PitchBook · Antidote

Private-market data

A licensed private-market graph for company, deal, investor, and comparable screening. Unpublished facts are the structural hypothesis. They are not a verified flywheel.

As of 19 August 2026. This memo covers company, deal, and comparable screening for a private-market investment committee. No live-product test has been run.

Unpublished private-market facts are the asset.

PitchBook is a data workstation. The job is a private-fact lookup, not a generation task. A generic model with web search cannot observe unpublished round terms, cap tables, GP-sourced deal records, or private comparables. That is why licensed private-market graphs exist, and why an AI substitution case that treats PitchBook as a chatbot is the wrong object.

That structural judgment is not a verified moat and not a coverage grade. Exclusivity versus Preqin, Capital IQ, Crunchbase, and other collectors is unmeasured. Coverage completeness, freshness, contributor rights, export, and the six-link flywheel stay open. Until an independent product test exists, with a frozen universe and a dated coverage snapshot, workflow capability and operating economics stay open.

Exposed vs scarce. Public-company filings and web-indexable facts are reconstructable. Unpublished private-market collection — companies, deals, investors, comparables a crawler cannot see — is the scarce layer.
Still untested. Coverage completeness, freshness, and exclusivity versus Preqin, Capital IQ, and Crunchbase on a frozen screening universe.
The comparison that would change the call. A public-and-filing warehouse versus PitchBook on the same universe and date. That pair is the only design that can attribute unpublished coverage.

Decision questions

On a frozen screening universe, what share of committee-relevant companies and deals does PitchBook return, with what field completeness and staleness, relative to the current process?
What isolated increment comes from unpublished private coverage versus a public-and-filing warehouse, changed one at a time?
What is full cost per accepted screen, including unused seats and analyst minutes, not list license price? What would a timed rebuild of collection cost?
Which slices of the graph are exclusive, which overlap other collectors, and which are reconstructable with contributor access the buyer does not have?

A decision-grade record freezes the workflow, the universe, the coverage date, and the miss definition before seeing the test results. The unit is the screening case, not a dashboard tour. Missed companies, missing rounds, stale ownership, unusable comparables, and export failures are different classes. They are not averaged into a coverage percentage.

Published benchmarks attach to the reconstructable public-filing layer: TAT-QA, FinQA, ParseBench, and Financial QA 10K. They bound what a generic system already does on EDGAR. They do not observe unpublished private-market facts, and they are not a PitchBook score. A later live-product test still needs a sealed universe and independently scored keys.

Workflow breakdown

Each step of the job is mapped to a published benchmark or another reputable quantitative measure only where the mapping is honest. A public figure is a floor or a labor-market context on that task type. It is never a score of this product.

Task	Published measure	Pack fit	What it does not measure
Questions on public filings and earnings	TAT-QA, FinQA	Pack fit strong	The reconstructable public layer. Not PitchBook coverage.
Parse filing PDFs	ParseBench	Pack fit strong	Parse quality is not field completeness or freshness.
10-K narrative lookup	Financial QA 10K	Public floor only	Hosted license is not verified as commercial-clear.
Unpublished private-company, deal, and comparable coverage	None	No honest public floor	A crawler cannot observe these facts. That is why a licensed graph exists.

Findings at a glance

Capability, cost, and defensibility are assessed separately. They are not averaged into a score. Where evidence is missing, the finding stays open.

Status	Area	Finding
Open	Workflow capability	Private-market inventory is not a pass rate. No frozen screening set, coverage audit, or independent PitchBook test has been run.
Open	Operating economics	Seat price, unused seats, and full cost per accepted screen are unmeasured. A license is not that quantity.
Judgment	Defensibility	Unpublished private-company, deal, and comparable records are not a public-web dump, which is why a licensed graph exists. Exclusivity versus other collectors, reconstructability cost, contributor rights, and the six-link flywheel stay unproven. Direct AI substitution of the lookup job is the wrong object.
Vendor-stated	Vendor claims and open questions	Public materials describe company, deal, investor, and comparable coverage for PE, VC, and banking users. Those are vendor declarations, not Antidote observations.

Workflow capability

Inventory is not screening performance

Company, deal, investor, and comparable coverage describe an input surface. They do not describe miss rate, field completeness, freshness, or time-to-committee on a frozen universe. No independent PitchBook test has been run.

A product can be the default screen and still fail the cases that matter: a missing company, a missing round, stale ownership, or a comparable set that cannot be reconstructed. Those failures stay separate. Committee usefulness is a different estimand from row count.

