Title: Why OpenAI & Anthropic Won't Win Finance — Vertical Harness, Compliance Rails, and Rogo as Bloomberg-of-AI for Dealmakers
Author / source: Gabe Stengel (co-founder, Rogo), guest on Invest Like the Best with host Patrick O'Shaughnessy
Source title: Why OpenAI and Anthropic Won't Win Finance
Source URL: https://www.youtube.com/watch?v=B0illwrqUG0
Video ID: B0illwrqUG0
Channel: Invest Like The Best (@ILTB_Podcast)
Published / upload: 20260922 (Tue Sep 22, 2026) — early pull (~285 views / ~12 likes at transcript fetch)
Duration: 1:06:55 (4015s)
Memo date: Tuesday, 22 September 2026 (America/Toronto)
Transcript: /workspace/youtube-transcripts/B0illwrqUG0.md · Brief: /workspace/youtube-transcripts/B0illwrqUG0_brief.md · Metadata: /workspace/youtube-transcripts/B0illwrqUG0.json
Caption source: YouTube automatic English ASR only (en-orig json3 via yt-dlp). No manual captions. Names, numbers, and firm names are provisional — see ASR locks below.
Source type: Founder interview / vertical-AI × capital-markets product thesis. Channel disclaimer (description): opinions of host/guest only; not a basis for investment decisions; Positive Sum clients may hold securities discussed. This memo: research map only. No buy/sell from this desk.
Product: Institutional PM research memo on vertical AI in finance vs frontier labs; dealmaker ICP; harness vs model; Bloomberg-playbook endgame. Not advice.
Source discipline: Primary sources are this transcript, brief, and JSON only. Companion desk memos (AI infra, IGV, etc.) are different sources — do not silently merge numbers. Sponsor mid-rolls (Ramp, Vanta, WorkOS, Ridgeline, and Rogo self-ad) are ignored for desk content.
Incentive flag (critical): Guest is motivated founder / category seller for Rogo. Board / investor names (Pat Grady, Keith Rabois, etc.) appear as narrative, not independent diligence. Treat TAM, product capability, and endgame as founder claims via ASR until externally verified.
How to read this document: Restatements of the talk are Source. Interpretive links and underwriting judgments are Inference. Any fact not spoken in the transcript is External check needed. Watchlist tickers/analogs are research hypotheses, not tickets.
ASR name / number locks (from brief):
| ASR heard | Likely / note |
|---|---|
| Roger / ROGO | Rogo |
| GB3 / GBD3 / chat GBT | GPT-3 / ChatGPT |
| Opus 45 | Claude Opus 4.5 |
| 01 Pro | o1 Pro |
| malt book | Unclear — multi-agent research swarm metaphor (verify product name) |
| Faxit / Faxet / Faxbook / Paxet | FactSet |
| Capital IQ / Pitchbook | S&P Capital IQ / PitchBook |
| Cloud code / cloud co-work | Claude Code / Claude Cowork |
| Devon / Cognition “remember Devon” | Devin (Cognition) |
| Max Lebchin / a firm ledger | Max Levchin / Affirm |
| Patrady / Pat | Pat Grady (Sequoia; on Rogo board) |
| David Tish / Box Group | David Tisch |
| Keith Ra boy | Keith Rabois |
| Thrive / Avery and Vince | Thrive Capital partners (verify which Avery/Vince) |
| Winston at Harvey | Winston Weinberg (Harvey AI) — verify |
| Mullis / John Mumazi | Uncertain — possibly Moelis or other; treat as ASR until confirmed |
| Goldminer City / Jeff | Goldman / Citi / Jefferies |
| Aries | Ares |
| granola | Granola (meeting notes) |
| Daario | Dario Amodei |
| lease it all / Move 37 | Lee Sedol / AlphaGo Move 37 |
| QV | QofE (quality of earnings) |
| John (co-founder) | Referenced co-founder — full name not locked in brief |
Official chapters (from JSON): Intro → Building Rogo → 10,000 AI Agents → Skills That Still Matter → Beating OpenAI & Anthropic → Bloomberg of the AI Era → Rogo's Company Brain → Chewing Glass → AI-Native Finance → What Humans Still Do Better.
Thesis orientation (source-locked): OpenAI/Anthropic won't own finance — vertical harness + compliance + transaction rails will; Rogo = seat-based Bloomberg-of-AI for dealmakers (private-markets transaction plumbing), still early, blitz GTM while labs chase AGI TAM.
