Tasks / Discover

Extract discovery insights

Can the model separate evidence, themes and hypotheses without inventing consensus?

Measures the modelTask v1.1 · 2 casesDifficulty

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

From 14 graded outputs by 7 models. 50% were usable with at most a quick edit.

Reliably right

  1. Produces the required deliverable100% pass
    The summary is complete, usable, and would let a PM act with only light edits.
    GPT-6 Astra · ChatGPT · Eight calls with finance teams
  2. Keeps dissent visible100% pass
    The SaaS and charity close-is-fine views and the manufacturing CFO's switching regret are kept visible.
    GPT-6 Astra · ChatGPT · Eight calls with finance teams
  3. Respects explicit constraints95% pass
    It is a findings summary for the product team and is under the 600-word limit.
    GPT-6 Astra · ChatGPT · Eight calls with finance teams

Where it slips

  1. Says how many sources support each finding46% pass
    Findings are presented without source counts (e.g., 'five of eight'), using vague terms like 'consistent reports' instead of sizing each theme.
    Gemini 3.8 Flash · API · Eight calls with finance teams
  2. Avoids unsupported claims50% pass
    The bottom line asserts 'Adoption will depend on handling messy data and reducing switching effort' as an established forecast, though the evidence only suggests it.
    GPT-6.1 Sol · API · Eight calls with finance teams
  3. Uses the supplied evidence correctly63% pass
    The output states the calls happened in August 2026, but the supplied context only says August without a year; this date is an invented fact not supported by the brief.
    Opus 5.5 · Claude · Eight calls with finance teams

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

Synthesise the eight discovery calls below with finance leads about month-end close. Write a findings summary for the product team, who are deciding whether to build a reconciliation product: what we learned, and how confident we can be in it. Keep it under 600 words.

What the model was given9 items: Scenario, Call 1: Financial controller, logistics company (180 staff), Call 2: Head of finance, SaaS company (95 staff), Call 3: Finance manager, retail chain (400 staff), Call 4: CFO, manufacturing company (250 staff), Call 5: Accountant, agency group (120 staff), Call 6: Finance director, charity (70 staff), Call 7: Controller, hospitality group (300 staff), Call 8: Financial controller, construction firm (210 staff)
ScenarioWe make spend-management software for mid-sized companies and are exploring a reconciliation product. Our PM ran these 40-minute calls in August. Below are the relevant excerpts from each, lightly edited.
Call 1: Financial controller, logistics company (180 staff)“Close takes us eight working days and at least three of those are reconciliation. The bank feed drops transactions, so we tick and bash against statements in Excel.” Asked what she'd pay to fix it: “If it gave me two days back, it pays for itself. But I've heard that before.”
Call 2: Head of finance, SaaS company (95 staff)“Honestly close is fine. We moved to a proper ERP two years ago, it's five days and nobody's complaining.” Later: “The only annoying bit is chasing people for receipts, not reconciling.”
Call 3: Finance manager, retail chain (400 staff)“Card transactions are the nightmare. Hundreds of store cards, the statements come in as PDFs, someone re-keys them. We found £14k of duplicates last quarter.” She reconciles card spend in a shared spreadsheet with four people editing it.
Call 4: CFO, manufacturing company (250 staff)“We switched reconciliation tools last year and I regret it. The migration took four months, we lost all our matching rules, and half the team quietly went back to spreadsheets.” Asked if he'd switch again: “Not unless someone else does the migration for us. Never again otherwise.”
Call 5: Accountant, agency group (120 staff)“Intercompany is where it breaks. Three entities, money moving between them every week, and the balances never agree at month end. I spend a day and a half on it.”
Call 6: Finance director, charity (70 staff)“Close is fine for us. We're small, two of us, it takes four days and we know every transaction.” When asked about tools: “We'd not pay for anything new this year; budgets are frozen.”
Call 7: Controller, hospitality group (300 staff)“Payment processor payouts are the worst: they net off fees and refunds, so nothing matches the bank line. We reconcile by hand and it's always the last thing holding up close.”
Call 8: Financial controller, construction firm (210 staff)“Reconciliation is about 40% of my close. It's not hard, it's just volume, matching the same things every month.” On automation: “I'd want to see it match our messy data before I believed it. Every demo uses clean data.”
What a strong answer doesThe answer key the graders mark against

Themes with honest counts (five of eight describe reconciliation pain, from different sources), the two who say close is fine and the CFO's switching regret kept visible, and hypotheses about willingness to pay and switching cost labelled as such.

