Tasks / Discover

Customer research call guide

Can the model write a guide that uncovers behaviour rather than opinions?

Measures the modelTask v1.0 · 2 casesDifficulty

What AI gets right here, and what you’ll still have to catch

From 9 graded outputs by 5 models. 50% were usable with at most a quick edit.

Reliably right

  1. Identifies material uncertainty100% pass
    Names specific unknowns (cost of lateness, ease of chasing, automation fit, willingness to pay) and says mixed findings mean targeted follow-up.
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?
  2. Designed to falsify100% pass
    Every learning goal has disconfirming questions, e.g., 'What did you do next?', 'When do you avoid sending another reminder?'
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?
  3. Asks what people did, not what they think100% pass
    Core questions ask for the most recent overdue invoice and specific actions taken, not opinions or predictions.
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?

Where it slips

  1. Fits the call33% pass
    Timings add up to 30 minutes but no must-ask questions are marked and no guidance on what to cut if time runs short.
    GPT-6.1 Sol · API · Would freelancers pay to stop chasing invoices?
  2. Respects explicit constraints81% pass
    The guide is clearly over 900 words, violating the explicit length constraint.
    Opus 5.5 · Claude · Would freelancers pay to stop chasing invoices?
  3. Different questions for user and signer81% pass
    The guide does not provide distinct question tracks for the daily warehouse operations manager and the executive who signed; it only differentiates churned vs at-risk customers, with the same questions for both roles within each track.
    GPT-6 Luna · API · Why are warehouses leaving? Two execs, two theories

Case viewer

Read the brief, then put up to three outputs side by side, each with the LLM judge’s verdict on every check. Highlights mark what a PM had to fix.

The brief

You're a PM at Paydeck. Next week, Mei (a designer) and Tomas (an engineer) will run 8 customer calls to test whether we should build Autochase next quarter. Neither has run a customer interview before. Write the call guide they'll use: the learning goals, the questions with rough timings for a 30-minute call, notes for first-time interviewers, who we should talk to, and what we'd need to hear to go ahead or to stop. Keep it under 900 words. Sam, the PM who owns Autochase, has drafted some questions and a recruiting plan. They're below with everything else we know.

About PaydeckInvoicing software for freelancers and small agencies (1 to 10 people). 48,000 accounts, 6,100 of them paying $12 a month for Pro.
The ideaAutochase: automatic, escalating reminders for overdue invoices, sold as a $10-a-month add-on. There is no prototype yet.
Sam's hypothesis“Late payments cost our users real money, and chasing them is painful enough that they'd pay $10 a month to automate it.”
What we know34% of invoices sent through Paydeck are paid after their due date; the median is 16 days late. The manual 'Send reminder' button was used by 12% of accounts in the last 90 days. 40 support tickets last quarter asked for automatic reminders.
Sam's draft questions1. How frustrating do you find chasing late payments? 2. Would an automatic reminder feature be useful to you? 3. How much would you pay for it: $5, $10 or $15 a month? 4. Do you think your clients would be annoyed by automated reminders? 5. What other features would you like us to build?
Sam's recruiting planEmail the 300 accounts that used the 'Send reminder' button most last month, and book the first 8 who reply.
What a strong answer does

A guide that splits Sam's hypothesis into goals that can each fail: do late payments actually cost users time or money, how do they deal with them today, would they pay to fix it, and does chasing harm client relationships. The core asks for the story of the last overdue invoice (what happened, what they did, how long it took, what it cost) and what they've already tried or paid for, instead of opinions or predictions. It replaces Sam's leading and hypothetical questions: no 'how frustrating', no 'would it be useful', no price menu; payment is tested through past spending or a real commitment (say, a paid pilot sign-up) at the end, after the problem questions and without pitching Autochase first. It fits 30 minutes with timings, coaches first-timers (ask for specifics, follow up with 'what happened next', stay quiet, don't pitch or defend), and fixes the recruiting plan: the heaviest reminder users will over-state demand, so it adds accounts with overdue invoices that never send reminders, and says who to exclude or balance. It ends with clear go and stop signals tied to what people did, such as most calls describing recent chasing that took real time plus a workaround they've tried or paid for, against most saying one nudge does the job.

