From 3 Days to Minutes: Apollo Rebuilds Portfolio Construction with AI Agents

In the world of financial investment, the tension between fragmented data and decision-making speed has long persisted. The partnership between Apollo's credit platform Reading Ridge and AI agent platform Rowspace demonstrates a path from "AI as an option" to "AI as mission-critical." According to on-site notes from a Bilibili creator at the AI Engineer conference, the two parties compressed portfolio construction work that previously took days into just minutes, providing a reusable blueprint for applying AI to core financial business.

From Days to Minutes: How AI Agents Reconstruct Portfolio Construction

Reading Ridge manages roughly $50 billion in assets, with its core business centered on CLOs (collateralized loan obligations). CLO products are rule-intensive—cash flows are strictly governed by in-depth indenture documents and multiple legal agreements. Portfolio managers must select from 200 to 400 individual credit names to build a portfolio and ensure compliance with complex waterfall distribution rules.

Under the traditional process, PMs had to manually review documents, build Excel models, and repeatedly validate data. A portfolio typically took two to three days from scratch to completion. Rowspace's AI agents take over the data engineering layer—extracting and reconciling information from document repositories, trade compliance systems, and real-time pricing sources, and automatically generating portfolio frameworks. According to Jack McDonald, PMs can now get a high-completeness base portfolio within minutes and then focus their energy on final-stage judgment and fine-tuning—"making the portfolio their own."

Another typical scenario is warehouse purchase activity. Reading Ridge relies on financing from multiple banks, buying and selling loans at any time in the fixed-income market. In the past, the team had to manually log into each bank account every week to reconcile settlement statements, with two to three people spending hours compiling report emails. Now, the same report only requires importing data into Rowspace, and a reusable visual report can be generated within a minute. The team's work shifts from "production" to "quality control."

CLOs (Collateralized Loan Obligations) and Waterfall Distribution Rules

A CLO (Collateralized Loan Obligation) is a structured financial product that pools large numbers of corporate loans and issues securities in tranches. Its core mechanism is the "waterfall": interest and principal generated by underlying loans are paid to different tranches of security holders in priority order—first AAA, then downward in sequence, and finally the equity tranche. Each tranche's payment conditions, trigger thresholds, and coverage tests (such as overcollateralization tests) are strictly defined by indenture documents—change one part and the whole structure is affected. Therefore, CLO portfolio construction is not just about selecting credit assets; it must ensure structural compliance under hundreds of rules. If a single data point or calculation logic is wrong, the entire waterfall's cash flow distribution will deviate, affecting returns for all investors from senior to equity tranches. This is precisely why Reading Ridge needs an AI system to ensure a traceable, auditable "golden source"—and the key reason general-purpose large models struggle with such tasks.

Why Can't You Just Throw Data into ChatGPT?

A natural question: if large models are so capable, why not just paste data into ChatGPT or Claude? Jack gave a clear answer in the conversation: general-purpose large models are trained on public information and excel at rapid research and building cognitive frameworks, but they do not own the enterprise's own data. In businesses like CLOs, a single data point may come from 20 different sources, only one of which is correct.

Rowspace's differentiation lies in "owning its own data" and building a verifiable technical layer on top of it. The team adopts an "embedded" service model—Michael and his colleagues sit with Reading Ridge's portfolio team and finance team every day, understanding each calculation logic, confirming which data source is the "golden source," and setting override rules in the system. This deep involvement makes the system's output traceable and auditable, giving the team confidence to make trading decisions based on it.

Jack particularly emphasizes that what sets Rowspace apart from other technology vendors is "boots on the ground"—not delivering a piece of software and leaving, but continuously being present to learn the information flow. Although this way of working is hard to scale and replicate, it is precisely the precondition for trusting AI in core financial business.

"Golden Source" and Data Override Rules

In financial data engineering, "golden source" refers to the source that is recognized as authoritative and correct for a given data point among many possible sources. For example, the current balance of a loan may appear simultaneously in custodian bank statements, internal trading systems, third-party data vendors, and email confirmations—but only one is the final settlement basis. Establishing the golden source requires business experts to confirm each item and set "override rules"—when multiple sources conflict, the system automatically selects the correct value according to preset priority. Rowspace's team sits with Reading Ridge's business team every day precisely to map each metric's data lineage, confirm authoritative sources, and solidify override logic. This mechanism makes system output auditable and traceable, and is the trust foundation for AI to move from "assisting research" to "supporting trading decisions." Without a golden source layer, AI output is just another data source requiring manual review.