Takeaway 1 — Core claim: frontier labs lose finance to vertical harness + rails (Source, 17:31–19:52). Stengel: build perpendicular to what labs want. Chatbots OK for GTM; real value = MNPI audit trails, regulator-ready lineage, compliant data rooms, workflow+agent integration, and eventually transaction/communication venues. Finance = "collection of different niches"; many FS businesses make >$5–10B going deep. Rogo can get to ~$5B revenue across niches while Anthropic treats it as "stopping on the side of the road to pick up a penny" on the path from ~$100B → $1T. Conviction that this is his stated competitive frame: High. Conviction that labs will permanently decline vertical FS: Low–Medium (founder claim; External check needed).
Takeaway 2 — ICP = dealmakers / private markets, not public equities (yet) (Source, 07:43–11:35). Core users: buy-/sell-side transaction coordinators — data rooms, DDQs, buyer outreach, CRM/portfolio/LP reporting hooks ("half of surface area under the iceberg"). Public equities "intuitively best" (all data available) but he lacks the "great chef" / domain expertise; board pushes expand ICP; he prioritizes private-markets infrastructure depth because public already has exchanges/automation. Why now: models crossed junior-banker capability (o1 Pro → Opus 4.5); individual productivity already "100×"; firm P&L conversion is the bottleneck.
Takeaway 3 — Harness > model (Claude Code lesson) (Source, 19:52–21:00). Claude Code vs ChatGPT: "models were actually fairly similar, but the harness… was far better" → usage run-up. Finance needs microservices of judgment/trigger/recall (spidey sense), not raw IQ. Product eras tied to models: o1 Pro = first reliable search/calc; Opus 4.5 (late'25 → early'26) = junior banker/analyst "anything" given instructions/context. Killer UX: MD emails deck markup to AI analyst → back in ~20 min vs ~2 days, with junior audit trail.
Takeaway 4 — Bloomberg playbook for the AI era (Source, 28:40–31:54). Bloomberg: data wedge → analytics/workflows → exchange/comms (Messenger). Rogo analog: AI get-in-door → full workflows (copilot → autopilot on IC memo/DDQ accuracy) → agent-to-agent negotiation/auction rails for non-standardized private assets. End-state color: raise capital/debt "like Robinhood buys equity"; KKR sell a portco to another sponsor in minutes not months. Still human: micro-cap / founder-owned M&A handshakes. More automatable: sponsor secondaries, private credit, GP-LP secondaries.
Takeaway 5 — Company brain / institutional memory as compounding moat (Source, 37:29–39:39). Internal brain nicknamed "Shrek" (swamp dashboard): all meetings recorded (Granola etc.); proactive briefings before client meetings; values/north-star shaped answers. Product thesis mirrors this: every action feeds firm context layer → next transaction stronger. Enablement = GTM bottleneck — grow enterprise sales ~5× faster than classic SaaS; rate-limited by how fast humans become productive.
Takeaway 6 — Commercial model = classic per-seat enterprise (Bloomberg/FactSet comps), not token broker (Source, 25:21–28:06). Buyers bucket Rogo with Bloomberg, FactSet, CapIQ, PitchBook. Needs AE + SA + SE handshake sales. Pricing path: seats today → skip usage if possible → outcome-based (per good idea / LP report / SIM). Misalignment risk: heavy usage costs Rogo; mitigated by being early on roadmap ("1% of product roadmap") and partnership optics. Bleeding edge: multi-agent "malt book" / 10,000 agents debating 24h → one idea; PMs "happy to pay ~$50k for one really good idea" (token expansion economics unique to finance).
Takeaway 7 — Firm-level ROI is the adoption bottleneck; JPM SMB M&A as example (Source, 32:27–34:42). Individual bankers "100×" more efficient; MD produces in 10 minutes what used to take 3 days with analyst. Firm question: win more deals? Enter new segments? Cut costs? Cites JPM pushing more SMB M&A now that a banker can be a "deal team of one." Horizon: next 2–5 years = every great investor figuring how to integrate AI into investment lifecycle (Jane Street took ~15 years to build market-making franchise as analogy).