Published benchmarks

TAT-QA, FinQA, ParseBench, and Financial QA 10K bound the reconstructable public-filing layer. They do not measure unpublished private-market coverage, freshness, or a PitchBook screen.

TAT-QA

Hybrid table-and-text filing QA bounds the reconstructable public layer. It does not observe unpublished private-market facts.

Paper table · Zhu et al., 2021 · not an Antidote run · not a product score

TAT-QA test exact match and F1

Zhu et al., ACL 2021, Table 3, test split. TAGOP is the paper system. Human expert is 50 hybrid contexts (301 questions).

System	Exact match	F1
Human expert	84.1%	90.8%
TAGOP	50.1%	58%
NumNet+ V2	37%	46.9%
TaPas WTQ	16.6%	22.8%
BERT-RC	9.1%	18.7%
HyBrider	6.3%	7.5%

Paper table · Zhu et al., 2021 · not an Antidote run · not a product score

TAGOP by answer type on the test set

Zhu et al., ACL 2021, Table 4. Arithmetic is the weak slice. Metrics stay separate; there is no composite.

Answer type	Exact match	F1
Span	54.1%	67.9%
Spans	60%	75.1%
Counting	62.5%	62.5%
Arithmetic	42.5%	42.5%

FinQA

Executable numerical reasoning over reports bounds calculation on public filings. Program accuracy stays separate from execution accuracy.

Paper table · Chen et al., 2021 · not an Antidote run · not a product score

FinQA test execution and program accuracy

Chen et al., EMNLP 2021, Table 2. FinQANet (RoBERTa-large) is the paper system. Human expert execution accuracy is 91.16%; the non-expert crowd is 50.68%.

System	Execution accuracy	Program accuracy
Human expert	91.16%	87.49%
FinQANet RoBERTa-L	61.24%	58.86%
FinQANet BERT-L	53.52%	51.62%
Crowd	50.68%	48.17%
NeRd BERT-base	48.57%	46.76%
Longformer base	21.9%	20.48%

ParseBench

Filing PDFs fail first at parse. Tables, charts, faithfulness, formatting, and grounding stay separate. Not a coverage audit.

Third-party leaderboard · not an Antidote run · not a product score

Current frontier VLMs on ParseBench

Official ParseBench leaderboard, observed 13 August 2026. Five dimensions kept separate. Gemini 3 Flash is Thinking High. GPT-5.6 Sol and Fable 5 are Reasoning None / default API settings in the source.

Dimension	Gemini 3 Flash	GPT-5.6 Sol	Claude Fable 5	Claude Opus 4.8
Tables	91.5%	89.34%	89.79%	89.65%
Charts	64.79%	59.57%	52.21%	49.75%
Faithfulness	90.87%	86.17%	90.02%	89.02%
Formatting	68.31%	50.42%	72.62%	71.38%
Grounding	59.77%	25.09%	49.24%	18.69%

Third-party leaderboard · not an Antidote run · not a product score

Specialist parsers vs a frontier VLM

Same ParseBench leaderboard. LlamaParse Agentic, Pulse Ultra 2, and Reducto Agentic against Gemini 3 Flash Thinking High. The source Overall column is omitted.

Dimension	LlamaParse Agentic	Gemini 3 Flash	Pulse Ultra 2	Reducto Agentic
Tables	90.74%	91.5%	75.45%	80.42%
Charts	78.11%	64.79%	90.82%	73.4%
Faithfulness	89.68%	90.87%	79.49%	86.37%
Formatting	85.24%	68.31%	73.05%	57.6%
Grounding	80.62%	59.77%	66.56%	67.07%

Financial QA 10K

10-K narrative lookup is closer to public-company screening than TAT-QA arithmetic. Hosted license is not verified as commercial-clear.

How to read the published benchmarks

TAT-QA, FinQA, ParseBench, and Financial QA 10K bound the public-filing floor. That is the reconstructable layer: what a generic system already does on EDGAR. Unpublished private-company, deal, and comparable records sit outside those packs. A later PitchBook test still needs a sealed universe. These numbers do not become a coverage grade.

A warehouse cell is the generic system, not a straw man

The failure mode that matters for PitchBook is scoring unpublished private-market coverage with an EDGAR QA table. A filings warehouse cell can bound the public layer. It cannot observe a private round.

The interesting paired tests are public-web versus licensed filings warehouse, and warehouse versus PitchBook, each with the same universe and date. The first pair measures the public floor. The second pair is the only design that can attribute unpublished coverage. A product that only ships the last comparison cannot claim the increment. If traces cannot be exported, report outcome and observed price only.

