# The Take-Home Case Study Log, One Dataset Carried From Brief to Panel Q&A

*Take your own dataset take-home from brief to panel with a dated clarifying email, a capped hour log, an 8 to 12 slide deck, and ten rehearsed Q&A answers.*

- Canonical URL: https://www.refolk.ai/candidates/guides/take-home-case-study-panel-log
- Pillar: Interviewing
- Format: Teardown
- Published: 2026-10-20
- Last reviewed: 2026-10-20
- Reading time: 22 min
- Keywords: how to do a take-home case study for a data analyst interview, take-home assignment presentation to panel tips, how many hours should a take-home assignment take, what to put in a take-home case study deck, questions asked after a take-home presentation, how to scope a take-home assignment

## Key takeaways

- Company hour estimates for take-homes run low: EasyApply tells employers that if a task takes them 3 hours, it will take candidates 4 to 6.
- In careerfair.io's analysis of engineering take-homes, 76% were scoped between 50 minutes and 4 hours with a median of 3 hours, and no equivalent study exists for non-coding case studies.
- Final Round AI calls 4.5 hours on a 3-hour assignment fine, but says 8 hours signals someone who cannot scope work, so the hour cap is part of what gets judged.
- Reported panels run 45 to 60 minutes with 10 to 25 minutes of questions, so Q&A preparation deserves its own 1 to 2 hour block with a written answer for each of ten likely questions.
- In Refolk's index there are 34.3 US data analysts for every US analytics manager, which suggests, but does not prove, that a case study panel is mostly peers plus one manager.
- Slide one of a take-home deck should name one action, because a DoorDash candidate who listed metrics without saying what to do was rejected for not being decisive.

This guide is for anyone working on a take-home case study: a dataset, a business question, a deck to hand in, and a panel presentation at the end. I follow one composite case from the invite to the last panel question. That includes the hour log, the analyses that got cut, the gold-plating mistake caught at hour four, and the questions that are hardest to answer well. Put your own brief next to it and copy one decision at a time.

## The worked case: brief, window, and panel

A dataset-and-deck take-home usually gives you several days, one business question, and a deck that you present to a small panel. The session runs 45 to 60 minutes, and 10 to 25 of those minutes are questions. Everything below is sized to that format.

The case is a composite. I built it from the ranges that candidates and prep sites report, not from one real invite, so treat its clock times as my plan rather than a measured study. Here is the brief:

- **The ask:** a two-page PDF with an orders dataset and one question: conversion fell last quarter, so why, and what should the team do?
- **The window:** five days from invite to panel.
- **The stated effort:** "should take about 3 hours."
- **The deliverable:** a slide deck, emailed before the panel.
- **The panel:** 60 minutes with four people. Roughly 25 minutes of presenting, and the rest is questions.

Every element sits inside reported ranges. Aced (formerly Exponent) says most companies give 4 to 7 days to complete the assignment, and the final stage is a 2 to 4 person panel lasting 45 to 60 minutes. Airbnb reportedly sends a 2 to 3 page PDF about seven days before the final round. Here is how reported formats compare by employer:

| Employer | Prep window | Panel session | Panel size |
|---|---|---|---|
| Intuit | 3 to 5 days | 45 to 60 min | 2+ |
| Adobe | few days to a week | 45 to 60 min (20 to 30 presenting) | not stated |
| Airbnb | ~7 days | under 1 h, then 30 to 45 min breakouts | ~5 |
| TikTok | 3 to 7 days | not stated | panel |
| Guidehouse (experienced hires) | 24 to 48 h | not stated | not stated |

These rows come from third-party prep pages, not from employer invites. Treat them as the reported norm, not a promise. Formats also vary within a single company: OpenAI data science roles reportedly get a separate 48-hour take-home built around an A/B-test analysis and submitted as a slide deck.

Nobody has published how often an invite includes an hour estimate rather than just a calendar window. If yours only gives a window, you have to estimate the hours yourself, which the next section covers.

