Yuri Gui

The AI Frontier in VC

Published March 15, 2026 · Updated June 14, 2026

Where AI capability ends and human judgment begins — a living record.

What AI can and cannot reliably do in venture capital, across all AI tools on the market — not only claude-vc.

Scope of claude-vc

Claude-vc is an open-source Claude Code plugin for deal screening, investment memos, cap tables, term sheets, financial models, and KPI reports. It does not aim to replicate proprietary platforms like Harmonic (sourcing), Affinity (CRM), PitchBook (data), Luminance (legal), or Standard Metrics (portfolio monitoring).

Frontier-grade VC AI today requires a dozen+ MCP servers, data subscriptions, and specialized platforms — beyond what one open-source plugin can maintain. This document maps the full landscape, including capabilities claude-vc doesn't cover.

Status key

  • Can Do — AI handles reliably; reviewer sanity-checks, doesn't rewrite. For quantitative outputs, arithmetic and structure are correct given inputs; assumption quality is still human judgment.
  • WIP — outputs need substantive rework, or simple cases work but edges/hybrids fail.
  • Cannot (capability) — no viable AI path today.
  • Cannot (regulatory) — technically possible, but licensing or fiduciary requirements prevent AI reliance. Marked with "Regulatory:" prefix in enabler; won't advance until regulation changes.
  • Enabler1 — the AI advancement or product that made (or would make) this possible.

Last reviewed: 2026-06-14


Deal Sourcing & Outreach

CapabilityCan DoWIPCannotEnabler
Inbound pitch triage and initial scoringExtended thinking, Harmonic signal detection
New investment opportunity identificationHarmonic, Grata, EQT Motherbrain
Founder relationship trackingAffinity AI, 4Degrees
Co-investor and syndicate coordinationCRM integrations, email MCPs
Back-channel reference callsNeeds real-world interaction
LP and stakeholder communicationsLLM drafting, email MCPs

Market Research

CapabilityCan DoWIPCannotEnabler
Industry landscape synthesis from public sourcesWeb search, Perplexity, Claude
Market sizing (TAM/SAM/SOM) from pitch materialsPDF vision, extended thinking
Competitive landscape mapping from web researchWeb search tool use
Sector trend analysis from public dataWeb search, long-context reasoning
Regulatory environment scanningWeb search, legal corpus MCPs
Real-time market data and indicesBloomberg MCP, Refinitiv MCP
Primary market research (surveys, interviews)DiligenceSquared AI voice agents
Emerging market and whitespace identificationHarmonic signals, Grata agentic search

Company Research

CapabilityCan DoWIPCannotEnabler
Pitch deck data extraction (PDF)PDF vision, multimodal understanding
Public company profiling from web sourcesWeb search tool use
Report generation (DOCX and markdown)Native file generation
Private company data accessPitchBook Navigator, Crunchbase, Grata
Deal screening with structured scoring (0-100)Parallel subagents, extended thinking
Investment memo generation (10-section format)Parallel subagents, long-context generation
KPI benchmarking by auto-detected company typePython tool use, extended thinking
Comparable company analysis with market dataPitchBook + Perplexity MCP
Factual claim verification against primary sourcesWeb search with source citations
Systematic data room cross-referencing1M context, self-verification

Product Assessment

CapabilityCan DoWIPCannotEnabler
Product claims extraction from pitch materialsPDF vision, multimodal understanding
Feature set and roadmap summarizationLong-context reasoning
Technical architecture assessmentCode analysis tools, computer use
User retention and engagement pattern analysisAnalytics platform MCPs
Product-market fit validationNeeds usage data access, user research
Hands-on product testing and UX evaluationNeeds computer use at scale

