01
Intake & normalization
Manual
Analyst reads the deck, copies numbers into a spreadsheet, files the data room.
Adversarial AI
Deck and supporting documents parsed deterministically. Numbers, claims, and citations indexed in a structured record.
How Venture CapitalistsUse AI for Due Diligence.
Most "AI for VC" tooling is descriptive — it summarizes decks, scrapes founders, and tags markets. That work is useful, but it does not change the decision. The shift now underway is from descriptive analytics to adversarial evaluative intelligence: systems that argue against the deck, surface friction, and produce an auditable committee package.
MANUAL REVIEW
~4 hours
Per deck, before committee. Output quality depends on analyst availability and recency bias.
ADVERSARIAL AI AUDIT
~180 seconds
Same deck. Five evaluator agents, deterministic claim extraction, immutable audit hash on the committee brief.
What actually changes
The point is not speed. The point is that compression frees partners to spend the saved hours on the parts of diligence that AI cannot do — references, founder conversations, sector pattern recognition. Speed without rigor is a regression; rigor without speed is the status quo. Adversarial evaluation aims for both.
- • Every claim in the deck is captured with its source — no summary loss.
- • Disconfirming evidence is surfaced by design, not by analyst willpower.
- • The committee brief follows a fixed structure, so memos are comparable across deals and across weeks.
- • A "what this does NOT capture" section is mandatory, not optional.
The five-step process
Each row contrasts the manual workflow with the adversarial AI workflow used by the SenseCore Protocol.
02
Claim extraction
Manual
Highlights are pulled by hand. Soft claims often pass through without scrutiny.
Adversarial AI
Every quantitative and qualitative claim is extracted and tagged with its source page and surrounding context.
03
Adversarial evaluation
Manual
Partner forms a view, then looks for confirming evidence. Disconfirming evidence is easy to skip.
Adversarial AI
Five evaluator agents argue against the deck on structure, market, capital, execution, and timing. Disagreements are recorded, not resolved by majority.
04
Friction surfacing
Manual
Risks live in the analyst's head and a memo paragraph titled 'Concerns'.
Adversarial AI
Friction points are listed explicitly, each linked to the claim or absence that produced them.
05
Committee package
Manual
Memo written from scratch. Format depends on who is on point that week.
Adversarial AI
Institutional brief generated to a fixed structure: claims, friction, blind spots, what this does NOT capture, and an audit hash.
How this plays out at the analyst level
Partners are not the bottleneck in a VC pipeline — analysts are. They carry the highest-volume, lowest-leverage parts of diligence: intake, claim extraction, friction mapping, and the first memo skeleton. Multi-agent adversarial protocols compress that structured layer so analysts can spend the saved hours on the parts AI cannot do — founder references, sector conversations, and pattern recognition across the partner's portfolio.
| Metric | Manual | Adversarial AI |
|---|---|---|
| Time per first-pass review | ~4 hours | ~180 seconds |
| Decks reviewable per analyst-week | 8–12 | 60–80 |
| Disconfirming evidence surfaced | Analyst willpower | By construction |
| Cross-deal comparability | Format drift | Fixed schema |
| Audit trail | Email threads | Hashed brief |
The failure modes move rather than disappear. Anchoring on the founder's narrative is broken by evaluator agents that argue against it. Recency bias from the last deck reviewed is removed because every brief follows the same schema. Quiet omissions — no GTM, no unit economics, no defensibility — surface as absences rather than getting lost in prose. Score drift across analysts collapses because the evaluators, not the humans, generate the structured layer. What remains is judgment: which frictions matter for this fund, at this stage, in this cycle.
Descriptive vs evaluative intelligence
THE TWO MODES
Descriptive systems answer: what does this deck say? They are summarizers. They produce a tidier version of the founder's narrative and are easy to integrate, which is why they dominate the current "AI for VC" category.
Evaluative systems answer: where does this deck break under pressure? They run multiple adversarial passes, record disagreements, and refuse to produce a single confidence number when the inputs do not support one. The output is friction, not flattery.
What this does NOT capture
- • Founder integrity. References still belong to humans.
- • Sector dynamics that live in conversations, not documents.
- • Macro and policy timing. AI can flag sensitivity; it cannot predict the cycle.
- • A recommendation. Adversarial evaluation is decision-support, not advice.
See an adversarial audit on a real deck
The SenseCore Protocol is the evaluative system behind Zurvek. The demo report walks through claim extraction, friction surfacing, and the committee brief on a representative deal.
Zurvek is a decision-support system. Verdicts are analytical, not advisory. Investment decisions remain the responsibility of the reader.