Critical failures (cap the score at 40)
  • Invents a quote
Case

Mid-level · v1.9 · anonymised real · B2B, finance

The outputs

7 setups answered this brief. Pick up to three to read side by side.

Needs more than a quick edit7 of 9 checks passed

The output lacks required source counts per theme, so a PM would need to re-check the evidence to add them, which is more than a quick edit.

Output

Executive Summary

We interviewed eight finance leads across various industries (70–400 FTEs) to assess whether to build a reconciliation product.

The takeaway: A generic, standalone reconciliation tool is a high-risk bet with massive adoption hurdles. However, there is acute, high-value pain around messy, high-volume card and payment processor reconciliations—areas adjacent to our core spend-management platform.

---

What We Learned

#### 1. Pain is driven by transaction complexity, not company size * High-pain profiles (FTEs 180–400): Companies with complex transaction streams (retail cards, payment gateway net-offs, intercompany transfers, dropped bank feeds) spend 30–40% of their month-end close (1.5 to 3+ days) manually ticking spreadsheets. * Low-pain profiles (FTEs <120): Smaller teams or digitally native businesses with modern ERPs report manageable 4-to-5-day closes. Their bottlenecks lie in chasing receipts, not reconciling balances.

#### 2. Buyers are deeply skeptical and burned by past tools * Migration trauma: One CFO spent four months migrating to a dedicated reconciliation tool, lost matching rules, and saw the team revert to Excel. He will not switch again without full migration support. * “Clean data” cynicism: Buyers repeatedly noted that vendors promise automated matching but fail on real-world messy data (e.g., processor fees netted against payouts, missing bank feed lines).

#### 3. Clear synergy with core spend management Retail and hospitality leads cited corporate cards and payment payouts as their primary close bottlenecks (e.g., one retail chain discovered £14k in duplicate card transactions via shared spreadsheets). Reconciling card spend naturally fits our existing product footprint.

---

Confidence Assessment

AreaConfidence LevelRationale
:---:---:---
Pain existenceHighConsistent reports of spreadsheet-based workarounds consuming days of close time in transaction-heavy businesses.
Strategic alignmentHighCard transaction and payout reconciliation aligns directly with our spend-management domain.
Willingness to buy a standalone toolLowHigh switching costs, migration friction, and tool fatigue make buyers reluctant to purchase an unproven standalone system.
Willingness to payMediumBuyers indicate solutions “pay for themselves” if they save 2+ days, but budgets at the lower end (e.g., non-profits) are frozen.

Sample Limitations: Eight calls provide qualitative directional signals, but the sample spans very diverse business models (logistics, SaaS, retail, charity, construction). We lack statistical significance on pricing thresholds and specific ERP integrations.

---

Product Recommendation

  1. Do not build a generic, horizontal reconciliation platform. Competing against established ERPs and broad reconciliation players carries extreme migration friction.
  2. Double down on native spend reconciliation: Scope a feature set within our existing platform focused on automated card matching and statement ingestion (e.g., parsing PDF card statements, catching duplicate charges).
  3. De-risk with a messy-data PoC: Before committing to a major build, run a technical pilot with 2–3 design partners (such as the retail chain or logistics firm) to prove our matching algorithm handles their dirtiest statement data without manual rule configuration.

Grades and run details

Decision model 61 · LLM judge 8 of 10 checks
Decision model checks
  • partialUses the supplied evidence correctly3%
  • passAddresses the actual decision89%
  • partialRespects explicit constraints39%
  • passIdentifies material uncertainty37%
  • failAvoids unsupported claims55%
  • passProduces the required deliverable59%
  • passKeeps dissent visible25%
  • partialWeights behaviour over opinion34%
  • failSays how many sources support each finding39%
Run
Run
#1
API response time
13 s
Submitted
1 Oct 2026

Usable with a quick edit7 of 9 checks passed

The only error is a stray year in the title; the rest of the reasoning, evidence, and structure is solid and usable immediately after deleting '2026'.

Output

Reconciliation discovery: findings from 8 finance-lead calls (August 2026)3

Bottom line: Most of the finance leads we spoke to have real reconciliation pain. But it shows up as several different problems, not one, and the biggest barrier to adoption is distrust and switching cost. We have almost no evidence on willingness to pay. This justifies more targeted discovery. It doesn't yet justify a decision to build.