Critical failures (cap the score)
  • Relies on hypothetical questions (would you use it, how much would you pay) as the main evidence for the decision
  • Describes or pitches Autochase before asking how the person handles late payments today
Case

v1.0 · synthetic · B2B SaaS, freelancers, willingness to pay, first-time interviewers

Usable with a quick edit11 of 12 checks passed

The guide is complete and usable; a quick edit to add a note about cutting the check-understanding section if time is short would fix the minor omission.

Output

Autochase: 30-minute customer call guide

Interviewers: Mei leads; Tomas takes notes and asks brief follow-ups at the end. Keep roles consistent across the eight calls.

Decision: Should Paydeck prioritize automatic, escalating overdue-invoice reminders as a $10/month add-on next quarter?

1. Learning goals

Test whether: - Late payments create recurring, material costs—not just irritation. - Chasing takes enough time or money that customers actively try to reduce it. - Reminders address the cause of delays, rather than disputes, approval processes, or clients lacking funds. - Customers would trust automation, and what control they require. - There is credible demand at $10/month, beyond polite enthusiasm.

Starting evidence is suggestive, not conclusive: 34% of invoices are late, with a median delay of 16 days. That does not tell us how many accounts experience serious pain. Reminder usage (12% of accounts) may miss chasing outside Paydeck. The 40 support requests show interest, not willingness to pay.

2. Who to recruit

Do not book the first eight respondents from the 300 heaviest reminder users. That would favor highly engaged users already chasing through Paydeck.

Recruit purposively: - 3 accounts: recurring overdue invoices; frequent manual reminders. - 3 accounts: recurring overdue invoices; little or no manual reminder use. - 2 accounts: mostly on-time invoices, as a contrast group.

Across those groups, aim for six Pro and two free accounts, with both freelancers and small agencies. Include no more than two support-ticket requesters. Speak to whoever actually handles collections; establish whether they also approve software spending.

Randomly invite accounts within each group, then fill group slots rather than accepting the first eight overall. Replace no-shows within the same group. Use a neutral invitation: “Help us understand how you manage invoices and payments.” If offering an incentive, keep it fixed and unrelated to feedback.

3. Call script and timings

0–3 minutes: Welcome and permission

“Thanks for helping. We’re learning how people manage payments—not testing you or selling anything. Honest criticism is useful. You can skip anything confidential.”

Ask permission before recording. Otherwise, take notes.

“What’s your role, who handles overdue invoices, and roughly how many invoices do you send each month?”

3–12 minutes: Reconstruct a real incident

“Tell me about the most recent invoice that wasn’t paid by its due date.”

Follow the timeline: - “When was it due, and when did you notice?” - “What did you do next? Then what happened?” - “Who else was involved?” - “What reason did the client give for the delay?” - “Was it paid? When?”

“Could you walk us through a redacted reminder or your tracking process?” Do not request client-identifying information.

If they have no recent late invoice: “Walk us through how your last invoice got paid. What helps payments arrive on time?”

12–19 minutes: Frequency, consequences, alternatives

“In the last three months, about how often has this happened?”

“For that example, roughly how much time did you spend chasing? What, if anything, did the delay cost you?”

“What happens if you don’t follow up?”

“How do you decide when and how firmly to chase? Have you ever chosen not to?”

“What have you tried to improve this? What did it cost, and why did you keep or stop using it?”

If relevant: “Do you use Paydeck’s reminder button? Why or why not?”

19–23 minutes: Automation fit and risks

“Which parts of that process need your judgment?”

“Tell me about a follow-up that went badly—or worked particularly well.”

“What situations would make sending another reminder inappropriate?”

Probe only after open answers: payments already made, disputed invoices, important relationships, tone, timing, approvals.

23–28 minutes: Concept and price

Only now introduce the idea:

“We’re considering Autochase: automatic reminders that escalate while an invoice remains overdue. There’s no prototype, and details aren’t decided.”

