Your AI-assisted GTM take-home needs a trail of your decisions

Whether you can use AI in a GTM take-home depends on that employer’s instructions for that assessment. When assistance is allowed, keep a short record of your decisions: the question you set out to answer, what the tool suggested, what you checked, and what you changed or rejected.

You can use our Markdown decision-log template alongside an assignment. The fully worked fictional example follows a growth candidate from an initial note through misleading suggestions to a defensible recommendation. It includes the source packet and calculations, so you can inspect every decision without opening another tool.

The example is a teaching artifact, not a real applicant’s submission or evidence that a particular disclosure improves hiring outcomes.

Start with the employer’s policy for this stage

Two policies show why a general statement that a company embraces AI is insufficient. We checked both pages on October 23, 2026:

Employer source	What it says	What that means for your assignment
Clay’s writing policy	Permits AI for brainstorming, drafting and proofreading, while making authors responsible for the ideas and wording. The article says the policy also applies to prospective hires preparing written interview materials.	Review every sentence and follow any additional instructions attached to your assignment. This writing policy does not describe permission for every interview activity.
Anthropic’s candidate guidance	Allows application refinement after a candidate-written first draft and interview preparation. Take-homes and live interviews exclude AI assistance unless Anthropic specifically permits it.	Permission to prepare does not extend automatically to the assessment itself.

If the instructions are unclear, ask a specific question: “Does the AI policy for this take-home allow brainstorming, editing, calculations or code generation, and is an assistance note required?” Until the scope is clear, do the assessed work yourself. A disclosure cannot turn prohibited assistance into permitted assistance.

Record the instruction’s source and date. Separately check what materials may be shared with a tool; permission to edit text does not establish permission to upload a private customer export.

Write your starting position before asking for help

Use a few lines, not a polished introduction. State the decision, the available evidence, the uncertainty and your initial view. This gives you something to compare with later suggestions.

In our fictional assignment, the candidate must recommend an onboarding improvement. The packet contains two nonrandomized groups: 20 of 100 accounts activated with standard onboarding; 12 of 40 activated after opting into a setup call. The candidate’s starting note is that the assisted group had a higher observed rate, but may have contained more motivated users.

That starting note matters when a fluent suggestion later credits the pilot with a 50% improvement, then recommends an email sequence. The candidate can identify both the unsupported causal claim and the switch from setup calls to emails.

Keep the decisions that changed the work

You do not need a transcript of every prompt. Keep the material choices in a small table:

Suggestion or question	Check	Decision
“The pilot increased activation by 50%.”	20/100 = 20%; 12/40 = 30%. The groups self-selected.	Report a 10 percentage-point observed gap. Do not attribute it to the intervention.
“83% of users struggle with importing.”	Fifteen of 18 support tickets concern imports; those tickets came from 12 accounts.	Use the ticket count as a lead for investigation. It is not the share of users affected.
“Read out a seven-day activation test after one week of enrollment.”	The final account enrolled has not had seven days to activate.	Wait for every enrolled account’s full observation window.

All numbers in this table come from the fictional download. The useful part is the check, including checks that leave you with an unanswered question.

Log accepted ideas too. In the worked example, the candidate accepts an in-product import checklist as a testable option, then narrows it to instructions for two known issues. They do not claim the checklist will fix every activation problem.

Attach evidence to the claims that survive

For each material claim, identify a source row, calculation or observation. Keep facts, assumptions and proposed actions distinct:

• Fact within the supplied packet: the assisted group activated at 30%.
• Limitation: assistance was optional, so the comparison does not isolate its effect.
• Hypothesis: making import instructions available during setup may help eligible accounts finish setup.
• Proposed test: compare the checklist with the existing flow using a defined assignment method and a complete observation window.

An unrun test belongs in the proposal. A fabricated customer quote, benchmark or conversion lift belongs nowhere in the submission. If the brief leaves a definition unclear, state the assumption and explain what would change if it were wrong.

For a technical work sample, the evidence might be a query, a test output or a replay. Our inbound audit project provides a separate dataset for practicing those checks; this guide focuses on documenting the decisions around your work.

Keep any assistance note specific and truthful

Follow the employer’s requested format. If a note is appropriate, state the tool you actually used, what it helped with, and what you independently checked. For example:

I used [actual tool and version, if available] to compare possible explanations and edit the final memo. I recalculated the rates from the supplied packet, rejected the causal claim, and chose the proposed test. The attached decision log identifies the material changes.

Use that wording only if it describes what happened. If AI wrote an initial draft or generated code, say so instead of describing the help as proofreading. Do not invent a human-first draft after the fact, or claim a check you did not run.

You do not need to submit a huge prompt archive unless requested. A concise record should help a reviewer trace the consequential choices without exposing unrelated private material.

Prepare to explain the recommendation without the tool

Pick one rejected suggestion, one accepted idea and one unresolved question. Explain each from the evidence. Then name a result that would make you change the recommendation.

In the worked example, a useful follow-up is whether accounts that contacted support eventually activated. The aggregate packet cannot answer that. Asking for the missing link is more informative than presenting a precise forecast from incomplete data.

Use the template to make your own work inspectable, and keep the employer’s requirements in charge. Compare assignments with the responsibilities in growth, GTM engineering, RevOps and content roles. For the hiring team’s side of a work sample, see our GTM engineer interview exercise.