Financial Analysis

CapabilityCan DoWIPCannotEnabler
Burn rate and runway analysisPython tool use
3-statement model generation (P&L, BS, CF)Python tool use, self-verification
Unit economics computation (LTV, CAC, payback)Python tool use, self-verification
Revenue projections (3-5 year forward)Python tool use, self-verification
KPI auto-detection and health assessmentExtended thinking, self-verification
Cohort and retention curve analysisAnalytics MCPs, Python tool use
Bulk portfolio-wide financial analysisChatFin, Chronograph, 1M context
Audit-grade financial statementsRegulatory: formal verification

Valuation

CapabilityCan DoWIPCannotEnabler
Pre/post-money round modelingCarta, Pulley, Python tool use
Multiples-based valuation with industry rangesPython tool use, web search
DCF analysis from user-provided assumptionsPython tool use, self-verification
Comparable company analysis with live dataPitchBook + Perplexity MCP
Precedent transaction analysisGrata, PitchBook + Perplexity MCP
Conviction weighting on valuation outputsNeeds calibrated confidence, human judgment

Deal Structuring & Negotiation

CapabilityCan DoWIPCannotEnabler
Cap table modeling and dilution analysisCarta, Pulley, Python tool use
SAFE and convertible note conversionCarta, Pulley, OCF standard
Multi-series liquidation waterfallCarta, Pulley, Eqvista
Exit scenario modeling at multiple valuationsCap table platforms, Python tool use
Term sheet red-flag identification (NVCA baseline)Spellbook (10M+ contracts reviewed)
Cap table platform sync (Carta, Pulley) via MCPCap table platform MCPs
Negotiation strategy and counter-offer structuringNeeds relationship context, game theory
Binding legal document generationRegulatory: legal licensing required

Technical Due Diligence

CapabilityCan DoWIPCannotEnabler
Technical claims summarization from materialsLong-context reasoning
Technology stack identificationWeb search, code analysis tool use
IP and patent strength evaluationIPRally, Patsnap, Patlytics
Codebase quality and architecture reviewCode analysis agents, computer use
Scalability and infrastructure assessmentNeeds system access, load testing
Technical team capability evaluationNeeds real-world interaction

Legal Due Diligence

CapabilityCan DoWIPCannotEnabler
Common provision pattern flaggingSpellbook, Luminance
NVCA baseline term comparisonSpellbook VC clause library
Contract review (SHA, IP assignments, employment)Luminance, Kira/Litera (64% Am Law 100)
Regulatory compliance analysisEmerging legal AI tools
Legal opinions and formal adviceRegulatory: legal licensing required
Cross-jurisdiction tax and structuringRegulatory: licensing + tax law MCPs

Financial Due Diligence

CapabilityCan DoWIPCannotEnabler
Individual financial document analysisHebbia Matrix, PDF vision
Private company financial dataPitchBook, Morningstar, Grata
Financial model internal consistency checksPython tool use, self-verification
Portfolio company data connectivityStandard Metrics, Chronograph MCPs
Data room systematic cross-referencing1M context, self-verification
Historical financials verificationNeeds SEC EDGAR MCP, audit trail access
Tax and transfer pricing analysisRegulatory: licensing + tax law MCPs

Portfolio Management

CapabilityCan DoWIPCannotEnabler
CRM and deal pipeline trackingAffinity AI (auto-capture, relationship intel)
One-time portfolio summary generationExtended thinking, structured output
Board deck preparation assistanceLong-context generation, native file output
Ongoing portfolio monitoringStandard Metrics, Chronograph, ChatFin
Scheduled recurring reportsStandard Metrics, Visible.vc automation
Anomaly detection and alertsChatFin anomaly engine
LP reporting preparationclaude-vc portfolio skill, Standard Metrics

Investment Decision & Closing

CapabilityCan DoWIPCannotEnabler
IC preparation materialsExtended thinking, long-context generation
Decision framework structuringExtended thinking, structured output
Invest/pass recommendationNeeds calibrated confidence, human judgment
Founder and team qualitative assessmentNeeds real-world interaction
IC facilitation and votingNeeds multi-user collaboration
Closing coordination and executionNeeds legal tooling, payment systems
Post-closing deliverable managementWorkflow automation, persistent agents

Changelog

Dates reflect when the enabler shipped, not when adoption matured.

