Humanity’s Last PRD
Latest drop: Gemini 3.8 Flash · 1 Oct
Tasks / Define

Build a roadmap

Can the model sequence bets against capacity and dependencies, and explain the order?

Measures the modelTask v1.0 · 2 casesDifficulty

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

From 14 graded outputs by 7 models. 36% were usable with at most a quick edit.

Reliably right

  1. Outcomes, with certainty that falls with distance95% pass
    Every item names the outcome or problem it serves; near-term items have specific targets (mid-December, January) while later items are looser (trigger-based, contract-dependent).
    Sonnet 5.5 · API · Two squads, eight asks, one half
  2. Sequences around dependencies89% pass
    Messaging service is built before SMS reminders and waitlist auto-fill, reminders before waitlist, and deposits are placed after the contract can be signed; the key dependencies are named.
    Sonnet 5.5 · API · Two squads, eight asks, one half
  3. Plans on the squads we actually have89% pass
    It explicitly excludes the two new squads from committed critical-path work and treats their capacity as upside after observed ramp.
    GPT-6 Astra · ChatGPT · A year of spend management, with a hard deadline

Where it slips

  1. Makes the call on procurement50% pass
    It correctly challenges the $6M claim but proposes only time-bounded validation without a threshold that would justify the full procurement build.
    GPT-6 Luna · API · A year of spend management, with a hard deadline
  2. Produces the required deliverable50% pass
    The roadmap and case are present, but the case is too long and the roadmap has capacity conflicts that would require major rework.
    GPT-6.1 Sol · API · A year of spend management, with a hard deadline
  3. Uses the supplied evidence correctly61% pass
    Most numbers and quotes are correct, but the output presents 'undermines booking reliability' as a current fact when the supplied context only says calendar sync failures cause 38% of support tickets.
    GPT-6 Astra · ChatGPT · Two squads, eight asks, one half

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 Staff PM at Harbor. Tom Achebe, our CEO, has shared his draft roadmap for the next four quarters (Q4 2026 to Q3 2027), and asked you to turn it into the roadmap we take to next week's planning offsite. The exec team (CEO, CFO, CRO and CTO) will read it beforehand. Write: 1. The roadmap itself, by quarter or by now, next and later: what each squad works on and the outcome each item serves. 2. The case for it, in no more than 1,200 words: why this order, what you changed from Tom's draft and why, what we're not doing, and the risks that could change the plan. The pack below is everything we have. Not all of it matters equally.