“For the invoice we discussed, where would this help, fail, or create extra work?”

“What would you need to control before trusting it?”

“The proposed price is $10/month, on top of your current plan. How would you decide whether to buy it? What would it replace? Who approves that spending?”

Ask interested participants whether they will join a follow-up pilot discussion. Record this as interest, not a purchase.

28–30 minutes: Check understanding

Summarize their situation and ask what you misunderstood. Tomas asks outstanding clarifications. Thank them.

4. First-time interviewer notes

  • Ask one question at a time; allow silence.
  • Prefer “Tell me about the last time” to predictions.
  • Don’t defend Autochase, suggest answers, or praise positive reactions.
  • Separate quotes, observed behavior, estimates, and interpretation.
  • Debrief immediately: pain, frequency, cost, workaround, automation fit, price evidence, contradictions.

5. Decision rules

Agree these directional thresholds before interviewing:

Advance: At least four of six overdue-account participants demonstrate recurring, consequential pain and reminder-suitable workflows; at least three Pro participants show credible $10 value grounded in current costs or spending and accept a concrete follow-up. Proceed to a scoped pilot and real pricing test—not a full-build commitment from interviews alone.

Stop or rethink: Pain is mostly minor, delays are not reminder-solvable, relationship risks outweigh savings, or interest disappears at $10.

Mixed evidence: Narrow the target segment and recruit more. Eight purposive interviews cannot estimate market demand.

Check by check

Got wrong · 1
  • Fits the callThe guide does not mark must-ask questions or say what to cut if time runs short, as required by the criterion.
Got right · 11
  • Uses the supplied evidence correctlyAll factual claims about the current situation are taken directly from the brief or supplied context.
  • Addresses the actual decisionThe guide commits to clear go/stop decision rules with thresholds and says what would change the answer.
  • Respects explicit constraintsThe output is under 900 words, includes all requested sections, and respects the form and reader.
  • Identifies material uncertaintyIt names specific unknowns (e.g., whether pain is minor, delays not reminder-solvable) and says how the interviews will resolve them.
  • Avoids unsupported claimsInterpretations are clearly labelled as suggestive, and no unsupported claims are presented as established fact.
  • Produces the required deliverableThe guide is a complete, usable call guide with learning goals, timed questions, interviewer notes, recruiting plan, and decision rules.
  • Talks to the right peopleIt explicitly avoids only the heaviest reminder users, explains why, and balances with users who have overdue invoices but don't chase in Paydeck.
  • Coaches first-time interviewersIt gives specific, usable instructions: ask one question at a time, allow silence, prefer past-behaviour questions, don't defend or pitch, and separate notes.
  • Designed to falsifyEvery learning goal has disconfirming questions, such as reconstructing a real late invoice and asking what happens if they don't follow up.
  • Asks what people did, not what they thinkCore questions ask for the most recent overdue invoice and what they did, with opinions and predictions kept secondary and at the end.
  • Doesn't lead the witnessQuestions are neutral and open; Autochase is introduced only after the problem exploration, with no pitching beforehand.

Grades and run details

Decision model 96 · LLM judge 12 of 13 checks
Decision model checks
  • passUses the supplied evidence correctly64%
  • passAddresses the actual decision52%
  • passRespects explicit constraints45%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims92%
  • passProduces the required deliverable94%
  • passTalks to the right people100%
  • passCoaches first-time interviewers86%
  • passDesigned to falsify89%
  • passAsks what people did, not what they think90%
  • passDoesn't lead the witness86%
  • partialFits the call51%
Run
Run
#1
Time to output
38 s
Submitted
30 Sept 2026

Needs more than a quick edit9 of 12 checks passed

The guide is strong and mostly usable, but it needs evidence fixes and a must-ask/cut plan before a PM could run it without re-checking assumptions.