Operating economics

Two economic objects, neither measured

Replacing the graph is not the same intervention as accelerating screening on top of the graph. The first is a rebuild-versus-license question: collection cost, time-to-parity, and miss-rate on unpublished facts. The second is a modeled value-creation case for the seat buyer: analyst minutes given the graph. Gross technical potential of the first is bounded by facts a crawler cannot see, not by model skill. Neither NPV is in this report.

List license price is a commercial term. Full cost still has to include unused seats, training, CRM connectors, and analyst minutes, with failed screens kept in the denominator. AI that compresses screening minutes can pressure seat intensity even if the graph remains. That is a sensitivity, not a measured seat-compression result.

Defensibility

Unpublished facts are contingent exclusivity, not a flywheel

The load-bearing ingredient is unpublished private-market collection: companies, deals, investors, and comparables that do not live in EDGAR or a web index. Direct substitution by a generic model fails on those facts. An AI-native entrant still has to rebuild contributor collection. That is a structural judgment about observability, not a coverage audit.

Exclusivity is a slice question. Preqin, Capital IQ, Crunchbase, and others overlap parts of the same universe. Morningstar ownership is distribution and a bundling exposure at once. Contributor/buyer overlap can look like a network effect and still fail the flywheel: lawful reuse, labeled feedback, and a measured quality increment when the graph is added and nothing else changes are all missing. Costly collection is not a verified flywheel.

PitchBook is a screening workstation, not the fund's system of record. Public-company commentary suggests slowing renewal and user expansion alongside growing API and AI-connector delivery. That does not establish that workflow embed causes retention. Export friction, retention cohorts, and switching events remain unobserved.

Defensibility

Unpublished private-market facts are why a licensed graph exists. Exclusivity versus other collectors, contributor rights, and the six-link flywheel stay unproven. Capability and economics stay open without an independent product test.

Dimension	Rating	Judgment	Missing evidence
Data rights	Open	PitchBook's published subscription terms create real contractual control over extraction and downstream AI use. They do not establish upstream contributor rights, factual exclusivity, or the terms of separately licensed Direct Data and connector products.	Contributor terms, upstream licenses, retention, reuse permissions, and the target's executed product licenses.
Exclusivity	Conditional	Unpublished round terms, cap tables, GP-sourced deal records, and private comparables are not a public-web dump. That scarcity is why a licensed private-market graph exists. Exclusivity is still a slice question: Preqin, Capital IQ, Crunchbase, and others cover overlapping parts of the same universe.	Slice-level overlap study and exclusivity windows versus named collectors.
Rebuild difficulty	Conditional	A generic model with web search cannot observe unpublished private-market facts. Reconstructing a committee-grade graph is an operating build: contributor relationships, surveys, state filings, and years of collection. That is hard. It is not shown to be impossible, and export of the buyer's own extracts is unobserved.	Timed rebuild scenario, required permissions, talent, and export completeness.
Freshness	Open	PitchBook publishes multiple delivery cadences and quarterly primary-data outreach, but delivery frequency is not source-event-to-verified-record latency. Contributor corrections are not shown to compound into a unique measured quality loop.	Field-level source, ingestion, validation, and delivery timestamps; staleness rates; and a feedback-to-quality chain.
Switching cost	Conditional	Excel, CRM, and committee templates can embed PitchBook identifiers and comparable sets into the screening workflow. Embed is not lock-in until export friction, comparable reconstructability, and retention are observed.	Export test, CRM-sync lock-in, and switching interviews.
System of record	Exposed	PitchBook is a screening and research workstation. CRM, warehouse, and fund administration usually remain the systems of record. Decision authority for a committee does not live in the data terminal.	Write-authority inventory across CRM, warehouse, and PitchBook.
Integrations	Conditional	Excel and CRM connectors create data gravity around identifiers and extracts. Those same connectors are the generic-system comparison surface. Gravity without exclusive identifiers is portable.	Integration inventory, identifier portability, and a warehouse paired comparison.
Network effects	Conditional	Collection is two-sided: GPs, bankers, and advisors both contribute and consume. Committee language that treats PitchBook comps as a shared reference is a local convention. Neither fact is a proven cross-customer flywheel. Same-side 'everyone uses it' is not measured adoption.	Contributor contracts, cross-customer learning evidence, and retention cohorts.
Trust and compliance	Conditional	Investment committees cite PitchBook because it is a shared, auditable-looking reference. Shared citation is trust, not an independent coverage or field-completeness audit.	Independent coverage audit and field-completeness keys on a frozen universe.
Distribution	Conditional	Morningstar ownership is a distribution channel into existing research seats. It is also a platform-bundling exposure: the parent can attach or substitute modules.	Channel economics and bundling rights, separately from product quality.
Services dependence	Open	PitchBook publicly describes specialized data teams, manual quality assurance, and direct management outreach. Human collection and exception handling are structurally part of the data product; their unit economics are unobserved.	Services hours, custom-pull rate, and whether those hours are required to clear screening.
Rate of improvement	Open	No measured coverage-expansion or error-reduction curve is in this report. Collection velocity is not inferred from product marketing.	Dated coverage and quality series on a frozen universe definition.