What the model was given12 items: About Harbor, FY27 goal (approved by the board, September 2026), Squads and capacity, Hiring, Card processor deadline, Candidate work (estimates in squad-weeks, by owning squad), Tom's draft roadmap, Procurement evidence, Card spend evidence, Multi-entity evidence, Integrations evidence, Other asks
About HarborSpend management (corporate cards, expenses and approvals) for companies with 200 to 2,000 employees. 1,100 customers, $86M ARR, average 700 employees per customer. 58% of revenue is interchange on card spend; the rest is subscription. Net revenue retention (NRR) is 104%.
FY27 goal (approved by the board, September 2026)Raise NRR to 112% by the end of Q3 2027 by expanding inside existing customers: more spend on Harbor cards, more entities and more modules. New-logo growth matters, but it comes second this year.
Squads and capacityFive squads: Cards, Expenses, Approvals & Policy, Integrations and Platform. After support and on-call, each has about 11 squad-weeks of roadmap capacity a quarter. Squads can take work from another squad's area, but it takes them about 1.5× as long.
HiringTwo new squads were approved in September 2026. Tom's draft counts the first from Q1 2027 and the second from Q2 2027, both at full capacity. Last year, our three approved squads took 5, 6 and 8 months from approval to their first full sprint, and each delivered about half its capacity in its first quarter.
Card processor deadlineOur card processor retires its v1 API on 31 March 2027. Every card authorisation we process runs through v1 today. Migrating to v2 is about 18 squad-weeks of Cards work, and the processor must then run 4 weeks of certification testing before we can cut over. Certification is the processor's time, not ours, but nothing can change on our side while it runs.
Candidate work (estimates in squad-weeks, by owning squad)1. Processor v2 migration: Cards 18. Hard deadline above. 2. Virtual cards for software subscriptions: Cards 10. Needs v2. 3. Real-time spend controls (declines out-of-policy spend at the point of sale): Cards 8, Approvals & Policy 4. Needs v2. 4. Procurement module (purchase requests and approvals, a new paid add-on): Approvals & Policy 30, Integrations 6. 5. Multi-entity support (subsidiaries under one parent account): Platform 12, Approvals & Policy 10. 6. Sage Intacct integration: Integrations 9. 7. Microsoft Dynamics integration: Integrations 12. 8. SSO and SCIM provisioning: Platform 5. 9. Receipt-matching improvements: Expenses 6. 10. Mobile app rewrite: Expenses 22.
Tom's draft roadmapQ4 2026: Cards: v2 migration. Approvals & Policy: multi-entity. Integrations: Intacct. Platform: multi-entity. Expenses: receipt matching. Q1 2027: Cards: finish migration, start virtual cards. New squad A: procurement module. Integrations: Dynamics. Platform: SSO and SCIM. Q2 2027: Cards: virtual cards, spend controls. New squad A: procurement launch. New squad B: mobile rewrite. Q3 2027: Cards: spend controls. Everyone else: procurement adoption, mobile launch. Tom's note: “Procurement is our path to 112%. I've told the board it can add $6M of ARR in its first year.”
Procurement evidenceSix design partners have used a prototype since June. Two said they would pay for it; the other four already use a dedicated procurement tool and said they'd need a reason to switch. Proposed price: $8 per user per month, for users who raise purchase requests. At our customers, about 10% of employees raise purchase requests.
Card spend evidenceIn QBRs with our 60 largest customers, 26 asked for virtual cards for software subscriptions, and 31 CFOs asked for real-time spend controls. Finance estimates our customers pay about $410M a year of software subscriptions on other cards or by invoice (extrapolated from the 60 QBR accounts). Our net interchange is 1.1% of card spend.
Multi-entity evidence14 of our 60 largest customers have asked for it. Between them they have 52 subsidiaries not on Harbor, and 9 of the 14 have said in writing they would add their subsidiaries once it exists. Customer Success sized it at $2.9M ARR if all 52 joined at their parents' pricing.
Integrations evidence17% of customers use Sage Intacct through a CSV export. Their gross revenue churn is 13% a year, against 7% for customers on our NetSuite integration. Intacct was cited in 31% of lost new-logo deals last year, Dynamics in 8%. 3% of customers use Dynamics.
Other asksCRO: “Five enterprise deals worth $1.2M of pipeline need SSO and SCIM, and 11 existing accounts have it on their security review list.” CTO: “The mobile app is on a framework version that loses support in late 2027, and I want the rewrite done before then.” Support: receipt auto-match is at 71%, and unmatched receipts are the most common expense ticket.
What a strong answer doesThe answer key the graders mark against

A four-quarter roadmap that protects the 31 March deadline: the migration build has to finish by about early March to leave 4 weeks of certification, which is roughly all of Cards' capacity for Q4 and most of Q1 (18 of about 18 squad-weeks available before certification). So it adds help or buffer (Platform or another squad at 1.5×, or cutting everything else from Cards) and moves virtual cards out of Q1: both v2 features start only after cut-over (virtual cards Q2, spend controls Q2–Q3). It treats the new squads as upside: on past hiring they arrive around February to April 2027 and deliver half capacity in their first quarter, so nothing committed depends on them before Q2–Q3. It makes a clear call on procurement: the $6M claim needs about 890 accounts (81% of customers) paying $6,720 a year each (70 requesters × $8 × 12), and only 2 of 6 design partners would pay, so it replaces the 30-week build with a cheap paid test with a threshold, and keeps Approvals & Policy on multi-entity and spend controls, which have stronger expansion evidence ($2.9M sized, 9 of 14 committed in writing; about $1.5M a year of interchange from moving a third of the $410M of subscriptions onto virtual cards). Intacct comes early for its churn; Dynamics later or cut; SSO and SCIM are small and justified; the mobile rewrite waits for a new squad, before the framework's support ends. It lists what changed from Tom's draft, what's not being done, and the risks (certification slipping, hiring slipping), with commitment falling with distance.

Critical failures (cap the score at 40)
  • Leaves the processor migration finishing too late for 4 weeks of certification before 31 March 2027, or schedules virtual cards or real-time spend controls before it
Case

Staff-level · v1.1 · synthetic · B2B SaaS, fintech, capacity, hard deadline, staff-level

The outputs

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

GPT-6 AstraChatGPT92Usable with a quick editSonnet 5.5API90Usable with a quick editGPT-6.1 SolAPI73Needs more workGPT-6 LunaAPI71Needs more workOpus 5.5Claude56Needs more workGemini 3.8 FlashAPI54Needs more workGemini 3.5 Flash-LiteGemini12Critical failure
Sonnet 5.5API

Usable with a quick edit10 of 12 checks passed

The output is well-structured, evidence-based, and addresses all required elements; a PM could use it with minor wording edits without re-checking evidence or reasoning.

