How to Use AI in a Take-Home or Interview Exercise in 2026 (When AI Is Allowed — and How You're Scored)

Short answer: In 2026, many take-home assignments and interview exercises now explicitly allow — or even require — you to use AI. When that's the case, employers aren't scoring whether you used it. They're scoring how: how well you framed the problem, steered the tool, verified its output, tailored it to their specific brief, and can explain your choices. The winning move is to use AI for the mechanical parts and spend your saved time on the judgment a tool can't provide. This guide covers all three situations — AI required, AI allowed, and AI banned — plus how to disclose your AI use so it builds trust instead of raising doubt.

The interview task changed, and most advice hasn't caught up

For years the take-home assignment ran on a simple assumption: you, alone, produce the work, and the quality of that work reveals your ability. AI broke that assumption.

By 2026, enough candidates use AI on assignments that pretending otherwise is naive — and a growing number of companies have stopped pretending. Instead of trying to police a tool that's now part of daily work, many have flipped the exercise on its head: use AI, we expect you to, and we'll evaluate how well you do it. Some assignments now come with an explicit line like "you may use any AI tools you like" or even "we expect you to use AI — show us your process."

That's a genuinely different test, and the old advice ("just be thorough and original") doesn't map to it. If you treat an AI-allowed take-home like a solo exam, you'll be slower than you need to be. If you treat it like a copy-paste shortcut, you'll submit confident-sounding work that falls apart the moment someone asks you a follow-up question. Neither is what strong hires do.

This is really the same skill employers are probing when they ask "how do you use AI in your work?" in the interview itself — except here you're not describing your judgment, you're demonstrating it in real time. Let's break down how to do that well.

Step zero: read the rules, and ask if they're unclear

Before you open a single tool, find out which of three regimes you're in. Getting this wrong is the one mistake that ends the process regardless of how good your work is.

• AI required or expected. The brief says something like "use AI tools and show your process," or the task is obviously impossible to do well in the time given without them. Here, not using AI — or using it and hiding it — works against you. They want to see you direct the tool.
• AI allowed. The brief permits AI but doesn't require it, or is silent in a context where AI is normal (most knowledge-work take-homes now fall here). Use it, use it well, and be ready to explain what you did.
• AI banned. The brief explicitly says no AI, or the exercise is clearly designed to test unaided fundamentals (a live whiteboard with the tool closed, a proctored assessment, a "we want to see how you think" instruction). Respect it completely. Using AI here is an integrity violation, and integrity flags don't get overturned by good output.

If the instructions don't make it obvious, ask — email the recruiter or note your question at the top of your submission: "Happy to use AI tools on this if that's welcome, or to do it unaided if you'd prefer to see that — let me know." Far from looking indecisive, this reads as someone who takes the rules and the relationship seriously. In 2026, asking whether AI is permitted is a maturity signal, not a weakness.

What evaluators actually score when AI is allowed

Once AI is on the table, the rubric quietly shifts away from raw output and toward the things a tool can't do for you. Understanding this is most of the battle. When AI is allowed, reviewers are looking for:

1. Problem framing. Did you understand what was actually being asked — including the unstated constraints — before you started generating? A good prompt starts with a well-understood problem. Candidates who paste the raw brief into a chatbot and ship the first response reveal that they never framed anything themselves.
2. Steering and iteration. Did you direct the tool toward a genuinely useful answer, or accept its first generic pass? Strong candidates iterate: they give context, push back on weak output, and ask for the specific thing the brief needs.
3. Verification. Did you check the output? This is the single highest-signal behavior. AI is confidently wrong often enough that submitting unverified work is the clearest possible tell of weak judgment. Facts, numbers, code behavior, citations — all of it needs checking against a real source.
4. Tailoring and judgment. Did you shape a generic draft into something specific to their context, or does your submission read like it could have been produced for any company? The tailoring is where your understanding shows.
5. Ownership. Can you explain and defend every choice in the final work? If a reviewer asks "why did you do it this way?" and the honest answer is "the AI suggested it and I didn't really look," you've lost — not because you used AI, but because you didn't supervise it.