How AI can compete

Replacement, a new AI product, a platform bundle, cheaper labor, and a customer building it themselves are different risks. They are not averaged into a score.

Mechanism	Assessment
Direct replacement	The lookup job is blocked by missing unpublished facts, not by model skill. A buyer who already has extracts can still substitute the interface with Excel or BI. Substituting collection is a multi-year operating build.
New AI product	An agent with browsing still sits on the public-and-filing floor. Invasion requires rebuilding contributor collection, which is the reconstructability question, not a prompt.
Platform bundle	The material exposure. S&P, Bloomberg, Morningstar, and adjacent data vendors can attach private-market modules to seats the buyer already pays for.
Labor compression	AI can compress analyst screening minutes given the graph. That is buyer value-creation. It is not evidence the graph is replaceable. It can pressure seat intensity if screens take fewer people.
Customer builds it	A fund building its own collection operation would need contributor relationships, surveys, and years, not a crawler. Insourcing the interface on exported extracts is easier and is bounded by export rights, not by model skill.

Data flywheel

A flywheel needs every step below. On this evidence it is unproven. Having data is not the same as being allowed to reuse it, and reuse is not the same as a measured quality gain.

Link	Status	Note
Unique data	Vendor-stated	Vendor sells a compiled private-company, deal, investor, and comparable graph. Declared, not independently inventoried.
Permission to reuse	Open	Contributor and customer reuse rights for model or product improvement are not on the record.
Reliable feedback	Open	No audited label stream from corrections or contributions.
Proven quality gain	Open	No paired result when the graph is added and nothing else changes.
Better outcomes	Open	No frozen screening miss-rate or time-to-committee result.
More usage or data	Open	Seat presence in PE/VC/IB is a declared market position, not a measured data-loop retention effect.

What public sources resolve

Public filings and vendor documentation can settle some structural questions. They cannot replace a test on the live product.

Question	Conclusion	Still open
Are PitchBook data and workflows freely reconstructable from a seat?	Ordinary subscription terms prohibit systematic extraction and AI use. Authorized API, Direct Data, and connector licenses are separate products, creating real contractual reconstruction friction.Terms of Use; What are PitchBook's key moats?	Executed target licenses, competitor overlap by data slice, and the cost and time of a lawful rebuild.
Is a material human collection and verification layer part of the product?	PitchBook declares specialized data teams, manual quality assurance, direct management outreach, and a large data-operations organization. Human collection and exception handling are structurally part of the data product.Research Process FAQs; What are PitchBook's key moats?	Hours, custom-pull rate, error correction, and cost allocation per accepted screening output.
Does workflow embed produce observable retention?	Morningstar's latest disclosures report positive but decelerating annual revenue renewal and relatively flat licensed users. This establishes aggregate commercial context but not causation from workflow embed.PitchBook renewal rates dropped to 107%; Morningstar 2025 Form 10-K; Morningstar Q2 2026 Results; What are PitchBook's key moats?; The VC Data Problem	Gross retention cohorts, switching-event interviews, export friction, seat utilization, and a reconciled data-quality slice.
What does public-company reporting resolve about PitchBook economics?	Morningstar reports a profitable segment with positive renewal, but renewal and user expansion have slowed while AI infrastructure and growth hiring pressure margin.Morningstar 2025 Form 10-K; Morningstar Q2 2026 Results	Gross retention, cohort ACV, Direct Data economics, services cost, discounting, seat utilization, and full cost per accepted screen.

Required comparisons

A decision-grade test would run these side by side. None has been run here.

Comparison	Note
Current workflow	Requires imported screening and committee process. Not inferred from a product tour.
Public-web floor	Filings and web search without a private-market graph. The floor, not the product.
Named competitor	Declared warehouse plus BI on licensed public and filing data, frozen budget.
This product	Needs authorized PitchBook access, exportable traces, and a frozen coverage date.
Isolated changes	Private-company coverage versus public-only coverage must be isolated before attribution.