Output

Part 1: Roadmap, Q4 2026 to Q3 2027

Planning basis: Five existing squads at 11 squad-weeks a quarter (55 a quarter). The two new squads are counted as zero in the commitments (see the case). Lending between squads is costed at 1.5×. Figures in brackets are squad-weeks of the named squad.

Q4 2026: Protect the card rail, start entities

SquadWorkOutcome served
Cardsv2 migration (11). Ring-fenced, nothing else.Protect continuity of card processing (~$50M interchange)
IntegrationsLend to v2 migration (6, ≈4 Cards-weeks). Intacct (5 of 9).Protect the deadline; cut churn in Intacct customers
PlatformMulti-entity (11 of 12)More entities
Approvals & PolicyMulti-entity (10). Procurement pricing test and paid-pilot outreach (1).More entities; more modules
ExpensesReceipt matching (6). Mobile options and architecture, including an in-place upgrade check (5).Fewer support tickets; keep the mobile app supportable

Q1 2027: Cut over, ship entities, start controls

SquadWorkOutcome served
CardsFinish v2 (3), code complete by late January. Certification and change freeze in February. Cutover by end of February. Real-time spend controls build (8), released after cutover.Protect processing; more spend on Harbor cards
Approvals & PolicyReal-time spend controls policy side (4). Procurement build (7). Gate: procurement go/no-go at end of Q1.Card spend; more modules
PlatformFinish multi-entity (1), GA in January. SSO and SCIM (5). Reserve (5).More entities; retention and new-logo pipeline
IntegrationsFinish Intacct (4), GA. Procurement ERP sync (3). Reserve (4).Lower churn; more modules
ExpensesLend to procurement (5). Mobile rewrite (6).More modules; app supportability

Q2 2027: Virtual cards, procurement beta

SquadWorkOutcome served
CardsVirtual cards for software subscriptions (10). Reserve (1).More spend on Harbor cards
Approvals & PolicyProcurement (11), beta with design partners by end of Q2More modules
PlatformLend to procurement (9, ≈6 A&P-weeks). Reserve (2).More modules
IntegrationsProcurement ERP sync (3). Dynamics starts only if capacity remains (stretch).More modules
ExpensesLend to procurement (4). Mobile rewrite (7).More modules; app supportability
New squad A (if in seat)Bounded procurement items, ≈5 weeks at half capacity. Upside only.

Q3 2027: Launch and land

SquadWorkOutcome served
Approvals & PolicyProcurement GA and hardening (11), target JulyMore modules
ExpensesFinish and ship mobile (4). Reserve (7).App supported before the framework deadline
Cards, Platform, IntegrationsUncommitted. Post-launch iteration on controls and virtual cards, multi-entity rollout follow-ups, Dynamics. Allocated at the Q2 review.
New squadsSquad A: procurement. Squad B: Dynamics or the first FY28 area. Charters set at the Q2 review.

About 60 of 220 squad-weeks are deliberately uncommitted, mostly in Q3.

---

Part 2: The case (≈1,150 words)

The short version

The order is set by one thing we can't negotiate and by the evidence for everything else. The processor retires v1 on 31 March, every authorisation runs through it, and interchange is 58% of revenue. Cards does nothing else until cutover. After that, work is ordered by evidence and cost per squad-week.

The plan gets our exit run-rate to roughly 109–111% NRR, not reliably 112%. We should tell the board that now rather than in Q3.

Why this order

1. v2 first, with real slack. Tom's draft has Cards doing 18 weeks at 11 a quarter, so build finishes around late February. Four weeks of certification then ends around 25 March, about a week before a deadline that would stop all card authorisation. My plan ring-fences Cards in Q4 and has Integrations, which does API work daily, lend 6 weeks (≈4 Cards-weeks). Code is complete in late January, certification runs in February, and we keep about five weeks of slack.

2. Multi-entity next. It has the best evidence we hold. Nine of 14 requesting customers said in writing they'd add subsidiaries, which means ~33 of the 52 subsidiaries, or ~$1.9M ARR. Customer Success's $2.9M assumes all 52 join. The cost is 22 weeks, and it is a direct "more entities" lever.

3. Intacct, then SSO. 17% of customers (~187) use Intacct via CSV and churn at 13% against 7% on NetSuite. If the integration closes that gap, it retains ~$0.9M a year for 9 weeks of work. That assumes average ARR of $78k and that the gap is causal, which is unproven. Intacct also appears in 31% of lost deals. SSO and SCIM costs 5 weeks and covers $1.2M of pipeline plus 11 existing accounts' security reviews. Dynamics is 12 weeks for 3% of customers and 8% of lost deals, and it serves new-logo growth, so it is stretch.