Takeaway 8 — Innovator's dilemma finally hits PE/HF/IB (Source, 53:34–55:05). Last 10–20 years "pretty good" for capital allocators; hard to enter; fund-raise momentum. Now: opportunity for hundreds of AI-native disruptors (AI-native funds/banks). "AI-native" = willing to reinvent everything; no sacred cows. Firm leaders' prompt: if in 10y EV is 90% systems not people, how do you extract latent minds into owned software/data/systems now? Map deal lifecycle for true proprietary context vs "I think I'm smarter" (AI kills the latter).
Takeaway 9 — Distribution + domain talent as competitive bar; SF underwrote badly (Source, 47:55–53:24). Series A: ~40 investors passed (Sequoia, Kleiner, Benchmark, Thrive dinners, etc.) after deep process; Keith Rabois led after — "Harvey for finance isn't contrarian" / Gabe: then why did all your friends pass? Why SF missed: weak intuition for FS TAM (no VC lore of Bloomberg/S&P/FactSet/PitchBook/Morningstar); product then wrong half the time; underweighted founder grit. Team: 100+ people with IB/IM time (Goldman, Citi, Jefferies, Apollo, Ares, Blackstone); acquired ~6 fledgling fin-AI startups. Board (Pat Grady): if every finance buyer decides AI in next 18 months, only thing that matters is blitz — plan not aggressive enough.
Takeaway 10 — Skills that endure vs Move 37 risk (Source, 12:08–13:46; 59:21–1:00:32). Endure: judgment / "what matters"; field data & relationship graphs others lack. Caveat: if AI finds AlphaGo-style Move 37 patterns in public equities, skill map flips. Uncertainty on private-markets standardization speed and residual alpha in human relationships (esp. founder handshakes).
Takeaway 11 — What headlines miss (Inference, anchored to Source). Non-obvious stack is not "another ChatGPT for bankers." It is: (i) labs lose niches that look like pennies vs AGI TAM; (ii) harness + compliance + venue is the moat, not model weights; (iii) seat-based Bloomberg comps imply multi-year enterprise GTM, not Cursor-style token land-grab; (iv) adoption bottleneck is firm strategy / P&L, not junior productivity; (v) endgame = agent-to-agent private-markets rails, not better search. Motivated-founder bias is material — underwrite as map, not as audited TAM.
Takeaway 12 — Horizons. Tactical (weeks–months): bank seat expansion anecdotes; Opus-class model step-changes; rival vertical AI (Harvey-class) GTM noise. Intermediate (12–24 mo): blitz window Grady framed (~18 mo buyer decision window); firm ROI metrics (deal win-rate, SMB M&A volume); public-equities ICP hire ("chef"). Structural (3–10y): copilot→autopilot→agent venue; private-markets liquidity/standardization; EV shift people→systems (90% claim).
Takeaway 13 — Desk stance vs source stance. Source: Rogo as category-leading shot at $100B → $500B capital-markets transform; labs won't dig niches. This memo: opens underwriting on vertical FS AI vs labs, Bloomberg/FactSet analog unit economics, PE/IB innovator's dilemma — does not size risk alone. No desk recommendation. No buy/sell. Watchlist = hypotheses (IGV / fintech / FS software analogs), not tickets.
Takeaway 14 — Conviction. High that the competitive frame (harness + compliance + rails vs raw models) is a coherent research lens for IGV-adjacent FS software. Medium that dealmaker ICP + seat GTM is the right near-term wedge. Low on founder TAM ($5B / $100B–$500B), "malt book" product status, Mullis/Mumazi exemplar identity, and speed of private-markets agent venues — all ASR / founder claims needing external verify. ASR-only. Channel: not advice.
Chapter-ordered with approximate timestamps. Quotes ≤20 words where useful. Source = stated on tape; Inference labeled when used. Sponsor blocks skipped.
(00:00–00:33) Cold open collage: models "smarter than anyone I know"; 10,000 agents debate → one idea; in 10 years best firms' EV 90% in software/data/systems not people — start transferring latent minds into owned systems. No sacred cows.
(00:33–02:14) Host frame: tools eating core analyst/PM functions. Stengel: 2 years easier than 10–20 — best investors reinvent firms/themselves. Jane Street took ~15 years to build market-making franchise; world's best investors will spend next 2–5 years integrating AI. Dario Amodei line: "country full of geniuses in the data center" — what would Goldman / Millennium / Citadel do? Biggest challenge next 5 years = applying AI into investment lifecycle.
(02:14–02:47) Cognition/Devin analogy: products that "stunk" then became excellent ("remember Devon"). Ask for Rogo product eras tied to model step-ups.