#### The take-home case study, invite to panel

1. **Read and estimate** - Write an hour number per piece and set a cap
2. **Clarify** - One dated email, with a fallback assumption per question
3. **Scope** - Five-line plan plus an out-of-scope list
4. **Build to the cap** - Analysis, then an 8 to 12 slide deck
5. **Submit** - Deck plus a one-line time-box statement
6. **Prep and present** - Ten written answers, then the panel

*Every stage ends in something written down, so nothing depends on a reply you may not get.*

## How many hours should a take-home case study take?

Budget from your own estimate, not the company's. For a brief that says 3 hours, plan on 4 to 6 real hours, and write down a hard cap before you open the data. The cap is part of what gets judged.

The company's number runs low for a structural reason: the person who wrote it already knows the answer. EasyApply says this openly in its advice to employers: "If it takes you 3 hours, it'll take candidates 4-6." The founder of Tapioca says strong candidates take around 2x the stated time, as a rule of thumb. careerfair.io's analysis of engineering take-homes found that 76% were scoped between 50 minutes and 4 hours, with a median of 3 hours. One take-home in that analysis was listed at 3 to 5 hours and took the engineer who was hired 7 to 8 hours.

All of that comes from coding take-homes and one founder's opinion. No public study measures non-coding case studies. Treat the multiplier as a planning assumption that you re-check at the end of your first hour, not as a law.

Practitioners set their caps in a fairly narrow band:

| Source | Default | Ceiling |
|---|---|---|
| hirecomb | 2 h | 4 h |
| rankid.dev | n/a | 3 h; negotiate at 4 to 8 h |
| careervision.coach | 3 h | 1 workday |
| resumemate | 3 to 8 h | 8 h if no time is given |
| askcruit | 4 to 8 h | 20+ h is a red flag |

The upper limit matters as much as the lower one. Final Round AI says "A 3-hour assignment that took you 4.5 hours is fine." It also says that eight hours on a three-hour task signals someone who cannot scope work or make tradeoffs under constraints.

**The case's decision.** The brief says 3 hours. I listed each piece and estimated it:

- 40 minutes to read and estimate
- 20 minutes for the email
- 30 minutes to scope
- 2 to 3 hours of analysis
- 1.5 hours for the deck
- 10 minutes to submit

That totals roughly 5 to 6 hours, which matches the EasyApply range. I set two caps. The build cap is 4.5 hours for analysis plus deck, which is Final Round AI's "fine" line for a 3-hour brief. The overall cap is 6 hours up to submission, which is 2x the stated figure. Q&A prep sits outside both caps, because it is preparation for a live interview, not part of the submitted work.

> **Rule:** Write the cap before you open the data
>
> The cap is a number with a clock time next to it, written in your log before hour one. When you reach it, you stop building and move to whatever is left: the deck, or the submission.

## The clarifying email: one message, dated, with a fallback per question

Send all your questions in one email before you start building, and log an assumption next to each one so you never stall while waiting for a reply. Practitioners on Fishbowl note that asking for clarification, or spotting important missing information, may itself be part of the test.

Cover four things:

1. **Metric definitions:** what "conversion" means in this dataset, including the numerator and denominator.
2. **Deliverable format:** whether they want PowerPoint, whether there is a slide limit, and whether to send the deck ahead.
3. **AI policy:** whether you may use AI tools and, if so, how to disclose it.
4. **Priority:** what to prioritise if time runs short.

The AI question has to be asked because employer policies conflict:

- Code.org allows AI on take-homes as long as you disclose your use to the interview team.
- GitLab requires take-home AI use to be cited and answered in your own words.
- Anthropic's guidance has said take-homes should be done without Claude unless stated otherwise.

These policies change. The Uncharted Career AI Interview Policy Tracker is an open dataset you can check, but the company's written reply is the only answer that binds you.

**Wait or proceed?** The sources disagree. Fishbowl says to start on educated assumptions while you wait. Skolar says to confirm the AI policy before doing any work at all. I think each is right about a different thing. Read, scope, and plan while you wait for a reply, but do not touch an AI tool until the policy is confirmed in writing or the brief already states it. Recruiter reply speed is not established publicly, so a plan that depends on a reply can lose two of your five days.