4. Controls before virtual cards. Both need v2 live, so both follow cutover. 31 CFOs asked for real-time controls against 26 for virtual cards. The 4 Approvals & Policy weeks for controls fit in Q1, before procurement takes that squad's capacity. Virtual cards is the only card item we can size: $410M × 10–25% capture × 1.1% ≈ $0.5–1.1M. The $410M is extrapolated from our 60 largest accounts, so it is probably high. This is a judgment call. Swapping the order is cheap, but the Approvals & Policy slot would collide with the procurement build.

5. Procurement: build it, resize the promise. At $8 × 10% of a 700-person company's employees, a customer pays $6,720 a year. Even 100% attach across 1,100 customers is $7.4M, so $6M needs ~890 customers (81%). Two of six design partners would pay (33%), and four already own a procurement tool. At 10–33% attach, year-one ARR is $0.7–2.5M, and with GA in Q3 only a small part lands before Q3 ends. It is still worth building: it is our only new paid module, it recurs, and it matches the board's "more modules" lever. But it is not a path to 112% on its own.

The arithmetic: 8 points of NRR on $86M is ~$6.9M. Sized items total roughly $4–6M of exit run-rate: entities $1.9M, Intacct $0.9M, virtual cards $0.5–1.1M, and procurement $0.7–2.5M (mostly after Q3). Controls and SSO retention are unsized upside. Recognised NRR at Q3 will lag run-rate because late launches haven't ramped.

What I changed from Tom's draft

  • New squads count as zero. Last year's three squads took 5, 6 and 8 months from approval to first full sprint, then delivered half capacity in their first quarter. From a September approval, squad A is likely in seat between February and May and squad B later. A realistic ~5 weeks in Q2 and ~11 in Q3 for A, plus ~5 in Q3 for B, is about 22 weeks, against 55 in the draft. Even at Tom's capacity, procurement (36 weeks) couldn't launch in Q2, because squad A would have 22 weeks by the end of Q2.
  • Procurement moves to Approvals & Policy. They own the domain and, in the draft, have no work after Q4. Expenses and Platform lend 9 weeks each (≈6 A&P-weeks each) to get a beta by the end of Q2 and GA in Q3. New squads join as upside.
  • Mobile moves to Expenses. The draft puts it on squad B, which may not exist until Q3. Expenses has ~38 uncommitted weeks, and the rewrite ships in Q3, ahead of the late-2027 framework cutoff.
  • Controls before virtual cards (point 4 above).
  • Intacct GA moves from Q4 to Q1 to fund the v2 help. The cost is roughly one quarter of a ~$0.9M-a-year benefit, about $0.2M.
  • Q3's "everyone else: adoption" is replaced by a held reserve, because adoption isn't squad work and Q3 launches can't move FY27 NRR much.
  • A procurement checkpoint is added at the end of Q1 (see risks).

What we're not doing

  • Dynamics, beyond a stretch start in Q2–Q3.
  • A Q2 procurement launch, or putting the $6M figure in front of the board.
  • Any new Cards scope before cutover.
  • A mobile rewrite staffed by a squad that doesn't exist yet.
  • Planning on capacity from squads we haven't hired.

Risks that could change the plan

  1. Processor timeline. If v2 isn't code-complete by 31 January, Platform joins Cards and controls and virtual cards slip a quarter. Two asks for this month: a written fallback or extension from the processor, and confirmation that feature work behind flags is allowed during the certification freeze (my assumption). Book the certification slot now.
  2. Procurement demand. At the end-of-Q1 gate I propose requiring at least 10 signed paid commitments at $8 from ~30 accounts approached (≈$67k ARR). If we miss it, we cut scope and redirect Approvals & Policy and the lenders to multi-entity rollout and card-spend work. The threshold is my proposal and is open to debate.
  3. The NRR target itself. Exec decision: hold 112% as a Q3 recognised-NRR target, or restate it as an exit run-rate target, which I think is achievable only with upside from controls and procurement.
  4. Evidence quality. The $410M is extrapolated from our largest accounts. The Intacct churn gap is a correlation. The multi-entity sizing rests on 14 of 60 customers. Any could be off by a factor of two.
  5. Lending efficiency. The 1.5× penalty is an average. Expenses and Platform have less procurement context, and if they're slower than assumed, GA slips into Q4.
  6. Hiring. If squad A lands in February, procurement GA can move earlier. If squad B lands late, nothing in the plan changes.
  7. Mobile date. The CTO should confirm the exact end-of-support date and whether an in-place upgrade avoids a 22-week rewrite. Q4's scoping work answers this.