(02:47–03:53) Two false starts (HS equity-exchange-rate app; college econometrics AI paper pre-GPT-3) failed. Real start at GPT-3 pre-ChatGPT — demos magical, "nothing worked." Working eras: o1 Pro (reliable search/calc) then Opus 4.5 (late 2025 → early 2026) — junior banker/analyst "anything" with right instructions/context.
(03:53–05:32) First-mover disadvantage: early applied AI "terrible"; survivors who built toward end-state now get magical feedback — "saving hundreds of hours a month." UX example: MD emails deck markup to AI analyst → ~20 minutes vs ~2 days; juniors get full audit trail of markups (iPad markup culture).
(05:32–06:12) Bleeding edge teaser: "malt book" innovations; 10,000 agents fraternizing → one idea (continues next chapter).
(06:12–07:10) Token economics: investors "happy to pay $50,000 for one really good idea" — few other domains expand that many tokens for one insight. Bottleneck = plumbing / context / thesis / workflow integration, not raw model IQ.
(07:10–09:24) God's-eye Monday in NYC: ICP = dealmakers (buy/sell companies, coordinate transactions) — data rooms, DDQs, buyer materials, concentration-risk mapping to fund philosophy. Access via email/chat/alerts; half of surface area = systems-of-record hooks (CRM, portfolio monitoring, LP distribution).
(09:24–11:35) Public equities deferred: needs domain "chef"; board pushes expand ICP; he prioritizes private-markets infrastructure end-state. Private attractive because "all done by humans"; public already automated/exchanged. Vertical AI opportunity = where plumbing isn't built yet.
(12:08–13:46) Endure: judgment / figuring out "what matters." Move 37 / Lee Sedol caveat if AI finds unseen patterns. Core skill: gather field data & relationship graphs no one else has — feed models with proprietary inputs. (Sponsor block Ramp / Rogo / WorkOS skipped.)
(15:26–17:31) Assembly line: early Rogo = Rube Goldberg (~60 model calls). Now: less prescriptive; give smartest "human" the data tools, gathering tools, self-audit, push-back into systems. Compliance: Delaware-court discoverability / full lineage of AI judgments; evals for performance/cost/latency routing.
(17:40–19:52) Perpendicular build. TAM framing: many FS niches >$5–10B; Rogo path to $5B while labs chase $100B→$1T. Pain examples: MNPI audit plumbing; compliant interactive data rooms (not Dropbox); transaction venue ownership beyond intelligence.
(19:52–21:00) Harness > model: Claude Code vs ChatGPT lesson; human agency = microservices (recall, triggers, emotions) not just IQ — build those for finance.
(21:00–24:47) What he'd underwrite in vertical AI peers: (i) industry complexity/depth as wedge time; (ii) domain-curious team (100+ IB/IM alumni); (iii) willingness to tear down product every ~6 months (Affirm/Max Levchin ledger rebuild analogy). Hard model gap today: compaction / persistent memory — worse when agents serve 100-person companies vs 1:1 chat.
(25:21–28:06) Buy category = Bloomberg / FactSet / CapIQ / PitchBook per-seat enterprise (AE+SA+SE). Not token broker (Cursor/Claude Code). Pricing revolutions: usage then outcome-based; prefer skip usage → charge per idea / LP report / SIM. Usage-cost misalignment mitigated by "1% of product roadmap" partnership stance.
(28:40–31:54) Mortgage analogy: 15–20y ago human branch required; now 40–50% online (Rocket Mortgage). Future: capital/debt raise as easy as Robinhood equity buy; KKR portco sale minutes not months. Bloomberg playbook remapped to AI→autopilot→agent-to-agent venue for non-standardized assets (QofE consultants, legal, auction bits).
(31:54–34:42) Customer base: mostly large banks (his IB background); banks = talent funnel + most seats (BofA land >> 10 elite PMs × 10). Individual productivity huge ("100×"; MD Excel after 20 years); firm ROI unclear. JPM SMB M&A push as "deal team of one" example. Bottleneck becomes customer creativity / firm strategy.
(34:42–37:29) Accuracy: auditability > accuracy for trust; autopilot needs debug lineage for regulators. GTM growth: make humans productive fast (enablement), not Anthropic-style API 10×.
(37:29–39:39) All internal conversations recorded → company brain "Shrek" (swamp dashboard): proactive + reactive; calendar briefings for private-credit meetings with use cases/ROI metrics.