**Clarifying email for a take-home case study**

```
Subject: Case study: four quick questions before I start

Hi [name],

Thanks for sending the case. I'm planning to have the deck to you by [date]. Four questions, each with the assumption I'll use if it's easier not to answer:

1. Metric: how is [conversion] defined in this data (numerator and denominator)?
   If I don't hear back, I'll use [completed orders / sessions] and state it on the assumptions slide.

2. Format: do you want PowerPoint, and is there a slide limit?
   If I don't hear back, I'll send PowerPoint, 8 to 12 slides.

3. AI tools: are AI tools allowed on this assignment, and if so, how should I disclose their use?
   Until I hear back, I won't use any.

4. Priority: if time runs short, which matters more to you: depth on the main question or breadth across possible causes?
   If I don't hear back, I'll go deep on the main question and list the rest as next steps.

Thanks,
[your name]
```

*Keep it to one email; swap in your own metric and deliverable, and keep the fallback line under every question.*

In the case, the reply answered the format and AI questions and said nothing about the metric. So the fallback definition went onto the assumptions slide, which is exactly what the fallback line is for.

> **Tip:** The fallback column is your assumptions slide
>
> Every fallback you logged in the email becomes a line on the assumptions slide. When the panel challenges an assumption, you can show it was stated in writing before you started.

## How to scope a take-home assignment: the five-line plan

A scope plan is five lines: the question, the key metric, the time period, the scope, and the deliverable. Under it goes an out-of-scope list of things you will deliberately not build. Prep guides consistently say to define the key metric (revenue, conversion rate, churn), the time period, and the scope before running any numbers.

**Five-line scope plan with out-of-scope list**

```
QUESTION:     [the business question, in one sentence, as the brief states it]
METRIC:       [one key metric, with the definition you are using]
PERIOD:       [the time window you will analyse, and the comparison window]
SCOPE:        [the 2 or 3 analyses you will run, in priority order]
DELIVERABLE:  [format, 8 to 12 slides, due date and time]

OUT OF SCOPE (and why):
- [item] :: [would not change the recommendation / cannot finish inside the cap]
- [item] :: [reason]
- [item] :: [reason]
```

*Fill this in after the email is sent and before any analysis; every out-of-scope item needs a one-line reason.*

In the case, the scope line held two analyses: where and when the drop happened, and which segment drove it. The out-of-scope list had four items:

- a forecast for next quarter
- a customer-level model
- an interactive dashboard
- cleaning every edge case in the data

The out-of-scope list does three jobs. It protects the cap. It becomes the next-steps slide. And it gives you a ready answer when someone on the panel asks why you did not look at something.

## The hour log, step by step

This is the procedure, followed by the case's actual hour log. Each step has a time budget and a "done" condition, so you always know whether you are on plan.

#### From brief to panel

1. **Read the brief and estimate hours** - Spend 30 to 45 minutes listing each piece and writing an hour number beside it. Done is a written estimate and a stated cap.
2. **Send one clarifying email** - Cover metric definitions, deliverable format, AI policy, and priorities if time runs short. Done is sent, dated, and a fallback logged per question.
3. **Write the scope plan and out-of-scope list** - Define question, metric, period, scope, and deliverable in five lines. Done is the plan plus a reasoned out-of-scope list.
4. **Run the analysis inside a time box** - Work roughly 2 to 4 hours, re-estimating at the end of hour one. Done is the core answer exists and the cap is reached.
5. **Build the deck** - Spend 1 to 2 hours on 8 to 12 slides that lead with the problem and the answer. Done is slide one naming one action.
6. **Submit with a time-box line** - Send the deck with one line stating your cap and what you would do next. Done is the email sent.
7. **Prepare the Q&A** - Write the ten likely questions and an answer to each, then rehearse aloud. Done is a written answer per question.
8. **Run the panel** - Present for 20 to 30 minutes to 2 to 5 people, then take 10 to 25 minutes of questions. Done is every recommendation defended live.