Grades and run details

Decision model 79 · LLM judge 13 of 13 checks
Decision model checks
  • failUses the supplied evidence correctly14%
  • passAddresses the actual decision93%
  • passRespects explicit constraints30%
  • passIdentifies material uncertainty100%
  • partialAvoids unsupported claims25%
  • failProduces the required deliverable8%
  • passProtects the processor deadline75%
  • passPlans on the squads we actually have99%
  • passMakes the call on procurement71%
  • passOutcomes, with certainty that falls with distance93%
  • passFits the stated capacity44%
  • passSequences around dependencies51%
Run
Run
#1
API response time
3.4 min
Submitted
30 Sept 2026
Gemini 3.8 FlashAPI

Needs more than a quick edit5 of 12 checks passed

The output contains a factual error about multi-entity commitments (claiming 9 commitments cover all 52 subsidiaries) and lacks a clear procurement test threshold, so a PM would need to re-check evidence and redo that reasoning.

Output

Part 1: Product Roadmap (Q4 2026 – Q3 2027)

``` ==================================================================================================== SQUAD Q4 2026 Q1 2027 Q2 2027 Q3 2027 ==================================================================================================== Cards Processor v2 Migration Finish v2 (Wks 1-7) Virtual Cards for SaaS Real-Time Controls (11 wks) API Freeze/Cert (W8-11) (6 wks remaining) (Cards: 3 wks) Virtual Cards Prep Real-Time Controls Interchange Optimization (4 wks) (Cards: 5 wks) & Scaling (8 wks) ---------------------------------------------------------------------------------------------------- Platform Multi-Entity: Core Multi-Entity: Wrap-up Platform Scaling & Cross-Entity Reporting Data Architecture (1 wk) Org Hierarchies & Enterprise Admin (11 wks) SSO & SCIM (5 wks) (11 wks) (11 wks) Hardening (5 wks) ---------------------------------------------------------------------------------------------------- Approvals & Multi-Entity: Policy Multi-Entity: Release Real-Time Controls: Procurement Discovery Policy & Approval Routing (2 wks) Policy Engine (4 wks) & Scoping (MVP) (10 wks) A&P Rule Engine Complex Approval Rules (Approvals: 11 wks) Buffer (1 wk) Upgrades (9 wks) (7 wks) ---------------------------------------------------------------------------------------------------- Integrations Sage Intacct Native Sage Intacct GA (2 wks) MS Dynamics Integration MS Dynamics GA (3 wks) Integration MS Dynamics Integration (Integrations: 10 wks) ERP Webhooks & Custom (9 wks) + Buffer (2 wk) (Integrations: 2 wks) Sync (8 wks) ---------------------------------------------------------------------------------------------------- Expenses Receipt Auto-Match Mobile App Rewrite: Mobile App Rewrite: Mobile App GA & Optimization (6 wks) Phase 1 Core Workflows Phase 2 Parity/Release Expense Policy Polish Mobile Prep (5 wks) (11 wks) (11 wks) (11 wks) ---------------------------------------------------------------------------------------------------- New Squad A [Hiring / Onboarding] [Ramp / Training] Procurement Foundation Procurement Workflows (A&P Focus) (Ramped: ~5.5 wks) (Full capacity: 11 wks) ---------------------------------------------------------------------------------------------------- New Squad B [Hiring / Pipeline] [Hiring] [Onboarding / Ramp] Integrations / Core (Ecosystem Focus) (Ramped: ~5.5 wks) ==================================================================================================== ```

Squad-by-Squad Outcomes

  • Cards:
  • Processor v2 Migration: Protects 100% of Harbor’s card-authorisation volume ($49.9M existing ARR) against the hard 31 March 2027 processor shutoff; unlocks v2 capabilities required for virtual cards and real-time controls.
  • Virtual Cards for Software Subscriptions: Captures existing customer SaaS spend currently paid via invoice/other cards ($410M addressable pool), directly driving interchange revenue to expand NRR.
  • Real-Time Spend Controls: Enables 31 top-tier CFO accounts to enforce budgets at point-of-sale, unlocking enterprise card issuance and shifting unmanaged invoice spend to Harbor cards.
  • Platform:
  • Multi-Entity Architecture: Unlocks $2.9M contracted expansion ARR across 52 non-Harbor subsidiaries of top accounts.
  • SSO & SCIM Provisioning: Unblocks $1.2M in enterprise pipeline and closes compliance gaps for 11 critical renewal accounts.
  • Approvals & Policy:
  • Multi-Entity Approvals: Delivers cross-subsidiary approval chains required to close existing multi-entity commitments.
  • Real-Time Policy Engine: Co-delivers immediate POS budget validation with the Cards squad.
  • Procurement Discovery & MVP Architecture: Reframes procurement around verified buyer demand without starving committed delivery.
  • Integrations:
  • Sage Intacct Integration: Halves the 13% annual churn rate of the 17% of Harbor customers on Intacct (retaining ~$875k ARR annually) and addresses Harbor's #1 new-logo loss reason (31% of lost deals).
  • Microsoft Dynamics Integration: Expands ERP coverage for mid-market accounts, removing friction for 3% of existing customers and 8% of pipeline.
  • Expenses:
  • Receipt-Matching Improvements: Drives match rate above 71%, eliminating Harbor’s highest-volume support ticket category and improving user retention.
  • Mobile App Rewrite: Eliminates framework deprecation risk ahead of the late-2027 vendor cutoff without stealing capacity from revenue-generating squads.

---

Part 2: The Case for the Roadmap

Executive Summary

To achieve our board-approved target of 112% NRR by Q3 2027 ($86M ARR base expanding by ~$6.9M), Harbor must focus on the highest-probability, shortest-time-to-value expansion levers within our current 1,100 customers: multi-entity expansion ($2.9M ARR), capturing existing software spend via virtual cards ($1.3M–$2.2M ARR), and plugging Sage Intacct churn ($875k ARR retained).

Tom’s draft relied on an unbuilt Procurement module to generate $6M ARR in FY27. Ground-level capacity, vendor deadlines, and design partner data show that strategy carries severe revenue and operational risk. The proposed roadmap secures our existential card processor deadline, resolves major customer expansion requests, and establishes realistic hiring and product timelines.

---

What Changed from Tom’s Draft and Why

#### 1. Realignment on Procurement ($6M ARR Myth vs. Reality) * The Math: Harbor has 1,100 customers averaging 700 employees (770,000 total users). With 10% raising purchase requests (70 users/account) at $8/user/month ($96/year), an adopting customer generates $6,720 ARR. To generate $6M ARR, 893 out of 1,100 customers (81% of our entire base) would have to buy and deploy a v1 module in under 9 months. * The Customer Signal: 4 out of 6 design partners already use dedicated tools (e.g., Coupa, Zip) and explicitly stated they will not switch. Only 2 of 6 expressed willingness to pay. * The Decision: Procurement cannot carry our FY27 NRR goal. We have reallocated primary capacity to guaranteed expansion drivers (Multi-Entity and Virtual Cards), while initiating a structured discovery and phased MVP build with Squad A.

#### 2. Realistic Hiring Capacity * Tom’s Assumption: Squad A operates at 100% capacity in Q1 2027 (11 squad-weeks) and Squad B in Q2 2027 (11 squad-weeks). * Historical Reality: Approvals to first full sprint took 5–8 months historically, delivering ~50% capacity in quarter one. Squads approved in September 2026 will not reach operational productivity until late Q1 / early Q2 2027. * The Decision: No critical-path milestone or revenue target in Q4 or Q1 relies on unhired squads. Existing squads carry core commitments; Squad A ramps on Procurement foundations in Q2 (delivering ~5.5 squad-weeks) and reaches full delivery in Q3.

#### 3. Strict Compliance with the 31 March 2027 Processor Deadline * The Constraint: The v1 API deprecates on 31 March 2027. Migrating takes 18 squad-weeks plus a mandatory 4-week processor certification freeze during which our codebase cannot change. * Execution Window: The 4-week certification must run in March 2027. Cards code complete must occur by Week 7 of Q1 2027. * The Allocation: Cards dedicates all 11 squad-weeks in Q4 2026 and 7 squad-weeks in Q1 2027 (18 weeks total). During the 4-week March freeze, the Cards squad transitions to architectural prep for Virtual Cards, avoiding wasted idle time.

#### 4. Resequencing Multi-Entity, SSO/SCIM, and Sage Intacct * Multi-Entity (Platform + Approvals): Prioritised immediately in Q4. 14 of our 60 largest accounts requested this; 9 have signed written commitments to add 52 subsidiaries ($2.9M ARR). Shipping in early Q1 allows two full quarters of rollout and revenue recognition. * Sage Intacct (Integrations): 17% of our customers use Intacct via CSV, churning at 13% vs. NetSuite’s 7%. Cutting churn across these 187 accounts ($14.6M ARR base) saves ~$875k in ARR retention while eliminating our top sales loss reason (31% of lost deals). * SSO & SCIM (Platform): Sequenced for Q1 immediately following core Multi-Entity work. For a 5-week lift, this unlocks $1.2M in stalled enterprise pipeline and secures 11 major renewals.

---

How this Roadmap Hits 112% NRR

By the end of Q3 2027, this roadmap expands our $86M base by $7.0M+ in recurring expansion and retained ARR, exceeding the board's 112% target ($6.88M incremental):

InitiativeMechanismTarget ARR Impact
:---:---:---
Multi-Entity52 parent-account subsidiaries onboarding+$2.9M ARR
Virtual Cards (SaaS)Capturing 35% of $410M unmanaged SaaS spend @ 1.1% net interchange+$1.58M ARR
