AI Engineer · Published 2026-10-16

AI Agents Cut 3-Day Portfolio Builds to Minutes — Jack McDonald, Apollo & Michael Manapat, Rowspace

privateReview: unreviewedConfidence: high
AI agentsstructured creditportfolio managementdata engineeringasset managementdata-provenanceenterprise-agentsHuman-in-the-loopdocument-processingsource-of-truthstructured-creditdomain-specific-agents
Open on YouTube ↗

Summary

Overview

  • Speaker: Michael Manapat & Jack McDonald
  • Channel: AI Engineer
  • Main topic: How AI agents and data engineering platforms transform private equity and structured credit operations at scale.
  • Purpose: To showcase practical enterprise AI implementation in structured credit and alternative asset management by pairing a software platform founder with an operating executive. Michael Manapat (Co-founder, Rowspace) and Jack McDonald (Principal and COO, Redding Ridge Capital / Apollo) discuss how financial institutions transition AI from a 'nice to have' to mission-critical operations. They examine real-world use cases such as portfolio construction and warehouse purchase activity, contrasting general-purpose models like Claude with domain-specific, data-backed enterprise AI architectures.

Topic Map

Introduction and Enterprise AI Context

  • Explanation: Introductions of speakers and overview of Rowspace as an agentic platform solving fragmented financial data problems for Apollo partners.
  • Key claims:
    • AI agents act as data engineers to reconcile information across disparate systems.
    • There is a wide gap between AI being 'nice to have' and mission-critical.
  • Examples:
    • Portfolio managers and traders needing faster, more accurate data integration.
  • Terminology:
    • Rowspace
    • Apollo
    • Redding Ridge Capital
    • Structured Credit
  • Why it matters: Establishes the operational reality of deploying AI in multi-billion dollar private asset management firms.

Portfolio Construction Use Case

  • Explanation: Comparing manual portfolio construction workflows against AI-accelerated workflows that reduce a multi-day process to minutes.
  • Key claims:
    • CLOs are test-specific and rules-based with deep indenture documents.
    • Portfolio construction went from taking 2-3 days in manual Excel to minutes using Rowspace.
  • Examples:
    • Portfolio managers reviewing indenture documents to build mock portfolios of 200-400 credit names.
  • Terminology:
    • CLO
    • Portfolio Manager
    • Waterfall
    • Indenture Documents
  • Why it matters: Demonstrates tangible productivity gains and accuracy improvements in complex financial compliance.

Warehouse Purchase Activity and Data Provenance

  • Explanation: Shifting from manual bank account statement review for trade reconciliation to automated, real-time visualization and verification.
  • Key claims:
    • Historical trade-wrecking processes required manual bank statement analysis and email recaps.
    • Rowspace automates report generation and visualization in under a minute.
  • Examples:
    • Weekly trade-rec activity tracking sale-purchase activity across multiple bank accounts.
  • Terminology:
    • Trade-rec
    • Bank Financing
    • Fixed Income
    • Data Provenance
  • Why it matters: Highlights how data provenance and automated verification replace error-prone manual auditing.

Trust, Accuracy, and General LLMs vs. Enterprise Platforms

  • Explanation: Addressing why generic models like ChatGPT are insufficient for mission-critical financial decisions without proprietary, verified data layers.
  • Key claims:
    • General LLMs are great for research, but lack proprietary internal data.
    • Rowspace sits on top of trusted internal data sources to override external generic information.
  • Examples:
    • Resolving conflicts when 20 different sources yield only one correct piece of financial data.
  • Terminology:
    • LLMs
    • Golden Source of Truth
    • Bloomberg
    • Data Quality
  • Why it matters: Explains the foundational requirement of trust and verification in enterprise AI deployments.

Future Outlook and Advice for AI in Finance

  • Explanation: Discussion on how AI will reshape software engineering, law, customer service, and asset management over the next 5 years.
  • Key claims:
    • Adopting AI for routine tasks frees capacity for higher-level creative and strategic thinking.
    • Financial institutions must build on the ground with domain experts to succeed.
  • Examples:
    • Investing time in sitting with front-office teams to understand complex workflows.
  • Terminology:
    • OPEX
    • Asset Management
    • Enterprise Adoption
  • Why it matters: Provides strategic guidance for organizations looking to integrate AI agents into complex operational workflows.