DateChangeCapability affectedDirectionEnabler
2022-09AI contract drafting and reviewCommon provision pattern flaggingHuman -> WIPSpellbook launch
2022-11AI-powered deal sourcing signalsNew investment opportunity identificationHuman -> WIPHarmonic Series A
2022-11Basic memo drafting and market summariesInvestment memo generationHuman -> WIPChatGPT (GPT-3.5)
2023-03Professional-grade investment analysisIndustry landscape synthesisWIP -> Can DoGPT-4 reasoning quality
2023-06Structured API tool callingKPI benchmarking, financial model checksHuman -> WIPOpenAI function calling
2023-07Full document analysis (100K context)Pitch deck data extractionHuman -> Can DoClaude 2
2023-09AI relationship intelligence in CRMFounder relationship trackingHuman -> WIPAffinity AI
2023-11Reliable structured data extractionKPI benchmarking, deal screeningWIP -> Can DoGPT-4 Turbo, JSON mode, 128K context
2024-03Pitch deck and chart visual analysisPitch deck data extraction, product claimsHuman -> Can DoClaude 3 vision + 200K context
2024-05AI-orchestrated multi-tool workflowsDeal screening, financial model generationHuman -> WIPClaude tool use GA
2024-06Cost-effective AI analysis at pipeline scaleInbound pitch triageWIP -> Can DoClaude 3.5 Sonnet
2024-08Perfect structured extraction (100% schema)KPI benchmarking, financial model checksWIP -> Can DoOpenAI structured outputs
2024-09Native PDF document processingPitch deck data extraction, document analysisHuman -> Can DoClaude PDF support
2024-10Computer-use automation for legacy toolsTechnical architecture assessmentHuman -> WIPClaude computer use beta
2024-10Multi-agent due diligence workflowsSystematic data room cross-referencingHuman -> WIPCrewAI maturity, AutoGen
2024-11Natural-language private company queriesPrivate company data accessHuman -> WIPPitchBook Navigator + OpenAI
2024-11Standardized AI-to-data connectivityPortfolio company data connectivityHuman -> WIPMCP (Model Context Protocol)
2025-02Deep financial reasoning on demand3-statement model generation, DCF analysisHuman -> WIPExtended thinking (Claude 3.7)
2025-02Agentic coding for custom VC toolsUnit economics computation, burn rate analysisHuman -> WIPClaude Code research preview
2025-03Real-time market intelligence in AISector trend analysis, competitive mappingHuman -> Can DoClaude web search
2025-05Production-grade agentic VC toolingReport generation, exit scenario modelingWIP -> Can DoClaude Code GA
2025-07Standard Metrics MCPPortfolio company data connectivity--Lookup-only; no anomaly detection or cross-fund aggregation
2025-09End-to-end document productionReport generation (DOCX), board deck prepHuman -> Can DoNative file generation
2025-11Private market data at scale via MCPPrivate company data accessWIP -> Can DoPitchBook Navigator MCP
2026-02Full data room in single contextSystematic data room cross-referencingHuman -> Can Do1M token context (Claude 4.6)
2026-03PitchBook data in conversational AIComparable analysis, precedent transactionsWIP -> Can DoPitchBook + Perplexity MCP
2026-03Fund-level return metricsIRR, MOIC, DPI, TVPI, PME computationHuman -> Can DoPython tool use, native file generation
2026-03XLSX export for financial outputsSpreadsheet generation for cap tables, modelsHuman -> Can DoNative file generation + skill flags
2026-03Parallel multi-agent deal screeningFull screening with 6 concurrent agentsWIP -> Can DoMulti-agent orchestration via subagents
2026-03Side-by-side company comparisonStructured comparison of 2-4 companiesWIP -> Can DoLong-context reasoning, structured output
2026-03Customizable due diligence checklistsStage+sector-specific DD checklistCannot -> Can DoLong-context generation, structured prompting
2026-03One-shot portfolio reporting for LPsLP-ready portfolio summaryCannot -> Can DoExtended thinking, native file generation
2026-04Multiples valuationMultiples-based valuation with industry rangesWIP -> Can DoPython tool use, web search
2026-04Self-verified financial analysis3-stmt model, unit econ, DCF, consistencyWIP -> Can DoClaude Opus 4.7 self-verification
2026-04Improved reasoning (GPQA 94.2%)KPI auto-detection and health assessmentWIP -> Can DoClaude Opus 4.7 extended thinking
2026-04Opus 4.7 vision (3.75MP, 3x prior)High-res document and diagram analysis--Claude Opus 4.7
2026-04GPT-5.4 reaches 1.05M token contextFull data room in single context (non-Claude)--GPT-5.4 (OpenAI)
2026-04Carta Plugins for ClaudeCap table platform sync via MCP--Carta-hosted only; Pulley/AngelList/Eqvista excluded
2026-04Gemini Deep Research Max previewLong-horizon research synthesis with native MCP--Gemini 3.1 Pro + announced (not GA) FactSet/S&P/PitchBook MCPs
2026-04GPT-5.5 GAFrontier baseline--1M ctx, ARC-AGI-2 85%; AA-Omniscience: 86% confab on errors → Opus 4.7 for citations
2026-041M-context reasoning on open weightsSelf-hosted full-data-room reasoning--DeepSeek V4-Pro
2026-04Affinity hosted MCP (beta)Founder relationship tracking--Scale+ tiers; internal signals only; no external discovery
2026-04Agentic finance AI at 250+ institutionsSystematic data room cross-referencing--Rogo Felix
2026-04AI contract redlining bundled in WordCommon provision pattern flagging--Microsoft Legal Agent for Word (early-access)
2026-05GPT-5.5 Instant default in ChatGPTInvestment memo quality--OpenAI claims 52.5% fewer hallucinations; AA-Omniscience contradicts
2026-05SEC EDGAR MCP servers in productionPublic comparables / IPO filings only--sec-edgar-mcp, EdgarTools — does NOT cover private companies
2026-06Fable 5 — most powerful public modelFrontier baseline--Anthropic Fable 5 (June 9); SOTA, but high-risk queries fall back to Opus 4.8
2026-06Fable 5 / Mythos 5 access suspendedFrontier baseline--US gov directive barred foreign-national access; Anthropic disabled both June 12