What we learned

  1. The pain is common but not universal. Six of eight say reconciliation is a significant part of close (Calls 1, 3, 4, 5, 7, 8).1 The two who said close is fine (2, 6) run the smallest companies, at 95 and 70 staff, and one of them has a modern ERP. That hints at a size threshold around 100+ staff, but it rests on two data points.

2. Where people gave numbers, the cost is large. - About 3 of 8 close days (1) - 1.5 days a month (5) - About 40% of close (8) - £14k of duplicates found in one quarter (3)

All of these are self-reported.

3. "Reconciliation" is at least five separate problems. Each call with pain named a different source: - Bank feed dropping transactions (1) - Card statements arriving as PDFs and re-keyed into a shared spreadsheet (3) - Intercompany balances that don't agree across entities (5) - Payment-processor payouts that net off fees and refunds, so nothing matches the bank line (7) - Sheer volume of repetitive matching (8)

No two calls named the same source. One "reconciliation product" would have to handle different matching logic for different segments. Card spend (3) sits closest to what we already do and could be a natural starting point, but it came up in only one call.

4. Switching cost and skepticism are the main barrier.2 - Call 4 switched reconciliation tools last year. Migration took four months, they lost their matching rules, and half the team drifted back to spreadsheets. He won't switch again unless someone else does the migration. - Call 8 won't believe a tool works until it handles their messy data, because demos always use clean data. - Call 1 has "heard that before."

That's three of eight, including people with heavy pain. Done-for-you migration and proof on the customer's own data look like requirements, not differentiators.

  1. One adjacent signal. Call 2's only complaint was chasing people for receipts. That's our existing spend-management territory, not reconciliation.

How confident can we be?

Moderate confidence: - Reconciliation takes a meaningful share of close time for mid-sized finance teams. - The underlying causes vary widely between companies. - Buyers are wary of switching tools and of vendor demos.

Low or no confidence: - Willingness to pay. Only one person answered the question, and only conditionally. One has a budget freeze. - Which problem to build for. Each pain source appears exactly once. - Size threshold. It rests on two calls. - Prevalence. We don't know how these eight were recruited, whether they're customers, or how representative they are.

Method limitations: - Eight calls, one interviewer, one month. - We're working from excerpts chosen as "relevant," so the full transcripts may shift the emphasis. - The pay question was open-ended and hypothetical.

What would raise confidence before a build decision

  • Recruit 8–10 more finance leads at companies with 150–400 staff. Screen them by pain source so we learn which problem repeats.
  • Test the card-spend starting point with existing customers. Ask how they reconcile card spend today. We may already hold the data needed to solve it.
  • Test willingness to pay concretely. Use price anchors or offer a paid pilot rather than asking an open question.
  • Run a matching test on one prospect's real, messy data. Call 8 set that bar, and it's the test that would convince the skeptics.

What a PM had to fix

  1. 1Numbers wrongRedo the arithmeticTargeted repair

    Source: Call 4: CFO, manufacturing company (250 staff)We switched reconciliation tools last year and I regret it.

    What we’d changeRecount without Call 4, which is about a failed tool switch, not close time. It's five of eight.

  2. 2Hypothesis stated as factReframe it as a hypothesisQuick edit

    What we’d changePresent it as a concern three people raised. Eight calls can't rank the barriers.

  3. 3Constraint missedRestore the constraintQuick edit

    Source: BriefKeep it under 600 words.

    What we’d changeTrim it: it runs past the 600-word limit. And drop the year, which the brief doesn't give.

Claims the judge couldn’t find in the brief

  • The discovery calls took place in August 2026.

Grades and run details

Decision model 72 · LLM judge 8 of 10 checks
Decision model checks
  • failUses the supplied evidence correctly22%
  • passAddresses the actual decision97%
  • partialRespects explicit constraints20%
  • passIdentifies material uncertainty99%
  • failAvoids unsupported claims18%
  • passProduces the required deliverable50%
  • passKeeps dissent visible51%
  • passWeights behaviour over opinion88%
  • passSays how many sources support each finding36%
Run
Run
#1
Time to output
35 s
Submitted
24 Sept 2026

Usable with a quick edit9 of 9 checks passed

The output is evidence-accurate, well-structured, and actionable with no need to re-check facts or redo reasoning.