Key Points

Eliminating the Data Gap in Structured Credit

  • Explanation: Financial firms deal with massive, fragmented, private datasets spread across thousands of loan pools and legal documents.
  • Evidence: Apollo manages approximately $50 billion in assets across structured credit ecosystems.
  • Practical implication: AI agents must ingest and reconcile multi-source data before high-stakes decisions can be automated.

Moving from Manual Excel to Automated Minutes

  • Explanation: Portfolio managers historically spent days building Excel models and mock portfolios manually.
  • Evidence: Process reduced from 2-3 days in manual Excel to minutes with real-time data.
  • Practical implication: PMs achieve 80-90% automated completion and spend their time on final strategic refinement.

The Necessity of Data Provenance

  • Explanation: General-purpose models hallucinate or lack proprietary context; enterprise AI requires a verified golden source of truth.
  • Evidence: For a single data point drawn from 20 sources, only one source is correct and verified by internal systems.
  • Practical implication: Enterprises must own their data layer and deploy domain-specific platforms rather than relying solely on off-the-shelf chat interfaces.

Frameworks, Models & Processes

Rowspace Enterprise AI Integration Framework

  • How it works: Combines deep boots-on-the-ground workflow analysis with automated data reconciliation to produce verified financial models.
  • Components:
    • Workflow immersion with domain experts
    • Multi-source data ingestion and validation
    • Automated document and spreadsheet parsing
    • Human-in-the-loop final review (80/20 rule)
  • When to use: When deploying AI into complex, highly regulated, private data environments like asset management and structured credit.

Examples & Case Studies

A portfolio manager needed 2 to 3 days in manual Excel to construct a mock portfolio from hundreds of credit names.

  • Illustrates: The inefficiency of traditional portfolio construction.
  • Lesson: AI agents can ingest deep indenture documents and generate 80-90% complete portfolios in minutes.

Weekly warehouse purchase trade reconciliation required manually reviewing bank statements and emailing management recaps.

  • Illustrates: The manual overhead in fixed-income operations.
  • Lesson: Visualizing reports and automating reconciliation reduces multi-hour weekly chores to repeatable minute-level processes.

Actionable Takeaways

  • Immediate:
    • Automate routine document review and data reconciliation tasks to free up team capacity.
    • Partner with domain experts on the ground to map out precise workflow bottlenecks.
  • Strategic:
    • Build or adopt platforms that ensure a verified 'golden source of truth' rather than relying solely on generic LLMs.
    • Embrace an open-minded culture toward technology adoption to maintain a competitive edge in alternative asset management.
  • Questions to investigate:
    • Where are the hidden data silos in your firm's document and compliance workflows?
    • How do you establish trust and verify AI output against internal proprietary systems?

Claims Worth Verifying

  • Rowspace reduces portfolio construction time from 2-3 days to minutes. (empirical)
  • Redding Ridge Capital manages approximately $50 billion in structured credit assets. (factual)

Notable Quotes

"There's a pretty wide gap between when AI is nice to have and when AI can actually become mission critical to a business." (at 2:16) "Claude and the open AIs of the world and like the chat GPTs, all the LLMs are very beneficial, but it's all general information." (at 12:09) "For one piece of data, that might come from 20 different sources and there's only one that's actually right." (at 13:41)

Compressed Summary

  • Rowspace acts as an AI data engineering platform for financial services partners like Apollo.
  • AI agents cut CLO portfolio construction from days to minutes.
  • Automating warehouse purchase trade reconciliation saves hours of manual bank statement auditing.
  • Generic LLMs provide general research, but enterprise AI requires proprietary data provenance and verification.
  • Keywords: ai agents, structured credit, portfolio management, data engineering, asset management
  • Core insight: Enterprise AI adoption in finance succeeds when AI agents act as rigorous data engineers atop verified internal golden sources of truth rather than relying on generic LLMs.

Core insights

6
Mental Modelmedium noveltymoderate evidence

In regulated financial workflows the agent's primary job is data engineering — reconciling conflicting values across many internal and external sources into an authoritative record — rather than reasoning over already-clean data. The reasoning layer is comparatively commoditized; the reconciliation and provenance layer is the hard, defensible part.