### The case's hour log

- **0:00 to 0:40, read and estimate.** Written estimate: 5 to 6 hours against a stated 3. Caps set at 4.5 build hours and 6 hours overall.
- **0:40 to 1:00, clarifying email.** Sent and dated, with four fallbacks logged.
- **1:00 to 1:30, scope plan.** Five lines, and four items on the out-of-scope list.
- **1:30 to 2:30, analysis one.** The conversion trend across the period, confirming when and where the drop shows. End-of-hour re-estimate: on plan.
- **2:30 to 3:30, analysis two.** The drop cut by segment, plus a written list of competing causes, including a data or tracking change and fraud.
- **3:30 to 4:00, the wrong turn.** I started a customer-level model because the data made it possible. At the hour-four check I asked one question: would this change the action on slide one? It would not, and it could not finish inside the cap. It was cut and moved to the next-steps list.
- **4:00 to 5:30, deck.** Ten slides.
- **5:30 to 5:40, submit.** The email included this line: "I capped this at about 5.5 hours, including setup. With more time, the next thing I would build is a customer-level model to test whether the drop is concentrated in new or returning buyers."
- **Stop at 5:40,** inside the 6-hour cap. Build time was 4 hours, inside the 4.5-hour cap.
- **Day three, separate block: 1.5 hours of Q&A prep.**

The hour-four wrong turn is the most common way these projects go over time. The analysis looks rigorous and feels like effort, but it does nothing for the answer. The opposite risk is real too: one candidate was criticised for not building more, despite being given a 4-hour limit. The priority question in the clarifying email is your defence against both. Whatever they answered, or whatever fallback you logged, is your written reason for the depth you chose.

## What to put in a take-home case study deck

Put the problem and your one recommended action on the first slide. Follow it with context, objective and assumptions, and a summary of recommendations, then one slide per analysis, then risks and next steps. Aim for 8 to 12 slides in total. With 20 to 30 minutes of presenting time, that is two to three minutes per slide, which leaves room for interruptions.

The case's ten slides:

1. The problem and the one action
2. Context
3. Objective and assumptions (taken from the fallback log)
4. Summary of recommendations
5. Analysis one: when and where the drop happened
6. Analysis two: which segment drove it
7. Competing causes considered, and how each was ruled in or out
8. The recommendation in detail, and what it would take
9. Risks
10. Next steps (the out-of-scope list, in order)

Some rubric dimensions have been published, but no weights. Scale AI scores three things: structure, quantitative reasoning, and a clear recommendation. A PM hiring manager's rubric, published by Aakash Gupta, asks how the document compares to internal documents, whether it shows clear insight, and whether it shows business sense. EasyApply publishes a weighted five-dimension take-home rubric, but only one dimension, Problem Understanding, was visible, and its weights were cut off. No published weights exist for framing, assumptions, rigour, storytelling, or Q&A. Treat any claimed percentage breakdown with suspicion.

#### What the panel reads, outermost first

1. **Recommendation** - One action, named on slide one
2. **Quantitative reasoning** - One slide per analysis, each tied to the action
3. **Structure** - Context, objective, summary, risks, next steps
4. **Assumptions** - Every fallback from the email, stated up front

*The panel meets the recommendation first and reaches the assumptions last, so every layer has to hold up when the one above it is challenged.*

## Questions asked after a take-home presentation

Expect challenges to your assumptions, questions about why you chose one approach over another, probes into the detail of your analysis, live questions about why you reached each recommendation, and an open closing question. Write an answer to each of ten likely questions before the panel. The Q&A is where the work actually gets tested.

| Question type | Reported at | What it tests |
|---|---|---|
| Challenge to an assumption | Intuit | Whether you stated it and can defend it |
| Why this approach over another | Intuit | Whether you made a choice or defaulted |
| Probe into analysis detail | Intuit | Whether you did the work yourself |
| Why this recommendation | Datadog | Whether the action follows from the numbers |
| "Is there something we haven't asked you that we should?" | PrepLounge report | Whether you can steer to your strength |

Intuit panels also reportedly interrupt in the middle of the walkthrough, which is another reason not to pack slides. Adobe gives 10 to 25 of its 45 to 60 minutes to questions. And the AI policies Joberney collected all come down to a live defence of the work. If you used AI with permission, the Q&A is where that use either holds up or falls apart.

Management Consulted cites a Harvard Business Review study claiming 76% of perception comes from delivery. The primary source is unverified, so I would not plan around that number. Rehearsing answers out loud is worth doing for a simpler reason: a written answer you have never said aloud tends to come out longer than it should.

### The case's ten questions

1. Why this definition of conversion?
2. Why this time period and comparison window?
3. Could the drop be a tracking or data change rather than real behaviour?
4. Could it be fraud?
5. Why did you cut by this segment and not another?
6. Why this recommendation over the runner-up?
7. What would it cost, and who would own it?
8. What would make you change your mind?
9. What would you do with another week?
10. Did you use AI tools, and how?

