puts the firm in a defensible position if a candidate is later removed from the process: they agreed to the terms.

Control 2: Verify the author in a short live follow-up

Fifteen minutes on video, within a few days of submission. Pick one cell in the model, one assumption in the memo, or one slide, and ask the candidate to change it and explain the effect. A candidate who built the work does this without thinking. One who did not will stall.

Run this follow-up with live detection active (see control 3), because the same AI overlay tools that feed answers during a first-round interview will be running during the follow-up.

Control 3: Use software that records what ran on the machine and reviews what was submitted

This is the part most firms are missing. Two layers:

During the take-home window. A lightweight desktop agent runs on the candidate's machine for the duration of the case study. It does not read files or capture content. It records whether restricted programs ran: AI assistants, overlay tools such as Cluely, Interview Coder and LockedIn AI, remote-access software, virtual machines. The output is a binary finding with the process name, hash and timestamp. If a restricted program ran, you know. If none did, the candidate has a clean record they can point to.

On the submission. Documents, spreadsheets, slide decks and code are analyzed for AI-generated content. These classifiers do not produce certainty, so the result is reported as a ranked signal with the evidence attached, not as an automatic rejection. Your team decides.

Both layers produce an integrity report that sits in the candidate's file alongside the submission and the follow-up notes. If the hiring committee asks why a candidate was advanced or removed, there is a record.

What this looks like in practice at a bank or fund

• Campus recruiting. Case study sent to 300 candidates with a 72-hour window. Agent active during the window; submissions reviewed on upload. Reviewers see the outliers first, ranked against the cohort, instead of reading 300 models cold.
• Lateral and experienced hires. Modeling test taken on a candidate's own laptop at home, or on a firm laptop in the office. The same agent runs in both settings. The follow-up interview runs with live detection.
• Search firms. The firm runs the take-home for the client and delivers the integrity report with the shortlist. The client gets evidence with each candidate rather than a recruiter's assurance.

What candidates see

Candidates are told what is monitored before they start, install a desktop agent of about 100 MB (macOS or Windows) for the window, and remove it afterward. No files are accessed, no screen is continuously recorded, no keystroke content is captured. Most candidates, in our experience, prefer a monitored take-home to losing the stage entirely to a timed live test, because it keeps the part of the process where they can show their best work.

Checklist

1. AI policy written into the case study instructions, with the statement that work is reviewed.
2. Live follow-up booked for every candidate who advances, with detection active.
3. Desktop agent active during the take-home window.
4. Submission run through AI-generated-content review.
5. Integrity report filed with the submission; hiring committee decides.