Notice what's not on this list: typing speed, unaided recall, and whether you technically could have produced the first draft yourself. Those stopped being the point. The exercise has become a test of judgment and direction — which, not coincidentally, is exactly what "AI fluency" means for real work in 2026.

A repeatable method for an AI-allowed take-home

Here's a process that consistently produces defensible, specific work instead of impressive-looking noise. It maps to the rubric above.

1. Frame the problem before you generate

Spend the first chunk of your time without the AI. Read the brief twice. Write down, in your own words, what success looks like, what the constraints are, and what could go wrong. Note any assumptions you're making. This is the part that makes everything downstream good — a sharp frame produces sharp prompts, and it's also the part reviewers can most easily tell you skipped.

2. Use AI for the mechanical first pass

Now bring in the tool for what it's genuinely good at: a structured first draft, boilerplate, test cases, alternative approaches, a quick synthesis of a lot of material. Give it your framing, not just the raw brief. The more context you provide — the audience, the constraints, the specific angle — the less generic the output. Iterate: if the first pass is vague, say so and push for specifics.

3. Verify everything, ruthlessly

This is where you earn the interview. Treat every factual claim, number, and functional detail as unverified until you've checked it. Run the code and test the edge cases the brief implies. Trace each statistic back to a real source. Read the argument as a skeptic would. The time AI saved you on drafting is the time you now spend here — and it's the difference between a submission that survives follow-up questions and one that collapses.

4. Make it yours

Rewrite the parts that read as generic. Cut what doesn't serve their specific problem. Add the domain judgment, the tradeoff you'd actually make, the caveat a real expert would include. By the end, the work should sound like a competent professional who understood the assignment — not like a tool's default output. If you can't confidently defend a sentence or a decision, either understand it well enough to defend it or remove it.

5. Document your process

Add a short note — a few lines is plenty — describing how you approached it and where you used AI. For example: "I used an AI assistant to generate a first-draft structure and a set of test cases, then verified the calculations against the provided dataset, corrected two errors it introduced, and rewrote the recommendation to fit your stated priority on retention over acquisition." This does two things at once: it discloses your AI use honestly, and it demonstrates exactly the judgment they're scoring.

Live and screen-shared exercises: narrate, don't hide

Real-time exercises where AI is allowed are becoming common, and they feel higher-stakes because someone's watching. The winning behavior is counterintuitive: make the AI more visible, not less.

• Think out loud. Say what you're about to ask the tool and what you expect back. "I'll have it draft the outline so I can focus my time on the analysis" tells the interviewer you're making a deliberate choice, not reaching for a crutch.
• Verify out loud. When the AI produces something, check it in front of them. "That number looks high — let me confirm it against the data before I trust it." Catching an AI error live is one of the strongest signals you can send.
• Own the decisions. When you accept or reject a suggestion, say why. The interviewer is watching you supervise a tool, which is exactly the skill they're hiring for.
• Ask if you're unsure. If it's not clear whether AI is welcome in a live exercise, ask before you open it. This is the same rule as step zero — it never counts against you.

An interviewer running an AI-allowed live task isn't trying to catch you. They're trying to see whether you're the kind of person they'd trust to use these tools well on the job. Show them that person. Some of these formats overlap with AI-conducted interviews, where an AI system is doing the evaluating — worth understanding the difference so you know who, or what, you're actually performing for.

How to disclose AI use so it builds trust

The instinct to hide allowed AI use is understandable and completely counterproductive. When AI was permitted, transparency is pure upside:

• Be specific, not apologetic. "I used AI to draft and generate test cases, then verified and rewrote" is confident. "I know I used AI, sorry if that's a problem" is not. You did nothing that needs an apology.
• Frame it as judgment. Every disclosure should show a decision: what you delegated, what you kept, how you checked. That framing turns "I used a tool" into "I used a tool well."
• Never claim unaided work you didn't do. If asked directly whether you used AI and it was allowed, say yes plainly. Getting caught in a small dishonesty about a permitted tool is a far bigger problem than the tool ever was.