Why it matters

If reconciliation is the bottleneck, engineering effort should go into source-of-truth resolution, conflict policy, and lineage tracking rather than prompt/agent-loop tuning. It relocates the hard problem from the model to the data plane.

Generalization

Any enterprise agent operating on multi-source operational data should be architected as a reconciliation engine with an explicit conflict-resolution policy per field, not as a chat wrapper over a data warehouse.

AI agents act as data engineers to reconcile information across disparate systems.
Open source video
For a single data point drawn from 20 sources, only one source is correct and verified by internal systems.
Open source video
Empirical Resultmedium noveltymoderate evidence

The useful operating point for these agents is roughly 80-90% autonomous completion with a human closing the final gap, rather than end-to-end autonomy. The economic win comes from compressing a 2-3 day manual build to minutes and reassigning human time to strategic refinement.

Why it matters

It gives a concrete target for agent product design: optimize the last-mile handoff surface (diffing, review UI, what the agent flags as uncertain) because that is where the residual human time is spent, not on pushing raw accuracy past 90%.

Generalization

For high-stakes analytical work, measure value as time-to-90%-draft plus quality of the human review interface, rather than as full-task success rate.

Portfolio construction went from taking 2-3 days in manual Excel to minutes using Rowspace.
Open source video
PMs achieve 80-90% automated completion and spend their time on final strategic refinement.
Open source video
Architecturemedium noveltymoderate evidence

Trust in an enterprise agent is a data-lineage property, not a model property. When 20 sources disagree, the platform must cite and override with the verified internal record; the model's parametric knowledge is explicitly treated as lower-authority than internal provenance.

Why it matters

This implies a precedence hierarchy enforced outside the LLM — internal verified source > external vendor (Bloomberg) > model priors — with attribution rendered to the user. Without it, outputs are unauditable and unusable for decisions.

Generalization

Give agents a deterministic source-precedence resolver and per-field provenance attached to every emitted value; treat model memory as a fallback tier, never the top tier.

Rowspace sits on top of trusted internal data sources to override external generic information.
Open source video
Enterprises must own their data layer and deploy domain-specific platforms rather than relying solely on off-the-shelf chat interfaces.
Open source video
Tradeofflow noveltymoderate evidence

General-purpose models are positioned as research assistants, not system-of-record components: they are strong at open-ended exploration but structurally lack the proprietary context required for decision-grade output, so they are deliberately kept out of the data-authority path.

Why it matters

This is a clean architectural boundary — the LLM is a reasoning/parsing engine, the platform owns truth. It suggests avoiding architectures where the model's own recall is implicitly trusted for numeric or contractual facts.

Generalization

Separate the reasoning engine from the truth engine; route all factual assertions through a retrieval/verification layer and treat unverified model output as a hypothesis to be checked.

General LLMs are great for research, but lack proprietary internal data.
Open source video
Mechanismmedium noveltymoderate evidence

Deep, rules-based legal/financial documents (indentures, waterfall tests for CLOs) are an unusually favorable agent substrate: the rules are explicit and checkable, so the agent can be validated against the document rather than trusted. This is why automation rates reach 80-90% in this domain.

Why it matters

Suggests a selection heuristic for where agents will work first: domains where the governing rules are written down and mechanically checkable, giving a free oracle for evaluation and self-verification.

Generalization

Prioritize agent deployments in domains with machine-checkable specifications (contracts, schemas, test suites, compliance rules) because they supply both the task definition and the verifier.

CLOs are test-specific and rules-based with deep indenture documents.
Open source video
Portfolio managers reviewing indenture documents to build mock portfolios of 200-400 credit names.
Open source video
Practicelow noveltymoderate evidence

Deployment succeeded because engineers embedded with front-office teams to map exact workflow bottlenecks before building, rather than shipping a generic platform capability. Workflow immersion precedes automation.

Why it matters

Enterprise agent quality is bounded by task specification fidelity; the spec lives in the operators' heads and in the artifacts they manipulate, so extraction of the spec is a first-class engineering phase.

