Applied AI / Private MarketsDesign Concept

Deals don't stall because teams lack dashboards. They stall when nobody sees the next blocker coming.

Deal OS is an AI-native transaction operating system for private-market secondary transfers. It surfaces blockers before they stop deals, makes every AI action explainable, and keeps humans in the approval seat.

Simulated concept · all transaction data, parties, and documents are fictional

Simulated scenario

A deal in motion.

A GP email arrives, the system identifies a missing document, and a blocker forms. Walk through how Deal OS detects, explains, recommends, and recalculates — with a human approving every action.

Closing readinessNot Ready
3/7 conditions met
Fund:Northbridge Growth Fund III
Seller:Harbor Family Holdings
Buyer:Meridian Secondary Capital
Target close:December 15, 2026
7 days to target close

GP email — incoming message

Fromsarah.okonkwo@northbridgefunds.com
SubjNorthbridge Growth Fund III — Transfer Consent Update
DateDecember 8, 2026, 09:14

We are progressing our review of the proposed transfer from Harbor Family Holdings to Meridian Secondary Capital.

To finalise the GP approval package, we require the following outstanding item:

Updated beneficial ownership documentation for Meridian Secondary Capital (Articles of Formation or equivalent, reflecting current ownership structure).

Our compliance team needs this document before we can complete our review and issue transfer consent. Please have this submitted by December 11, 2026.

— Sarah Okonkwo, LP Relations, Northbridge Capital Management

Simulated email · fictional content

AI detection

Document requirement identified

The GP email requests beneficial ownership documentation for Meridian Secondary Capital before transfer consent can be issued.

Blocking condition

Buyer ownership documentation

Required for GP approval package → transfer consent → closing

Days to target close

7

Document must be submitted by Dec 11 to keep closing on track.

The thinking behind it

Three decisions that shaped the design.

01

Design around dependencies, not steps

Most deal management tools model transactions as checklists. The real problem is dependency chains — one missing document blocks five downstream items, and nobody sees the cascade. I chose to make dependencies the primary data model, not a secondary annotation.

Trade-off

Dependency graphs are harder to build and maintain than flat checklists. Teams accustomed to status columns need a mental shift.

02

Make intelligence explainable, not just actionable

A system that says 'this deal is blocked' is useful. One that says 'this deal is blocked because the GP approval package is missing the buyer ownership document, which arrived in an email three days ago' is trusted. I required every AI output to link to its source.

Trade-off

Source-grounded explanations cost more to build and surface more complexity to users. Some users prefer a clean summary.

03

Separate assistance from authority

The system can recommend, draft, and flag — but it cannot approve, send, or close without a human step. LP transfers, GP consent, legal execution, and investor communications require a person to confirm. Automation handles preparation; humans retain authority.

Trade-off

More friction per action. Reduces the throughput gains from automation, but avoids the much larger cost of an unchecked action in a regulated transaction.

Explore the system

Architecture and detail.

Intake Agent

Parses incoming deal flow from email, CRM webhooks, and broker submissions. Extracts fund name, size, fund type, and timeline. Flags ambiguity for human review.

Intelligence Agent

Matches buyer and seller profiles against investment preferences, eligibility constraints, and prior transaction history. Ranks candidates and explains each recommendation.

Diligence Agent

Extracts structured data from LP agreements, capital account statements, and fund documents. Identifies missing sections and inconsistencies. Produces source-grounded summaries — no hallucinated facts.

Execution Agent

Monitors transaction tasks, counterparty responses, and deadline commitments. Builds the dependency graph. Surfaces blockers and prepares communication drafts for broker approval.

Closing Agent

Tracks document receipt, verification, GP consent status, and regulatory conditions. Maintains a live readiness score. The score does not advance unless conditions are verified, not just received.

Proposed architecture · not implemented production infrastructure

Most deal management tools model transactions as linear stages. Deal OS models them as a directed dependency graph: each condition either satisfies or blocks downstream conditions. The closing readiness score is a function of the graph state, not a manually maintained field.

When a document arrives, the system marks the document condition as received — not verified. Verification is a separate step that may require compliance checks, GP review, or legal confirmation. A received document does not advance the closing score until verification is confirmed.

GP consent is treated as a distinct authority node. The system can track whether the GP approval package is complete, but it cannot grant or infer consent on the GP's behalf. Transfer consent remains blocked until the GP acts.

Data Sources

Email / CRM / broker submissions / fund documents / LP agreements / capital account statements / compliance databases

Context Processing

Entity resolution · document classification · timeline extraction · counterparty matching · permissions

Intelligence Layer

Dependency graph construction · blocker detection · evidence retrieval · recommendation engine · confidence scoring

Execution Layer

Draft preparation · broker approval flows · task orchestration · deadline tracking · audit trail

Closing Layer

Document receipt tracking · verification gates · GP consent monitoring · readiness score · settlement readiness

Proposed architecture · not production infrastructure

Phase 1 — Document and dependency intelligence

AI-assisted document extraction, gap detection, and blocker visibility across a single transaction type. Human review required before any communication.

Phase 2 — Transaction orchestration

Dependency graph engine, multi-party task coordination, broker-approval communication drafts, and closing readiness scoring across fund types.

Phase 3 — Portfolio intelligence

Cross-deal pattern analysis, counterparty profiling, origination signals, and recurring blocker detection. Feedback loop from execution back to deal sourcing.

How success would be measured if this were a production system:

  • —Time from onboarding to first signature — median days per deal
  • —Blocker resolution time — median hours from blocker creation to resolution, by type and owner
  • —Closing readiness score delta — difference between score at deal open vs. at signing
  • —Document round-trips — median number of document versions exchanged per closing
  • —Off-platform communications — proportion of closings where parties reported communicating outside the system
  • —Return rate within 48 hours of new blocker — proportion of assigned owners acting within 2 days

What the system cannot do

  • —Grant GP consent or issue transfer consent on behalf of the GP
  • —Verify a document as legally sufficient — it can flag receipt and surface content, but legal verification requires human review
  • —Predict closing timelines — it surfaces current state, not future certainty
  • —Substitute for legal, compliance, or regulatory counsel

Regulatory considerations

Secondary market transfers in private markets are regulated transactions. Any production implementation would require legal review of communication flows, consent mechanisms, and data handling in each jurisdiction. This concept does not model regulatory compliance — it models the product workflow layer above it.

Want to go deeper?

Building something in private markets, FinTech, or regulated workflows?