The only situation where you don't mention AI is one where you didn't use it — because it was banned, and you respected that.

The mistakes that get strong candidates rejected

Even capable people torpedo AI-allowed assessments in predictable ways. Avoid these:

• Prompt-and-paste. Submitting the first AI response with no framing, verification, or tailoring. It's the most common failure and the easiest to spot.
• Shipping unverified facts or broken edge cases. One confidently wrong number or one unhandled case the brief clearly implied signals that you don't check your work. In some roles, that's disqualifying on its own.
• Ignoring the specific brief. Generic output that would fit any company tells the reviewer you outsourced your thinking. The tailoring is the test.
• Being unable to explain your own submission. If a follow-up question ("walk me through why you did this") exposes that you don't understand your own work, the submission is worthless no matter how polished it looked.
• Using AI when it was banned. No amount of quality survives an integrity flag. When the rules say no, the answer is no.
• Over-engineering to look impressive. Adding complexity the brief didn't ask for, because the AI happily generated it, reads as poor judgment, not sophistication. Solve the actual problem.

When AI is banned: respect it, and it's fine

If an exercise explicitly prohibits AI, that's a legitimate design choice — often the company wants to see unaided fundamentals, or is testing something specifically about how you personally reason. Honor it fully. Close the tools. Do the work yourself.

This isn't a disadvantage if you've built real competence, because the fundamentals that make you good at directing AI are the same ones that make you good without it. That's the deeper reason genuine skill still matters in an AI world: you can't verify what you don't understand, and you can't fake understanding under follow-up questions. Preparing to be genuinely capable — not just fast with a tool — is what makes you strong in all three regimes at once.

A quick worked example

Imagine a marketing analyst take-home: "Here's three months of campaign data. Recommend where to shift next quarter's budget, and explain your reasoning. You may use AI tools."

• Weak approach: paste the prompt and a data summary into a chatbot, get a plausible-sounding recommendation, format it, submit. It reads generically, one of the two key numbers is subtly miscalculated, and when asked "why not the other channel?" the candidate has no real answer.
• Strong approach: the candidate first writes down what "good" means here (incremental return, not vanity metrics), frames the real question, then uses AI to speed through the arithmetic and to draft alternative framings. They verify every figure against the raw data — catching an error the AI introduced — pressure-test the logic, and rewrite the recommendation to reflect a judgment call the data alone doesn't settle. They add a three-line note on how they used AI and what they checked. In the follow-up call, they defend every number and explain the tradeoff they chose. Same tool, same time budget — completely different signal.

The gap between those two isn't AI access. It's judgment, verification, and ownership. That gap is the entire modern assessment.

Where this fits for a career-changer

If you're using these interviews as a path into an AI-adjacent role — not just surviving them in your current field — the take-home is a preview of the job itself. The roles that career-changers are winning in 2026 are overwhelmingly about directing and verifying AI, not building it from scratch. Getting good at AI-allowed assessments and getting good at the AI-adjacent work itself are the same project. The same is true of the portfolio you bring to the table: work samples that show AI-assisted judgment — not just "I used ChatGPT" — are exactly the proof these exercises are looking for in live form.

Prepare for the assessment the way you'd prepare for the job: build enough genuine competence to recognize good work, use AI to move faster on the mechanical parts, and make verification and ownership your default. Do that and the exercise stops being a trap and becomes the easiest place in the whole process to stand out.

The honest bottom line

In 2026, "can I use AI on this?" is often the wrong question — the real one is "how do I use it well enough to stand out?" When AI is allowed, employers aren't testing whether you reached for the tool; they're testing whether you framed the problem, steered the tool, checked its work, tailored the result, and can defend every choice. Do the mechanical parts fast, spend your saved time on verification and judgment, disclose your process plainly, and respect the rules when AI is off-limits.

The candidates who treat an AI-allowed assessment as a shortcut submit fast, forgettable, sometimes-wrong work. The ones who treat it as a chance to show their judgment submit work that holds up under any follow-up question. That's the whole game now — and it's a game you can prepare for.