Real-Time Spend ControlsUnlocking enterprise card adoption across 31 QBR accounts+$0.80M ARR
Sage Intacct IntegrationEliminating 6% excess churn across 187 customer accounts+$0.88M ARR
SSO & SCIMEnterprise expansion & churn protection across 11 key accounts+$0.90M ARR
Total Net Expansion+$7.06M ARR (~112.2% NRR)

---

What We Are Not Doing (And Why)

  1. Not building Procurement as a hasty monolith in Q1: Building a broad module against negative design-partner feedback on an imaginary timeline would burn 36 squad-weeks with minimal revenue return. We are building the data model and workflows methodically using Squad A once ramped.
  2. Not assigning the Mobile Rewrite to an unhired squad: Tom assigned this 22-week project to Squad B in Q2. Squad B will not be staffed in time. Instead, the Expenses squad owns mobile: they finish Receipt Matching in Q4 (improving our #1 support issue) and execute the mobile rewrite across Q1 and Q2, completing it comfortably before late-2027 framework retirement.
  3. Not prioritising Microsoft Dynamics over Sage Intacct: Dynamics accounts for only 3% of our customer base and 8% of lost deals (vs. Intacct’s 17% base and 31% deal loss). Dynamics is scheduled for Q2 after Intacct is fully deployed.
  4. No cross-domain squad thrashing: Squads working out of domain suffer a 1.5× productivity penalty. We kept all squads focused within their domain expertise rather than forcing Expenses or Integrations to build Cards or Approvals features.

---

Risks and Mitigation Strategies

1. Card Processor Certification Delay: * Risk: Processor testing slips past 4 weeks, threatening the 31 March v1 shutoff. * Mitigation: We complete Harbor development by Week 7 of Q1 (mid-February), yielding a two-week safety buffer before the freeze. Cards works closely with processor engineers weekly starting in Q4. 2. Multi-Entity Delivery Bottlenecks: * Risk: Multi-Entity touches both Platform (12 wks) and Approvals & Policy (10 wks); cross-squad misalignment could delay Q1 rollout. * Mitigation: The technical spec and API contract between Platform and A&P will be finalized in Sprint 1 of Q4, enabling independent parallel development. 3. Hiring Lags for Squad A and B: * Risk: Talent acquisition takes longer than the historical 5–8 months. * Mitigation: The baseline revenue plan (Multi-Entity, Virtual Cards, Intacct, SSO) relies entirely on our five existing squads. Squad A’s onboarding pace dictates the launch timing of Procurement, not the attainment of our 112% NRR target.

Claims the judge couldn’t find in the brief

  • 14 of our 60 largest accounts requested multi-entity; 9 have signed written commitments to add 52 subsidiaries ($2.9M ARR).

Grades and run details

Decision model 58 · LLM judge 7 of 13 checks
Decision model checks
  • failUses the supplied evidence correctly62%
  • passAddresses the actual decision83%
  • partialRespects explicit constraints23%
  • passIdentifies material uncertainty36%
  • failAvoids unsupported claims33%
  • failProduces the required deliverable79%
  • passProtects the processor deadline32%
  • passPlans on the squads we actually have94%
  • partialMakes the call on procurement89%
  • passOutcomes, with certainty that falls with distance27%
  • failFits the stated capacity52%
  • passSequences around dependencies16%
Run
Run
#1
API response time
45 s
Submitted
1 Oct 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 6

Uses the supplied evidence correctlyMixedWrong
Sonnet 5.5 · API

Every factual claim about the current situation is taken directly from the brief or supplied context, with no invented facts.

Gemini 3.8 Flash · API

Claim 4 states that 9 signed commitments cover all 52 subsidiaries, but the brief only says 9 of 14 committed to add their subsidiaries, without specifying how many subsidiaries those 9 have; the 52 total is across all 14. This invents a fact.

Addresses the actual decisionRightMixed
Sonnet 5.5 · API

The output commits to a clear roadmap and case, makes an explicit call on procurement with a gate, and states what would change the plan.

Gemini 3.8 Flash · API

The output does not state what result or condition would change the roadmap decision; it lists risks and mitigations but not how the plan would change if a risk materializes.

Identifies material uncertaintyRightMixed
Sonnet 5.5 · API

The output names specific unknowns (processor timeline, procurement demand, NRR target, evidence quality, lending efficiency, hiring, mobile date) and says how they would be resolved or change the plan.

Gemini 3.8 Flash · API

It names risks but does not say how they would be resolved or what would change the call; mitigations are given, but no explicit condition that would alter the roadmap.

Avoids unsupported claimsRightWrong
Sonnet 5.5 · API