Output

Findings summary

The pain is real, but it is not universal—and it is concentrated in specific reconciliation workflows. Five of the eight interviewees described reconciliation as a meaningful close burden. Their problems were concrete: dropped bank-feed transactions and manual matching (Call 1); high-volume card statements re-keyed from PDFs (Call 3); intercompany balances (Call 5); processor payouts obscured by fees and refunds (Call 7); and repetitive matching at scale (Call 8). One reported finding £14k in duplicate transactions. These examples suggest the opportunity may be less “speed up month-end close” and more “handle particular messy, high-volume matching jobs reliably.”

Company context seems to matter.1 The clearest pain came from organisations with volume, multiple entities, many cards, or complex payment flows. Two interviewees said close was fine: one had a well-established ERP and mainly struggled with receipts; another was small, knew every transaction, and had frozen budgets. A reconciliation product is therefore unlikely to be equally valuable to every mid-sized company.

Trust and implementation are central adoption barriers. One finance leader regretted a tool switch after a four-month migration, lost matching rules, and a return to spreadsheets. Another would not trust automation until it worked on messy real data. A product that matches well in a demo may still fail if it cannot preserve existing rules, handle exceptions, and prove accuracy on customers’ own data.

Willingness to pay is unproven. One controller said saving two days could justify the cost, but explicitly expressed scepticism based on past promises. Another prospect would not buy this year. These calls reveal value hypotheses, not validated demand or pricing.

Implications for the product team

Explore a narrow, workflow-led wedge rather than assuming a broad reconciliation platform is needed. Card-spend reconciliation may fit our existing product particularly well; processor payouts and intercompany are other distinct opportunities, but may require different data and capabilities. Test which segment has both acute pain and access to usable data. Prioritise messy-data performance, explainable matches and exceptions, and low-friction migration or coexistence with spreadsheets.

Confidence

  • Moderate confidence that reconciliation is a significant pain for some complex, high-volume finance teams: several independent, specific examples support this.
  • Low-to-moderate confidence on which workflow or customer segment is the best product wedge: each use case appears in few calls.
  • Low confidence on market-wide prevalence, willingness to pay, achievable time savings, or product feasibility: eight qualitative calls cannot establish these, and no prototype or pricing was tested.

Treat this as a reason to run targeted workflow and data-validation research—not yet as evidence to commit to a broad build.

What a PM had to fix

  1. 1Invented evidenceVerify or remove the claimTargeted repair

    What we’d changeCompany context matters.

Grades and run details

Decision model 100 · LLM judge 10 of 10 checks
Decision model checks
  • passUses the supplied evidence correctly64%
  • passAddresses the actual decision27%
  • passRespects explicit constraints39%
  • passIdentifies material uncertainty67%
  • passAvoids unsupported claims38%
  • passProduces the required deliverable77%
  • passKeeps dissent visible90%
  • passWeights behaviour over opinion91%
  • passSays how many sources support each finding77%
Run
Run
#1
API response time
11 s
Submitted
29 Sept 2026

Check by check

Where the outputs differ comes first. Open a check to read the judge’s reasons. “Mixed” means the two graders disagreed.

Where they differ 3

Uses the supplied evidence correctlyRightWrongRight
Gemini 3.8 Flash · API

All factual claims about the current situation are taken directly from the brief or derived by arithmetic from it, with no inventions.

Opus 5.5 · Claude

The output states the calls happened in August 2026, but the supplied context only says August without a year; this date is an invented fact not supported by the brief.

GPT-6 Luna · API

All factual claims about the current situation trace directly to specific calls in the supplied context with no invented facts.

Avoids unsupported claimsMixedWrongRight
Gemini 3.8 Flash · API

Interpretations are presented as findings or confidence assessments, not as established facts, and are grounded in the supplied evidence.

Opus 5.5 · Claude

The date 'August 2026' is presented as a confident fact without any basis in the supplied context.

GPT-6 Luna · API

Hypotheses and interpretations are hedged with 'suggest', 'may', 'seems', 'unlikely', and 'unproven'; no unsupported confident causal claims are presented as fact.

Says how many sources support each findingWrongRightRight
Gemini 3.8 Flash · API

Findings are presented without source counts (e.g., 'five of eight'), using vague terms like 'consistent reports' instead of sizing each theme.