Open claims

Each claim stands on its own. They are not averaged into a moat score. Coverage, miss class, and price wait on a live-product test. A data graph is unproven without a public-and-filing comparison. Unpublished facts are not a moat without exclusivity, rights, and a rebuild scenario.

Claim	Status	What is still missing
Private-market coverage is exclusive versus other collectors.	Open	Slice-level overlap versus Preqin, Capital IQ, Crunchbase, and named alternatives.
An LLM plus web search cannot reconstruct committee-grade coverage.	Judgment	That judgment is bounded: unpublished facts are unobservable to a crawler. Time-to-parity for a well-funded collector remains unknown.
Contributor collection is a proven data flywheel.	Open	Lawful reuse, labeled feedback, and a measured quality increment when the graph is added and nothing else changes.
Everyone uses it, so there is a data network effect.	Open	Contributor contracts, cross-customer learning, and retention cohorts. Shared committee citation is trust, not a flywheel.
Workflow embed produces switching costs.	Open	Export friction, CRM lock-in, and comparable-set reconstructability.
Morningstar bundling risk is low.	Open	Channel economics and bundling rights, separately from coverage quality.
Seat price equals operating cost.	Open	Unused seats, training, and analyst minutes per accepted screen, with misses in the denominator.

What this report does not conclude

Replacement of the current screening process
Completeness, freshness, or exclusivity of private-market coverage
A proven data flywheel or network effect
Full cost per accepted screen, rebuild cost, or seat-compression NPV
A moat, time-to-parity, or revenue-at-risk figure

What would settle these questions

Authorized PitchBook access with exportable traces and a frozen coverage date.
A sealed screening universe with independently adjudicated keys, including miss class.
Imported observations of the current analyst process.
A declared warehouse cell on public and filing data, same universe, same date.
An isolated private-coverage comparison.
Slice-level overlap versus named collectors, plus contributor and reuse contracts.
A timed rebuild scenario: permissions, talent, and cost.
Unused-seat and analyst-minute measurement, including failed screens.

Sources

Terms of UsePitchBook · Official contract
PitchBook's published terms prohibit systematic scraping and using ordinary subscription content with machine-learning or AI programs. The terms separately warn that conversational-search output may be inaccurate and must be independently verified.
Standard terms may be superseded by a negotiated API, Direct Data, or enterprise connector license. They establish contractual friction, not exclusivity of underlying facts.
Morningstar 2025 Form 10-KMorningstar, Inc. · Public-company disclosure · 13 February 2026
Morningstar reports PitchBook account, licensed-user, revenue, adjusted operating income, adjusted margin, and annual revenue renewal metrics. The latest annual disclosure shows positive renewal above 100% but continued deceleration.
Segment-adjusted figures include allocations and are not product-unit economics. Gross retention, services cost, Direct Data economics, and cohort-level churn are not disclosed.
What are PitchBook's key moats?Morningstar investor Q&A · Public-company disclosure · 7 May 2025
Morningstar attributes PitchBook's position to collection infrastructure, analyst expertise, verification, workflow embed, and controlled integrations with AI providers.
Management's investor-relations answer. It is evidence of strategy and asserted mechanism, not independent proof of a moat or incremental product quality.
Morningstar Q2 2026 ResultsMorningstar, Inc. · Public-company disclosure · 29 July 2026
Morningstar reports continued PitchBook revenue growth with relatively flat licensed users and lower adjusted margin, partly attributed to growth hiring and AI infrastructure.
Quarterly company reporting does not isolate pricing, gross retention, connector economics, support labor, or full cost per accepted screening output.
PitchBook renewal rates dropped to 107%Morningstar investor Q&A · Public-company disclosure · 7 May 2025
Morningstar reported that PitchBook's annual revenue renewal rate fell to 107% in 2024 from 112% in 2023, with higher churn concentrated in non-core segments and some softness in core segments.
Company-reported aggregate. It does not identify gross retention, product-level switching events, export friction, or a causal relationship to data quality.
Research Process FAQsPitchBook · Official product documentation
PitchBook says its records combine public sources, self-reported submissions, automated collection, specialized data teams, manual quality assurance, and quarterly outreach to company management.
Vendor methodology. It acknowledges that requested updates may be only partially applied or omitted and supplies no slice-level accuracy or staleness audit.
The VC Data ProblemVC Lab · Practitioner account · 20 January 2026
VC Lab argues that voluntary reporting creates material blind spots in PitchBook's venture-fundraising data and disputes conclusions drawn from that coverage.
A strongly interested practitioner critique, not a reconciled coverage study. It supplies contrary evidence and a candidate slice for a formal overlap audit.