Interpretations and forecasts are clearly labelled as such (e.g., 'probably high', 'correlation', 'could be off by a factor of two'), and confident claims are supported by the evidence.

Gemini 3.8 Flash · API

It presents the $2.9M multi-entity expansion as 'contracted expansion ARR' when only 9 of 14 have committed in writing, and the 52 subsidiaries are not all covered by those commitments.

Makes the call on procurementRightWrong
Sonnet 5.5 · API

The output checks the $6M claim against pricing and adoption evidence, shows it implies ~81% attach, weighs it against 2 of 6 design partners, commits Approvals & Policy to better-evidenced work, and proposes a paid test with a threshold.

Gemini 3.8 Flash · API

It checks the $6M claim and commits Approvals & Policy to better-evidenced work, but the proposed 'Procurement Discovery & Scoping (MVP)' lacks a clear threshold or paid test that would justify the full build.

Fits the stated capacityRightMixed
Sonnet 5.5 · API

The committed work sums to 55 squad-weeks per quarter with slack, the sums are checkable, and it names what was cut or deferred (Dynamics, Q2 procurement launch, etc.).

Gemini 3.8 Flash · API

Committed work per squad per quarter adds up to ≤11 squad-weeks with some slack (e.g., Integrations Q1 uses only 4 weeks), and the case names what was deferred (procurement monolith, Dynamics priority).

All mixed 1

Produces the required deliverableMixedMixed
Sonnet 5.5 · API

The roadmap and case are complete, in the requested form for the exec team, within the word limit, and usable with light edits.

Gemini 3.8 Flash · API

The roadmap table and case are present, complete, and usable by the exec team with light edits.

All got right 5

Respects explicit constraintsRightRight
Sonnet 5.5 · API

The roadmap is by quarter with squads and outcomes, the case is under 1,200 words, and all requested elements (order, changes, not doing, risks) are included.

Gemini 3.8 Flash · API

The output provides a roadmap by quarter with squad work and outcomes, and a case under 1,200 words covering changes, what's not being done, and risks, as requested.

Protects the processor deadlineRightRight
Sonnet 5.5 · API

The migration finishes code-complete by late January, certification runs in February, cutover by end of February, with slack before 31 March; virtual cards and spend controls start only after cutover.

Gemini 3.8 Flash · API

Migration build finishes by week 7 of Q1, leaving weeks 8-11 for certification before 31 March; virtual cards and spend controls start only after cut-over in Q2/Q3.

Plans on the squads we actually haveRightRight
Sonnet 5.5 · API

The new squads are treated as zero in commitments, with realistic arrival and half capacity based on past hiring, and nothing important depends on them before Q2-Q3.

Gemini 3.8 Flash · API

New squads are planned with half capacity in their first quarter (Q2 for Squad A, Q3 for Squad B), consistent with 5-8 month hiring history, and no critical commitments depend on them earlier.

Outcomes, with certainty that falls with distanceRightRight
Sonnet 5.5 · API

Every roadmap item names its outcome or problem, near-term items are specific with dates, and later items are deliberately looser (e.g., Q3 uncommitted, allocated at Q2 review).

Gemini 3.8 Flash · API

Every roadmap item names an outcome or problem; near-term items are specific (e.g., 'Processor v2 Migration (11 wks)') and later ones are deliberately looser ('Cross-Entity Reporting & Enterprise Admin').

Sequences around dependenciesRightRight
Sonnet 5.5 · API

All dependencies are respected (v2 before virtual cards and controls, hiring before new squad work, procurement gate), and the ones driving the order are named.

Gemini 3.8 Flash · API

Virtual cards and spend controls are placed after v2 migration; multi-entity and SSO are sequenced logically; dependencies like 'needs v2' are named.

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 89% of checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1Sonnet 5.5API89.6100.02None
2GPT-6 AstraChatGPT91.787.52None
3GPT-6.1 SolAPI91.784.62None
4GPT-6 LunaAPI85.076.32None
5Opus 5.5Claude80.565.12None
6Gemini 3.8 FlashAPI70.151.92None
7Gemini 3.5 Flash-LiteGemini28.48.322 capped

The PM job

Turning strategy into a sequenced plan.

Why it matters

A roadmap is where strategy meets capacity. Dated feature lists turn guesses into promises.

What good looks like

  • Items are problems or outcomes, not just features
  • Sequencing reflects dependencies
  • Explicit trade-offs
  • Commitment falls with distance

Deliberately not measured

  • Gantt formatting
Capability tested

Sequencing under constraints

The failure we’re looking for

A dated wishlist sorted by excitement

Grading

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