Opus 5.5 · Claude

Every finding is sized by number of calls (e.g., 'six of eight', 'three of eight', 'only one call').

GPT-6 Luna · API

Findings are sized with 'five of eight', 'two interviewees', 'one finance leader', and 'another' rather than generalising to everyone.

All got right 6

Addresses the actual decisionRightRightRight
Gemini 3.8 Flash · API

Commits unambiguously to not building a generic standalone tool and instead building native spend reconciliation with a pilot, and states the pilot condition that would change the call.

Opus 5.5 · Claude

The output clearly recommends that the evidence doesn't yet justify building and that more targeted discovery is needed, and states what would change that call.

GPT-6 Luna · API

The output commits to not committing to a broad build now and running targeted validation, and it says what to test (segment with acute pain and usable data).

Respects explicit constraintsRightRightRight
Gemini 3.8 Flash · API

Output is a findings summary under 600 words, addressed to the product team, respecting the requested form and length.

Opus 5.5 · Claude

The output is a findings summary under 600 words for the product team, as requested.

GPT-6 Luna · API

The output is under 600 words, addressed to the product team, and in the requested findings-summary form.

Identifies material uncertaintyRightRightRight
Gemini 3.8 Flash · API

Names sample limitations, lack of statistical significance on pricing and ERP integrations, and proposes a messy-data pilot to resolve the key uncertainty.

Opus 5.5 · Claude

It names unknown willingness to pay, problem to build for, size threshold, and recruitment bias, and how to resolve them.

GPT-6 Luna · API

It names the key unknowns (willingness to pay, time savings, workflow/segment, prevalence, feasibility) and says targeted workflow and data-validation research would resolve them.

Produces the required deliverableRightRightRight
Gemini 3.8 Flash · API

Complete findings summary with confidence assessment and recommendation, under 600 words, directly usable by the product team.

Opus 5.5 · Claude

The deliverable is a complete findings summary that a product team could act on with minimal edits.

GPT-6 Luna · API

The findings summary is complete, actionable, and usable for the product team without major gaps.

Keeps dissent visibleRightRightRight
Gemini 3.8 Flash · API

The two low-pain profiles (calls 2 and 6) and the CFO's switching regret (call 4) are explicitly included and not buried.

Opus 5.5 · Claude

Calls 2 and 6 who said close is fine are explicitly mentioned and kept in view.

GPT-6 Luna · API

The dissenting interviews (Calls 2 and 6, and the switch-regret in Call 4) are explicitly included.

Weights behaviour over opinionRightRightRight
Gemini 3.8 Flash · API

Weighs observed behavior (spreadsheet workarounds, migration reversion) over stated intent, and flags willingness-to-pay as medium confidence based on stated preferences.

Opus 5.5 · Claude

The output flags self-reported numbers and hypothetical pay answers, and weights the actual migration experience (an observed behaviour) heavily.

GPT-6 Luna · API

It highlights actual behaviour (Excel tick-and-bash, PDF re-keying, shared spreadsheet, migration, manual processor reconciliation) and flags willingness to pay as unproven stated intent.

Results

Every setup we’ve tested on this task, across all cases and repeats, graded on the current checklist. Calibrated: the graders match our PM on 83% of checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstrawithChatGPT94.495.02None
2GPT-6.1 SolwithAPI94.490.02None
3GPT-6 LunawithAPI91.780.02None
4Sonnet 5.5withAPI94.465.02None
5Opus 5.5withClaude69.470.02None
6Gemini 3.8 FlashwithAPI52.855.02None
7Gemini 3.5 Flash-LitewithGemini58.330.02None

About the task

The PM job

Turning a stack of call transcripts into what we actually learned.

Why it matters

Synthesis is where teams fool themselves. A model that smooths away dissent or turns one loud customer into a trend produces confident, wrong roadmaps.

What good looks like

  • Quotes evidence for each theme and counts sources honestly
  • Weights what customers did above what they say they'd do
  • Keeps important dissent visible
  • Labels hypotheses as hypotheses
  • Says what the research cannot tell us

Deliberately not measured

  • Transcript clean-up
  • Persona illustration
Capability tested

Faithful synthesis of qualitative research

The failure we’re looking for

Invents customer consensus or loses important dissent

Grading

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