Apollo's AI-native DNA: Culture Matters More Than Technology

According to host Mark Mirchandani, Apollo had already established a路线 of "rebuilding the company in an AI-native way" before Claude Code and Cursor became widespread, and insisted on platformization rather than point solutions. Jack attributes this foresight to Apollo's culture: as a leading institution in alternative asset management, management has consistently emphasized innovation, flat hierarchy, and rapid adaptation to market changes.

This culture manifests in organizational behavior as an open attitude toward new technology—"don't reject, adopt first." AI is introduced into workflows not to replace judgment but to free up human capacity for work requiring deeper thinking. Jack notes that the company has many middle- and back-office support functions with significant efficiency improvement potential, and AI can take over routine tasks within them.

However, Jack maintains a clear-eyed view of AI's overall progress in finance. He believes that by the standard of "AI truly transforming the industry's form," finance is still in its early stages. Current large models are indeed valuable in scenarios like rapid research and quick fundamental overviews of companies, but such applications have relatively low added value. The truly difficult part is embedding into core processes such as trading, compliance, and valuation, which requires technology teams to understand business details far beyond the surface.

AI-native and Platform Strategy

"AI-native" means rebuilding from organizational design, workflows, and technical architecture with AI capabilities as a precondition, rather than layering AI tools on top of existing processes. The opposite is "AI-enabled"—introducing AI only in local areas to improve efficiency. Apollo chose the former, emphasizing "platformization" rather than point solutions: rather than purchasing a separate AI tool for each department, build a unified data and agent platform so different teams can develop applications on the same foundation. The advantage of platformization is unified data standards, reusable permissions and auditing, and capabilities that can migrate across teams; the cost is large upfront investment and slow returns, requiring long-term management commitment. Jack attributes this choice to Apollo's innovation culture and flat decision-making mechanism, which also explains why the company could establish an AI-native path before Claude Code and Cursor became widespread.

Is AI Excel or a Moat?

The conversation raised a sharp question: will AI ultimately become a general-purpose productivity tool available to everyone, like Excel, or can it form a structural competitive advantage? Jack's answer leans toward the latter, but with conditions.

In actively traded markets, competitiveness depends on process rigor and information timeliness—knowing where available liquidity is, how assets are priced in real time, and what they are worth. If AI can ensure such information is accurate, consistent, and not fragmented, it can directly translate into trading advantage. In private markets, the same logic holds but evolves more slowly. Jack points out that as firms gain better information access and teams have full confidence in data accuracy, the lead brought by AI will gradually accumulate into competitive barriers. The downside risk is equally clear: decisions made based on incorrect information will propagate along the chain to institutions and investors—"that's the last thing you want to happen."

For the future, Jack's expectations center on two points. First, complete confidence in output information—Reading Ridge serves more than 400 global investors, and any data error could have wide impact. AI systems must free teams from "spending hours on manual review." Second, a shift in team mindset: technology should no longer be an afterthought but the starting point for how employees think about problems. Once routine tasks are automated, human capacity can turn to asking "why does this process exist, and how should it be improved?"

Advice for AI + Finance Practitioners

At the end of the conversation, Jack offered a plain but crucial piece of advice for those hoping to enter the AI + finance intersection: financial institutions are extremely complex, and many seemingly odd or redundant process arrangements often have specific compliance or business reasons behind them. Spending time truly sitting beside business teams and understanding their workflows and judgment logic is the dividing line for whether any tool can be deployed. The fact that Rowspace's product could be delivered ahead of its roadmap largely stems from this day-after-day presence.

For technology teams, finance is not a field that can be conquered with general-purpose capabilities alone; for finance teams, AI is not plug-and-play magic. Deep collaboration between both sides is the real reason days of work can be compressed into minutes.

One-line Summary

Apollo's Reading Ridge used Rowspace's AI agents to compress CLO portfolio construction from days to minutes, revealing that the key to deploying AI in core financial business lies in deeply embedded understanding of data and processes.

Paragraph Summary

Apollo's credit platform Reading Ridge manages about $50 billion in assets, and its core CLO business is rule-intensive with complex data sources. Through partnership with Rowspace's AI agents, portfolio construction work that previously took two to three days—involving manual document review and Excel modeling—has been compressed to minutes, with PMs only needing to complete final judgment and fine-tuning. The article notes that general-purpose large models cannot replace an enterprise's own data platform; Rowspace establishes an auditable "golden source" layer by being on-site daily to understand data sources and calculation logic. Apollo's AI-native culture and platform strategy put it ahead, but finance overall remains in early stages. Whether AI becomes a moat depends on information accuracy and process rigor, and the key to deployment lies in technology teams deeply understanding business complexity.