Output

Autochase Call Guide: 8 calls, 30 minutes each

What we're trying to learn

Sam's hypothesis has three parts, and each can fail separately: 1. Cost: Late payment hurts in money or time, not just annoyance. (34% of invoices are late, but that doesn't mean it hurts.) 2. Behavior: People already try to fix it. Only 12% of accounts used the manual Send reminder button, so we need to know why the other 88% didn't. 3. Willingness: They'd let software message their clients, and pay $10 a month on top of $12 for Pro.

Interviews can't settle price. They can tell us whether it's worth testing.

Who to talk to

Sam's plan (the 300 heaviest button users, first 8 to reply) would only find people who already like reminders and respond fast to email. Recruit against quotas instead:

  • 2 heavy button users (Sam's list is fine for these)
  • 3 accounts with several late invoices in the last 90 days who never used the button
  • 1 who filed an Autochase support ticket
  • 2 agencies (3+ people) and 2 solo freelancers. These overlap with the groups above.
  • Mix of Pro and free accounts, with at least 2 free.

Screen for at least one invoice paid late in the last 90 days. Pick people to fill the quotas, not first come first served. Offer a $50 gift card.

Call structure

0:00–0:03 Intro. "We're learning how people get paid, and we're not selling anything. There are no right answers, and critical feedback helps most. Okay to record?"

0:03–0:05 Context. What do you do, how many clients, how many invoices a month?

0:05–0:15 The last late invoice (the core of the call) 1. Tell me about the last time a client paid you late. Probe: how late, how much, what happened first, what did you do, how did it feel? 2. Walk me through what you do between sending an invoice and getting paid. Probe: what tool, how often, who does it? 3. What did that lateness cost you? Probe: hours, cash flow, bills or contractors you had to delay.

0:15–0:21 Current behavior 4. (Button users) What made you use Send reminder? (Non-users) What do you do instead, or why not just let it go? 5. Is there a client you would never chase? Why? 6. Have you tried anything else, such as late fees, deposits, shorter terms or other tools? What happened?

0:21–0:27 Concept. Read this, don't demo it: "Suppose Paydeck automatically emailed your client at set points after the due date, getting firmer each time." 7. What's your first reaction? What worries you? 8. What would have to be true for you to switch it on? Probe: tone, control, per-client exceptions, seeing the emails first.

0:27–0:30 Money and wrap 9. What do you pay for today that helps you get paid or run your finances? 10. "We'd likely charge extra for this. Would you want early access when it's ready?" Note the exact response. Hesitation counts as data. 11. Anything I should have asked? Anyone else we should talk to?

Why Sam's questions changed

  • "How frustrating is chasing?" and "Would it be useful?" invite polite yes answers. Past-behavior questions don't.
  • "$5, $10 or $15?" anchors people on our options and measures nothing real. Questions 9–10 are better evidence.
  • "Would your clients be annoyed?" asks for a guess about someone else. Question 5 asks what they've actually avoided.
  • "What other features?" is off-topic. Drop it.

Notes for first-time interviewers

  • Split roles. One person asks and the other takes verbatim notes. Swap halfway through the 8 calls.
  • Ask about the past, not the future. "What did you do last time?" beats "Would you ever...?"
  • Don't pitch or rescue. Don't name Autochase until minute 21. If they struggle to answer, wait, then ask "Can you say more?" Silence is fine.
  • Follow the story. Ask "why?" and "what happened next?" more than you ask the next scripted question. The script is a backstop.
  • Discount compliments. "That sounds great" is worth little. Specific past actions, money spent and workarounds built are worth a lot.
  • Stay neutral if they criticize the product or the idea. Don't defend it.
  • Debrief within 30 minutes of each call. Write down three surprises and the best quotes. Do a practice run with a colleague first.

What we'd need to hear

Count only the people who had a late invoice.

Go (build, or run a paid pre-sale test): - At least 5 of 8 describe a specific recent late payment that cost real money, stress or more than about an hour a month. - At least 5 already chase manually or with workarounds, including people who skip the button, and can explain why the button doesn't work for them. - At least 4 would turn on automation given controls we could build, and at least 3 say yes to early access for a paid add-on.