Questions 3 and 4 are the ones most likely to go badly, because they test whether you considered a cause you did not chart. A DoorDash candidate missed fraud as a hypothesis and was rejected. Slide 7 exists so that your answer can start with "yes, here is how I ruled it in or out." Question 9 is answered by the out-of-scope list.

For the closing question, have one example from your own history ready that connects to the posting's hardest requirement. Finding that requirement is the friction: [Refolk](/candidates) scores how well your history fits the posting, and the weakest-fitting requirement is the one a panel is most likely to probe. That is where your closing answer should go.

### Who is likely to be in the room

**34.3 - US data analysts for every US analytics manager**

In Refolk's index there are 65,223 current Data Analyst profiles in the United States and 1,901 current Analytics Manager profiles.

| Segment in Refolk's index | Count | Exact-title share, top-25 sample |
|---|---|---|
| Data Analyst, US | 65,223 | 15/25 (60%) |
| Data Analyst, UK | 16,464 | 21/25 (84%) |
| Analytics Manager, US | 1,901 | 18/25 (72%) |

That ratio, combined with panels of 2 to 5 people, suggests most panels are mostly peers with one manager. This is an inference, not a documented fact. The practical consequence is to expect detail questions from people who do this analysis every day, and one question about the decision from the person who owns it. The sample shares come from only 25 profiles, so treat them as indicative. The UK pool is about a quarter the size of the US one, and some UK employers in the sample are small companies whose invites may follow a different format. That part is unverified.

> The deck gets you into the room. The ten written answers decide what the room thinks of it.

## How this goes wrong: failure modes and false signals

Most take-home failures come from one of seven patterns: trusting the company's hour estimate, gold-plating, waiting on a reply, analysis without a decision, a missed alternative explanation, AI use you cannot defend, and reading a warm room as a pass. Each one has a check you can run before the panel.

- **Using the company's hour estimate as your plan.** The person who designed the exercise already knows the answer, and reports suggest overruns of 2x are normal. What it looks like when it fails: still on analysis one at hour three. Check: re-estimate at the end of hour one and cut scope, not sleep.
- **Gold-plating.** Extra breadth can read as poor judgment. Final Round AI says eight hours on a three-hour task signals someone who cannot scope work or make tradeoffs. What it looks like when it fails: an extra model that does not change slide one, like the case's hour-four wrong turn. Check: ask in the clarifying email what to prioritise if time runs short, and test every new analysis against the action.
- **Waiting for a reply to the clarifying email.** Reply speed is unknown, so a stalled email eats days. What it looks like when it fails: day three, no reply, and nothing built. Check: log a fallback assumption beside every question and start on it.
- **Analysis without a decision.** The DoorDash candidate listed metrics but was not decisive about what action should be taken. What it looks like when it fails: a strong deck whose summary slide says "it depends." Check: slide one names one action.
- **Missing the alternative explanation.** The same candidate missed fraud as a hypothesis. What it looks like when it fails: a confident causal story with no "ruled out" slide. Check: list the competing causes, including data changes and fraud, before you present.
- **AI use you cannot defend.** Datadog tells candidates to expect live "why" questions. What it looks like when it fails: a recommendation you can read out but cannot explain. Check: explain every recommendation aloud without your notes.
- **Reading a positive room as a pass.** In candidate reports on Blind, interviewer warmth does not predict the outcome. What it looks like when it fails: relaxing your other loops after a friendly panel. Check: treat warmth as noise and keep your search running.

#### Breadth against decisiveness

Horizontal axis runs from Narrow analysis to Broad analysis. Vertical axis runs from Vague conclusion to One clear action.

| Quadrant | What it means |
| --- | --- |
| Narrow and decisive | The target: defend scope with the out-of-scope list |
| Broad and decisive | Over the cap: cut what does not move slide one |
| Thin and vague | Reads as not enough effort: add the competing-causes slide |
| Broad and vague | The DoorDash pattern: pick one action and lead with it |

*The panel rewards a narrow analysis that ends in one action over a broad one that ends in a list.*

> **Watch out:** Effort is judged both ways
>
> One Blind poster reports failures on PM take-homes put down to "not enough effort," while Final Round AI treats eight hours on a three-hour task as a scoping failure. The only safe position is a stated cap, a written priority, and a deck that shows both.