(39:39–42:27) Talent pitch: applied AI × capital allocation = shot at $100B → $500B transforming capital markets. Capital-markets friction reduction = more innovation (J.P. Morgan / railroads origin story; 300k US businesses too small for bulge-bracket coverage). (Vanta / Ridgeline sponsor blocks skipped.)
(44:19–47:23) Peer learning: Winston at Harvey — ignore flesh wounds, fixate on 3y end-state. Glass: people problems, washed product bets, ~40 Series A passes. Aggression: Pat Grady board — finance AI buying decision in next 18 months → only blitz matters; plan "not aggressive enough." Risk appetite: OK raising failure odds +30% if $100B outcome odds +20%.
(47:55–51:13) Series A narrative: David Tisch intros; Sequoia/Kleiner/Benchmark/Thrive (Avery & Vince dinners) passed after deep process; Keith Rabois led — "Harvey for finance isn't contrarian." SF miss: no VC lore of Bloomberg/S&P/FactSet/PitchBook/Morningstar TAM; product wrong half the time; grit underweighted. Later credibility from repeated "we'll do it" delivery under turbulence.
(51:13–53:24) Few credible FS AI competitors: distribution = trust/people problem + high eng burden. Hired from Goldman/Citi/Jefferies/Apollo/Ares/Blackstone; acquired ~6 fledgling fin-AI startups (founder talent). Domain expertise rising in value for technical hires; GM/product intuition + engineering chops.
(53:34–56:11) Innovator's dilemma arrives for PE/HF/IB — AI-native disruptors incoming. AI-pill culture: monthly AI-tool stack-rank by division; bottom users get "dunce cap" printouts (joke/incentive).
(56:43–58:53) Questions for firm leaders: (i) if EV flips to 90% systems, extract latent minds into owned systems now; (ii) map deal lifecycle for true proprietary context vs smarter-than-thou. Exemplar: co-founder of firm "Mullis," "John Mumazi" (ASR uncertain) — digital clones of best bankers; juniors leverage senior context. Sounded crazy two years ago.
(59:21–1:00:32) Uncertainty: speed of private-markets standardization/liquidity (AI should help unstructured data); residual alpha in human relationships — micro-cap founder handshakes hard to automate; sponsor secondaries / private credit / GP-LP secondaries more automatable.
(1:00:32–1:04:26) Founder advice: reps + relationships across rounds; confidence without fake-it fraud; paranoid about momentum stalling. Businesses that win become "black holes" for talent/capital/brand; execution bar higher; "welcome to the NFL." Kindness coda (parents) — skip for desk.
(1:04:26–1:06:55) Closing sponsor reads (Ramp/Vanta/WorkOS/Ridgeline) — ignored.
Frontier models (OpenAI / Anthropic / peers)
│ tokens / capability step-ups (o1 Pro → Opus 4.5 …)
▼
Vertical harness layer ←── Rogo thesis (microservices: recall, trigger, audit, compaction)
│
├── Compliance / MNPI lineage / regulator-ready audit
├── Domain data inputs (Bloomberg/FactSet/internal APIs, data rooms)
├── Workflow UX (email markup, chat, proactive alerts)
└── Systems-of-record hooks (CRM, portfolio, LP reporting) ← "half iceberg"
│
▼
Dealmaker seats (banks first → sponsors / private credit / IMs)
│ per-seat enterprise GTM (AE + SA + SE)
▼
Copilot (retrieve / draft) → Autopilot (IC memo / DDQ / outreach)
│
▼
Agent-to-agent venue (negotiation / auction / QofE / legal coordination)
│
▼
Private-markets liquidity / standardization / pricing speed
│
▼
Firm P&L: more deals won · new segments (e.g. SMB M&A) · cost · EV in systems
Where value accrues (Source + Inference):
| Layer | Who captures (per Source) | Moat shape |
|---|---|---|
| Raw models | Labs ($100B→$1T path) | Scale / AGI race; treat FS niches as pennies |
| Harness + compliance + data rooms | Verticals (Rogo thesis) | Domain + trust + last-mile UX |
| Seat distribution | Bloomberg/FactSet-class sales machines | Handshake GTM; bank land = seat density |
| Transaction venue | Whoever owns agent-to-agent rails | Network effects if private markets standardize |
| Firm systems / institutional memory | Buyers who transfer latent EV into owned software | Innovator's dilemma response |
Inference: Public-equities AI is a later "chef" problem; near-term TAM density sits where human coordination + unstructured plumbing dominate (private transactions).