Generalization

Budget explicit discovery cycles where engineers sit with operators and capture artifacts (spreadsheets, emails, statements) as fixtures for both the agent build and its regression suite.

Investing time in sitting with front-office teams to understand complex workflows.
Open source video
Workflow immersion with domain experts
Open source video

Deep dives

4

Per-field precedence and arbitration across conflicting internal systems of record

Research question

When multiple internal systems disagree on the same field (not just internal vs. external), what deterministic precedence and arbitration policy yields the lowest field-level error rate, and how much of that policy can be learned from lineage rather than hand-specified?

Why

The pass-1 insight establishes that trust is a lineage property enforced outside the model, but the described hierarchy (internal > vendor > model priors) collapses as soon as two internal systems conflict. If arbitration is undefined, the 'authoritative record' the agent emits is unauditable, and the reviewer is forced back into full re-derivation, wiping out the time savings that justified the deployment.

AI agents act as data engineers to reconcile information across disparate systems.
Open source video
Rowspace sits on top of trusted internal data sources to override external generic information.
Open source video
There is a wide gap between AI being 'nice to have' and mission-critical.
Open source video
Source video

Review-surface information design for auditing a 200-400 name artifact in minutes

Research question

Which review-surface design — full provenance-annotated table diff, exception-only queue, or flag-confidence triage — minimizes time-to-signoff without increasing missed-error rate on a large multi-name artifact?

Why

The residual human effort sits entirely in the last-mile handoff, so the interface, not the model, is the binding constraint on realized value. If the agent's uncertainty flags are imprecise, the reviewer must redo the whole artifact and the days-to-minutes gain is fictive; this makes flag precision and diff rendering first-class engineering targets rather than UI polish.

Portfolio managers reviewing indenture documents to build mock portfolios of 200-400 credit names.
Open source video
Portfolio construction went from taking 2-3 days in manual Excel to minutes using Rowspace.
Open source video
Source video

Operationalizing '80-90% completion': field accuracy, artifact completeness, or rework rate

Research question

Does a single 'percent complete' figure predict realized time savings, or does rework rate (fraction of agent-produced fields the expert rewrites) dominate the outcome regardless of headline completion numbers?

Why

Without a defined error taxonomy the productivity claim is unfalsifiable and cannot be used to decide where to invest engineering effort. A rework-centric metric would redirect effort from raising raw accuracy to improving which fields get flagged, and would let teams compare platforms on a common basis instead of on demo narratives.

Portfolio construction went from taking 2-3 days in manual Excel to minutes using Rowspace.
Open source video
Rowspace automates report generation and visualization in under a minute.
Open source video
Source video

Specification explicitness as the predictor of achievable agent automation rate

Research question

Across domains, does automation rate scale with the availability of a machine-checkable specification (written rules, schema, testable compliance condition), and what is the automation ceiling in document-grounded work that lacks such a verifier?

Why

The structured-credit result may be an artifact of indentures being explicit and mechanically checkable, which supplies a free oracle for both task definition and self-verification. If so, domain-selection for agent pilots becomes a measurable property of the corpus rather than a judgment call, and teams stop burning quarters on domains where no verifier exists.

CLOs are test-specific and rules-based with deep indenture documents.
Open source video
Portfolio managers reviewing indenture documents to build mock portfolios of 200-400 credit names.
Open source video
Source video

Article ideas

3

Your Agent's Hard Problem Is Reconciliation, Not Reasoning

In decision-grade enterprise work the reasoning layer is commoditized and the reconciliation-and-provenance layer is the entire product; teams that tune prompts and agent loops while leaving source precedence unspecified ship confidently wrong numbers at machine speed.

Angle

Reframe the engineering budget: show that when a value is drawn from many sources only one is authoritative, so the deliverable is an explicit per-field conflict policy plus rendered lineage — and a chat wrapper over a warehouse is structurally disqualified from decision-grade use.

Source video

The 80% Agent Is the Product

Optimizing for full autonomy is a misallocation; the economically dominant design target is time-to-90%-draft plus the precision of the uncertainty surface, because the residual 10-20% is where all remaining human cost lives and a bad flag forces a full re-derivation.