## After the panel: what to do next

Log the date of the panel, keep your other applications moving, and do not read either silence or warmth as a verdict. Decision timing is reported only anecdotally, and detailed feedback is rare for legal reasons, not for reasons about you.

Here is what is known about timing. One person was rejected a week after submission with no explanation. A media-company data analyst final round took about a week. The median time from submission to decision is not established publicly. One PM on Blind reports a 70% pass rate on take-homes, but that is one person's count and should not be treated as a base rate.

Practitioners on Hacker News report that corporate counsel will basically refuse to let hiring teams tell a candidate why they were rejected. So your post-mortem has to come from candidate reports. These are the recurring rejection reasons:

- not committing to an action
- missing a hypothesis such as fraud
- "not enough effort"
- no rapport with the panel
- losing to a candidate with extra skills

Hold your deck up against that list before your next take-home, not after it.

The next application needs a resume and cover letter tailored to a new posting. Doing that by hand while waiting on a decision is where searches stall. Refolk drafts both from your own history and scores your fit for each posting, so you can keep sending applications while this one is pending.

#### Before the panel slot

- [ ] A written hour estimate exists and was made before the data was opened
- [ ] The clarifying email was sent and dated, with a fallback assumption beside every question
- [ ] The AI policy is confirmed in writing or stated in the brief, and any use is disclosed as it requires
- [ ] The scope plan has five lines and the out-of-scope list gives a reason for each item
- [ ] The hour log stops at or before the stated cap
- [ ] Slide one names one action
- [ ] The deck is 8 to 12 slides, with assumptions stated before the first analysis
- [ ] A slide lists the competing causes, including data changes and fraud, and how each was ruled in or out
- [ ] The submission email states the cap and the next thing you would build
- [ ] A written answer exists for each of ten likely questions, rehearsed out loud
- [ ] You can explain every recommendation without notes

## Frequently asked questions

### How many hours should a take-home case study take?

Plan for more than the brief says, then cap it. EasyApply tells employers that a task taking them 3 hours takes candidates 4 to 6, and practitioner caps mostly fall between 2 and 8 hours. Final Round AI calls 4.5 hours on a 3-hour assignment fine and treats 8 hours as a scoping failure. Write your own estimate and a hard cap before you open the data, and state the cap when you submit.

### Is it OK to ask clarifying questions on a take-home case study?

Yes. Send one email with every question in it, since practitioners note that asking for clarification may be part of what is being tested. Ask about metric definitions, deliverable format, AI policy, and what to prioritise if time runs short. Log a fallback assumption beside each question and keep working, because nobody has published how fast recruiters reply, and waiting can eat days of your window.

### Can I use AI tools on a take-home assignment?

Only if the employer allows it, and policies conflict. Code.org allows AI if you disclose it to the interview team, GitLab requires AI use to be cited and answered in your own words, and Anthropic's guidance has said take-homes should be done without Claude unless stated otherwise. Confirm in writing before you start, and make sure you can defend every recommendation live without notes.

### What questions do panels ask after a take-home presentation?

Expect challenges to your assumptions, questions about why you chose one approach over another, probes into the detail of your analysis, and live questions on why you reached each recommendation. Intuit panels reportedly interrupt mid-walkthrough. A documented closing question is whether there is something they have not asked that they should. Nobody has published how often each type comes up, so prepare written answers to the ten most likely.

### How many slides should a take-home case study deck have?

Aim for 8 to 12. Lead with the problem and your answer, then add slides for context, objective and assumptions, and a summary of recommendations, then one slide per analysis, then risks and next steps. With 20 to 30 minutes of presenting time, that works out to two or three minutes per slide, which leaves room for interruptions.

### Why didn't I get feedback after being rejected on a take-home?

Usually because of legal caution, not because of anything about you. Practitioners report that corporate counsel will refuse to let hiring teams tell candidates why they were rejected. Candidate reports name recurring reasons instead: not committing to an action, missing an alternative hypothesis such as fraud, not enough effort, and no rapport with the panel. Check your own deck against those.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/take-home-case-study-panel-log*