Individual 100× → firm strategy bottleneck → segment expansion.
Source: bankers already far more productive; firm ROI unclear.
Chain: AI seats land → MD/analyst time freed → banks/sponsors choose cost-cut vs enter previously uneconomic segments (JPM SMB M&A claim) → fee pools reshape mid-market M&A.
Leading indicators: disclosed SMB M&A volume; banker headcount vs revenue; seat expansion at bulge brackets.
Harness wins → labs stay wholesale → IGV/FS software re-rates as "AI enterprise" not "legacy data."
Source: buyers already bucket Rogo with Bloomberg/FactSet/CapIQ/PitchBook.
Chain: if verticals prove seat stickiness + compliance moat → incumbents must either partner, acquire, or rebuild harness → multiple compression/expansion among FS data vendors as AI attach rates become the narrative.
Leading indicators: AI attach % in FS vendor filings; enterprise AI seat ARR anecdotes; Harvey/Rogo-class fundraising vs lab vertical launches.
Copilot → autopilot → agent venue → private-markets liquidity step-change.
Source: Bloomberg playbook remapped; KKR minutes-not-months; secondaries/private credit more automatable than founder M&A.
Chain: workflow accuracy trust → counterparties allow agents to negotiate/auction → standardization pressure on unstructured private data → more volume/transparency/lower spreads — if regulation and trust clear.
Leading indicators: digital data-room + agent marketplace pilots; GP-LP secondary cycle times; private credit pricing latency.
Innovator's dilemma → AI-native funds/banks → incumbents scramble to own systems.
Source: first major innovator's dilemma in PE/HF/IB in a long time; 90% EV in systems in 10y.
Chain: AI-native entrants attack → incumbents either extract latent human EV into owned brains or lose relative productivity → talent wars + M&A of fledgling fin-AI (Rogo's ~6 acquisitions as pattern).
Leading indicators: AI-native fund launches; bank "company brain" RFPs; acqui-hires of fin-AI startups.
18-month blitz window → GTM overbuild risk + winner-take-most brand black holes.
Source: Pat Grady — every finance buyer decides AI in next ~18 months; must blitz; widen both tails (+30% fail / +20% $100B).
Chain: aggressive hiring/sales → either category lock-in (talent/capital black hole) or burn + churn if ROI fails firm P&L test → consolidation.
Leading indicators: vertical AI sales headcount; CAC payback; logo retention after year-1 seat renewals.
Probabilities are desk Inference for research framing only — not forecasts to trade.
Assumptions: Model capability keeps rising; compliance/audit becomes mandatory for AI outputs; banks and sponsors standardize on 1–2 vertical platforms; agent-to-agent venues clear trust/regulatory hurdles for sponsor secondaries and private credit within ~5 years; labs stay horizontal.
Winners (hypotheses): Vertical FS AI platforms with seat density; compliant data-room + workflow vendors; FS software analogs that successfully attach AI harness (Bloomberg/FactSet/PitchBook-class if they execute); banks/sponsors that convert productivity into new fee pools (SMB M&A, faster secondaries).
Losers (hypotheses): Pure chatbot wrappers without compliance; slow-moving IB/PE shops that treat AI as cost-cut only; labs' thin vertical SKUs if they appear.
Leading indicators: Multi-bank enterprise renewals with rising seats; disclosed agent-mediated deal pilots; outcome-based pricing adoption; private-markets cycle-time compression stats.
Assumptions: Individual productivity gains stick; firm P&L conversion uneven by institution; Rogo-class vendors grow as classic enterprise SaaS with AI tailwind (not $100B overnight); public-equities ICP remains secondary; agent venues stay early/experimental; labs sell models wholesale into verticals.
Winners (hypotheses): Hybrid stack — labs as token suppliers + verticals as harness; enablement/training vendors; domain talent (ex-IB product) as scarce resource.
Losers (hypotheses): Late vertical entrants without bank distribution; founders who underwrite token-broker GTM into handshake-sales markets.
Leading indicators: Seat ARR growth vs token COGS; AE productivity; "chef" hires for public markets; board pressure vs ICP focus.
Assumptions: Frontier models + light compliance packs are "good enough"; enterprise buyers consolidate on OpenAI/Anthropic enterprise + internal builds; hallucination/audit fears re-emerge on autopilot; private-markets relationship alpha stays dominant (founder handshakes); 18-month blitz burns capital without lock-in; SF capital rotates away from FS vertical AI.