Angle

Argue against autonomy roadmaps by treating the review interface — diffs, exception queues, provenance display, flag precision — as a first-order systems concern with measurable effects on signoff time and missed-error rate.

Source video

Automate Where the Rules Are Already Written Down

High automation rates are a property of the corpus, not the model: domains whose governing rules are explicit and mechanically checkable hand the agent both its task definition and its verifier, so specification explicitness should be the primary screening criterion for agent pilots.

Angle

Convert a sector anecdote into a selection heuristic — score candidate domains on rule explicitness and verifier availability before committing engineering effort, and predict which ones will stall without one.

Source video

Project ideas

3

Precedence Engine

gatehouse

A deterministic per-field source-precedence resolver with rendered lineage reduces expert rework rate on a multi-source artifact by a measurable margin versus a retrieval-only baseline that lets the model choose among conflicting values.

Proof of concept

Ingest four realistic source tiers (internal system of record A, internal system of record B, external vendor feed, document-extracted values) for ~50 fields; the agent emits value, chosen source, conflict set, and confidence; the review surface shows only conflicted or low-confidence fields with the rejected alternatives and their provenance.

Measurement

Field-level accuracy against a hand-verified ground truth; fraction of emitted fields rewritten by an expert; misses on a seeded conflict set (recall of deliberately injected wrong values); time-to-signoff on the full artifact.

Source video

Exception-Only Review Bench

movement-lab

An exception-only review queue with calibrated confidence flags cuts time-to-signoff on a 200+ item artifact by more than half versus a full provenance-annotated table diff, without increasing missed-error rate.

Proof of concept

Generate a 200-name artifact from a rules-based document set with a known ground truth; render three review surfaces (full diff table, exception-only queue, confidence triage) and run the same reviewers across each; log signoff time, edits, and errors escaped.

Measurement

Time-to-signoff per surface; edits per artifact; missed-error rate measured against the seeded ground truth; reviewer-reported trust and re-derivation incidents.

Source video

Spec-Explicitness Automation Curve

beyond-evals

Achievable agent automation rate is predicted by the availability of a machine-checkable specification in the target corpus, not by model capability, so automation rate should decline monotonically as verifier availability falls across domains at fixed model and effort budget.

Proof of concept

Hold model and tooling constant; run the same document-grounded agent on four tasks spanning a fully checkable rule set, a partially checkable set, a schema-constrained extraction task, and a judgment-heavy narrative task; grade each against its available oracle and record where self-verification is impossible.

Measurement

Automation rate per task; gap between agent self-reported confidence and graded correctness per task; fraction of outputs the agent can mechanically self-verify; slope of automation rate versus a verifier-availability score.

Source video

Architectural implications

3

The agent's value concentrates in reconciling fragmented private datasets across loan pools, legal documents, and bank statements before any decision is produced.

Before

Agent sits directly on top of a model plus ad-hoc retrieval; data quality is assumed or handled downstream by humans.

After

A dedicated ingestion/validation/provenance layer sits between sources and the reasoning layer, with per-field precedence rules.

Consequence

Most engineering effort shifts to connectors, entity resolution, and conflict policy; the LLM becomes a replaceable component and model swaps become low-risk.

Source video

Human-in-the-loop is scoped to the final 10-20% review rather than spread across the workflow.

Before

Humans check intermediate steps and re-derive results, or humans are excluded entirely and outputs are unverified.

After

Agent produces a complete near-final artifact plus the delta/uncertainty surface; human reviews only flagged items.

Consequence

The review interface (diffing, provenance display, flag precision) becomes a core product surface and a primary lever on realized time savings.

Source video

Generic chat interfaces are rejected for mission-critical financial decisions in favor of domain platforms that own the data layer.

Before

Off-the-shelf LLM chat as the delivery surface for enterprise analysis.

After

Domain-specific platform with embedded verified data, reproducible report generation, and visualization.

Consequence

Vendors who do not own or integrate the source-of-truth layer are structurally locked out of decision-grade use cases.

Source video

Tradeoffs and failure modes

3

Model knowledge vs. verified internal data

Benefit

General LLMs provide broad research capability and fast exploration without bespoke data engineering.

Cost or risk

They lack proprietary context and can assert generic or stale facts where only one internal source is correct, producing confidently wrong decision inputs.