Methodology

Changelog conventions

  • Entries reflect industry-wide capability shifts, attributed to the model release, product, or integration that enabled them.
  • Regressions logged as Can Do → WIP or WIP → Cannot with date and cause. None recorded yet — reflects the doc's youth (March 2026 baseline), not an assumption of one-way progress.

Known limitations

  1. Enabler attribution is best-effort, not causal. A capability became reliable around the time the enabler shipped.
  2. Anthropic-skewed. The changelog over-indexes on Anthropic releases (what we test directly); other providers covered when materially frontier-changing.
  3. Self-verification (Opus 4.7, April 2026) means the model writes checks for its own outputs — test assertions, re-reads, cross-checks — before reporting. Catches arithmetic and structural errors, not bad assumptions.

Footnotes

  1. Enabler = the AI product, feature, or integration that made (or would make) a capability possible. Either intrinsic (available to any user of the model) or integration-gated (requires third-party subscription or MCP).

    Common intrinsic: extended thinking, Python tool use, web search, PDF vision, native file generation, 1M context, structured output, self-verification.

    Common integration-gated: PitchBook Navigator, Carta/Pulley, Spellbook, Harmonic, Grata, Standard Metrics/Chronograph, Hebbia Matrix.

    A "Can Do" with an integration-gated enabler requires that subscription — it is not universally available.


Source: claude-vc/docs/frontier.md