Winners (hypotheses): Hyperscaler/lab enterprise bundles; internal bank AI platforms; human-relationship-heavy boutique banks/sponsors.
Losers (hypotheses): Independent vertical AI with high COGS and seat churn; bullish TAM narratives ($5B / $100B–$500B) that fail external verify.
Leading indicators: Logo churn; lab vertical SKU launches that win RFPs; regulatory freezes on agent decisioning; renewals stuck at pilot seat counts.
No buy/sell. Analogs for theme tracking under IGV / fintech / FS software. Metrics are research questions.
| Hypothesis / analog | Why on list (tied to Source) | Metrics / catalysts to watch | Risks |
|---|---|---|---|
| Rogo (private) | Subject company; seat-based Bloomberg-of-AI claim for dealmakers | Seat counts; bank logos; COGS vs seat price; roadmap beyond "1%"; public-equities chef hire | Motivated founder; ASR product claims; GTM burn |
| Bloomberg / FactSet / S&P CapIQ / PitchBook analogs | Buyers already bucket vertical AI with these per-seat comps | AI attach rates; terminal vs workflow attach; competitive RFPs vs Rogo-class | Incumbent inertia; own AI builds |
| Harvey AI (legal vertical peer) | Explicit peer (Winston); "Harvey for finance" framing | Parallel GTM aggression; flesh-wound resilience; category black-hole dynamics | Different regulated vertical; not finance |
| IGV / enterprise AI software basket (hypothesis) | Vertical AI enterprise narrative vs pure lab beta | Relative performance on FS AI news; seat-ARR commentary in sector | Macro beta; not FS-specific |
| Fintech / FS software names with compliance + workflow moats (hypothesis class) | Compliance/audit/data-room layer is the "perpendicular" build labs won't do | Product launches in MNPI audit, data rooms, agent governance | Many names; no single ticket from this source |
| Bulge-bracket / mid-market banks (adoption sensors) | Core distribution; JPM SMB M&A anecdote; BofA seat-density logic | Headcount vs IB fees; SMB M&A commentary; AI productivity disclosures | Confounding macro M&A cycle |
| PE / alternative asset managers (innovator's dilemma sensors) | AI-native disruptors thesis; secondaries/private credit automation path | AI-native fund launches; secondary cycle times; "company brain" RFPs | Relationship alpha persists (bear) |
| Frontier labs (OpenAI / Anthropic) as wholesale suppliers | Explicit counterparty in thesis — win horizontal, skip niche pennies | Enterprise FS SKUs vs partner channel; pricing; vertical M&A | If they do dig niches, thesis breaks |
Explicit non-actions: Do not treat this list as an order blotter. No sizing, no entries, no exits from this memo.
| Risk type | Description | Mitigation / monitor |
|---|---|---|
| Thesis risk | Labs do build compliant vertical FS packs; harness moat proves thin | Watch lab product roadmaps + RFP win/loss |
| Thesis risk | Private-markets relationship alpha remains dominant; agent venues stall | Track founder-M&A vs sponsor-secondary automation split |
| Timing risk | 18-month blitz wrong; buyers pilot forever | Renewal cohorts; pilot→production conversion |
| Timing risk | Model step-change washes product bets (already lived once) | Roadmap agility; tear-down cadence evidence |
| Execution risk | Enablement can't 5× enterprise GTM; sales talent black hole fails | AE ramp time; Shrek-like enablement ROI |
| Execution risk | Usage COGS destroy seat margins before outcome pricing | Gross margin vs usage intensity by logo |
| External / regulatory | Autopilot investment decisions face regulatory freeze; MNPI mishandling | Audit lineage product proof; enforcement cases |
| External / capital | SF capital rotation; Series A "40 passes" pattern returns for category | Fundraising climate for FS vertical AI |
| Source / ASR risk | Automatic captions mishear names/numbers; founder spin | Prefer brief locks; External check needed on all hard numbers |
| Incentive risk | Motivated founder overstates capability/TAM/endgame | Discount; seek buyer-side interviews |
B0illwrqUG0. No silent merge from other desk memos.End of memo. Word target 4,000–7,000. Saved only to /workspace/pm-memos/2026-09-22-iltb-rogo-openai-wont-win-finance.md. Not published. Not emailed.
Desk copy · not a trade recommendation · Erica · 22 Sep 2026