General LLMs are great for research, but lack proprietary internal data.
Open source video
Source video

80/20 automation boundary

Benefit

Captures most of the time savings (days to minutes) while avoiding the cost of chasing full autonomy.

Cost or risk

The residual 10-20% requires expert attention; if the agent's uncertainty flags are imprecise, reviewers must redo the whole artifact, erasing the gain.

PMs achieve 80-90% automated completion and spend their time on final strategic refinement.
Open source video
Source video

Boots-on-the-ground customization vs. platform reuse

Benefit

Deep workflow immersion with domain experts yields accuracy in complex, regulated processes.

Cost or risk

Per-client workflow modeling is expensive and does not generalize; scaling requires templating the immersion output.

Financial institutions must build on the ground with domain experts to succeed.
Open source video
Source video

Open questions

4

How are conflicts resolved when internal verified sources disagree with each other, not just with external vendors?

Why unresolved

The summary describes internal data overriding generic/external information, but does not describe precedence among multiple internal systems.

Research direction

Prototype a per-field precedence and arbitration layer with recorded lineage and measure disagreement rates across internal systems of record.

Source video

How are the 80-90% completion and days-to-minutes figures measured — by artifact completeness, numeric accuracy versus a ground-truth portfolio, or PM acceptance without rework?

Why unresolved

Only headline productivity numbers are given; no evaluation methodology, sample size, or error taxonomy is described.

Research direction

Build a benchmark for document-grounded portfolio construction with field-level accuracy and rework-rate metrics.

Source video

What does the review surface look like so that a PM can audit a 200-400 name portfolio in minutes rather than re-deriving it?

Why unresolved

Human-in-the-loop is named as a component but its information design is not specified.

Research direction

Compare review interfaces (full diff, exception-only queue, provenance-annotated table) on time-to-signoff and missed-error rate.

Source video

Which parts of this architecture transfer outside structured credit, given the advantage that indentures are explicit and rules-based?

Why unresolved

The summary notes the rules-based nature of CLO documents but does not test domains lacking machine-checkable specifications.

Research direction

Measure automation rate as a function of specification explicitness across domains with and without a mechanical verifier.

Source video

Key claims

7
causalVerification needed

Portfolio construction time dropped from 2-3 days in manual Excel to minutes using the Rowspace agent platform.

Evidence

Portfolio construction went from taking 2-3 days in manual Excel to minutes using Rowspace.

Question

What was the measured baseline and post-deployment time across how many portfolios, and was the output quality equal or better?

Source video
opinionVerification needed

Agents deliver 80-90% of portfolio construction automatically, with humans doing final refinement.

Evidence

PMs achieve 80-90% automated completion and spend their time on final strategic refinement.

Question

Is 80-90% measured as artifact completeness, field accuracy, or fraction of work needing no rework?

Source video
factualVerification needed

For a given data point drawn from many sources, only one source is correct and internally verified.

Evidence

For a single data point drawn from 20 sources, only one source is correct and verified by internal systems.

Question

What fraction of fields actually have a single unambiguous correct source, and how are ambiguous fields handled?

Source video
causalVerification needed

Enterprise AI must sit on trusted internal data to override external or generic information.

Evidence

Rowspace sits on top of trusted internal data sources to override external generic information.

Question

Does internal-source precedence measurably reduce decision errors versus retrieval-only baselines?

Source video
comparativeVerification needed

General LLMs are suitable for research but insufficient for mission-critical decisions due to missing proprietary data.

Evidence

General LLMs are great for research, but lack proprietary internal data.

Question

On decision-grade tasks, what is the accuracy gap between a general model and the data-grounded platform?

Source video
causalVerification not requested

CLOs are rules-based with deep indenture documents, making them amenable to agent automation.

Evidence

CLOs are test-specific and rules-based with deep indenture documents.

Source video
causalVerification needed

Automated reconciliation turns multi-hour weekly manual trade-rec work into repeatable minute-level processes.

Evidence

Visualizing reports and automating reconciliation reduces multi-hour weekly chores to repeatable minute-level processes.

Question

How many reconciliation errors are caught by the automated path versus the prior manual bank-statement review?

Source video

Connections

4