Stop or rethink: - Most shrug lateness off, or say it comes from a few clients they'd never automate against. - The real problem is something else, such as clients not opening invoices, wanting deposits, or needing a payment link, and reminders wouldn't fix it. - Most won't let software contact their clients, or wouldn't pay anything for it.

Mixed results (for example, it only matters for agencies, or only with tone controls): narrow the target segment and run a second round of calls.

Even a clear "go" from 8 calls only justifies a cheap next test, such as a fake-door button or a pre-sale. It doesn't justify committing the quarter.

Check by check

Got wrong · 1
  • Fits the callTimings add to 30 minutes, but it does not mark must-ask questions or say what to cut if time runs short.
Mixed · 2
  • Uses the supplied evidence correctlyIt uses most supplied numbers correctly, but invents or overstates current-situation details such as an 'Autochase support ticket' population and treats recruiting-plan inferences as facts.The two graders disagreed on this one.
  • Avoids unsupported claimsIt presents some interpretations as established, especially that Sam's list would only find reminder-likers/fast email responders and that Autochase tickets exist.The two graders disagreed on this one.
Got right · 9
  • Addresses the actual decisionIt commits to a clear call: 8 calls can justify only a cheap next test, not committing the quarter, and gives go/stop conditions.
  • Respects explicit constraintsIt is a call guide under 900 words with learning goals, timed questions, interviewer notes, recruiting, and go/stop signals.
  • Identifies material uncertaintyIt names material unknowns—cost, behavior, willingness, client-contact concerns—and says how calls plus a cheap test would resolve them.
  • Produces the required deliverableMei and Tomas could run the calls from it with light edits.
  • Talks to the right peopleIt explains the heaviest-button-user bias and gives a quota mix including non-button users with late invoices, ticket requesters, agencies, solos, and free accounts.
  • Coaches first-time interviewersIt gives specific first-timer instructions: role split, verbatim notes, silence, follow-ups, no pitching/defending, debriefs, and a practice run.
  • Designed to falsifyEach learning goal has disconfirming questions about recent late payments, actual chasing behavior, client-contact limits, and past spending/early-access response.
  • Asks what people did, not what they thinkThe core asks for the last late invoice, what happened, what they did, what it cost, and what they already tried.
  • Doesn't lead the witnessIt removes Sam's leading/hypothetical questions, keeps the concept late, and avoids pitching before current-behavior questions.

Claims the judge couldn’t find in the brief

  • There is an 'Autochase support ticket' population to recruit from.
  • Sam's plan would only find people who already like reminders and respond fast to email.
  • We can identify accounts with several late invoices in the last 90 days who never used the button.

Grades and run details

Decision model 96 · LLM judge 9 of 13 checks
Decision model checks
  • passUses the supplied evidence correctly26%
  • passAddresses the actual decision42%
  • passRespects explicit constraints47%
  • passIdentifies material uncertainty100%
  • passAvoids unsupported claims62%
  • passProduces the required deliverable96%
  • passTalks to the right people99%
  • passCoaches first-time interviewers100%
  • passDesigned to falsify95%
  • passAsks what people did, not what they think82%
  • passDoesn't lead the witness60%
  • partialFits the call44%
Run
Run
#1
API response time
32 s
Submitted
30 Sept 2026

Results

Every setup we’ve tested on this task, across all cases and repeats, graded on the current checklist. Provisional The checklist is still being calibrated against our PM.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstrawithChatGPT95.892.31None
2GPT-6.1 SolwithAPI92.192.62None
3Opus 5.5withClaude94.169.22None
4Sonnet 5.5withAPI86.484.62None
5GPT-6 LunawithAPI90.273.921 capped

About the task

The PM job

Preparing for customer calls.

Why it matters

Leading questions produce the answers you wanted. A good guide is designed to prove you wrong.

What good looks like

  • Past-behaviour questions
  • No leading prompts
  • Clear learning goals
  • Could disprove the team's hypothesis

Deliberately not measured

    Capability tested

    Question design

    The failure we’re looking for

    Hypothetical, leading questions

    Grading

    Decision model and LLM judge, calibrated against a blind PM review