How to Apply AI Agents in Finance Like Apollo

AI agents in finance earn their keep at Apollo's Redding Ridge, where a mock CLO portfolio that took two to three days in Excel now takes minutes. Jack McDonald, Principal at Apollo and COO of Redding Ridge, described the shift with Michael Manapat, co-founder of Rowspace, at AI Engineer New York in October 2026.

Where AI Agents Became Mission-Critical at Apollo's CLO Manager

AI agents in finance became mission-critical at Redding Ridge, the structured credit and CLO manager owned by Apollo, in two specific workflows: building mock CLO portfolios and running weekly trade reconciliation. Jack McDonald, Principal at Apollo and COO of Redding Ridge, said the unit manages roughly $50 billion in assets and tracks about a thousand loan names across many fund entities. His comments came in an October 2026 fireside chat at AI Engineer New York with Michael Manapat, co-founder of Rowspace, a cloud platform that uses agents to do data engineering work for financial firms.

The problem Redding Ridge faced was data fragmentation. Private credit data does not trade like public equities, McDonald explained. A single data point can arrive from twenty different sources, and only one of them is correct. Information sits in indenture documents, legal agreements, trade systems, and bank statements. Before any automation could work, someone had to establish what the firm calls the golden source of truth for each calculation.

That framing matters for engineers evaluating similar projects. The value did not come from a general language model answering questions. It came from agents that reconcile, deduplicate, and route the firm's own information so portfolio managers and traders can act on it.

How AI Agents Cut CLO Portfolio Construction From Days to Minutes

Portfolio construction at Redding Ridge dropped from two to three days of manual Excel work to minutes, according to McDonald. A CLO is rules-based by design. Money flows through a defined waterfall, and compliance tests live deep inside indenture documents and multiple legal agreements.

Historically, a portfolio manager building a mock portfolio would read those documents, assemble an Excel model, and populate it with 200 to 400 individual credit names. McDonald said that process took two to three days per portfolio before the firm adopted the agent-based platform. Now the manager pulls current data, generates the portfolio in minutes, and spends human judgment on the final refinement.

The division of labor is the useful detail. The agent output reaches what McDonald called 80 to 90 percent completion. The portfolio manager then applies discretion to shape the last 10 percent and make the portfolio their own. Data flows in from document libraries, trade and compliance systems, and real-time pricing feeds. This is a supervised workflow, not an autonomous one, and McDonald presented it as a first-hand account from his own operation rather than an independently audited benchmark.

Automating a Weekly Trade Reconciliation That Took 2–3 People

A weekly trade reconciliation that consumed two to three employees for many hours each week now runs as an automated, repeatable report produced in about a minute. McDonald described the old process: Redding Ridge uses bank financing to buy loans in the fixed income market, and staff had to go into bank accounts manually, read statements to see where trades settled, reconcile buy and sell activity, and email a recap to management.

The new process produces the same report as a visualization, repeatable on demand. The team's role shifted from producing the report to quality-checking it each day. McDonald characterized the result as a routine, rinse-and-repeat process where human effort moved from assembly to review.

For engineers, the pattern generalizes: the highest-return automation target was a recurring, well-defined report with a known output format, not an open-ended reasoning task. The firm kept humans in the loop precisely because the report informs management decisions.

Why Plugging the Firm Into ChatGPT or Claude Not Enough

General-purpose assistants such as ChatGPT, OpenAI's assistant, and Claude, Anthropic AI assistant, are useful for research and orientation, McDonald said, but they work off general information and none of the firm's own data. For a firm making trading and portfolio decisions, that gap is disqualifying.

His reasoning breaks into three points:

• General models have no access to the firm's trade order management system, performance system, Bloomberg feeds, or pricing indices, so they cannot know which source is authoritative.

• One data point can come from twenty sources with only one correct value. Trust requires that someone validated the correct source and encoded that override logic into the system.

• Outputs feed decisions that reach more than 400 global investors. A plausible-sounding wrong number is worse than a slow human process, because errors propagate to the enterprise and the investor base.

The practical takeaway is that accuracy in this domain is a data engineering problem before it is a model problem. The model layer matters, but the moat is the reconciliation layer that decides which number is true.

Why Embedding Engineers With the Front Office Built Trust

Trust came from daily, in-person collaboration between the vendor's engineers and Redding Ridge's portfolio and finance teams. McDonald said the Rowspace team sat with his staff essentially every day, learning the information flow, understanding the calculations, and learning which data source to trust when sources conflict.

He contrasted this with other technology vendors the firm had worked with, which delivered products without that ground-level immersion. In his account, the embedded approach is why the Rowspace product ran ahead of its roadmap and expectations, and why the firm was willing to let agents handle work feeding investor-facing outputs.

His advice to anyone applying AI to a complex domain: what looks novel or out of place in a financial workflow is probably there for a reason. Financial organizations are nuanced and complicated, so spending time on the ground with the professionals who use the tool is, in his words, time well spent. This is speaker first-hand experience, not an independent study of implementation methods.

Is AI in Finance a Commodity or a Durable Competitive Advantage?

McDonald's position is that AI will be a durable competitive advantage for firms that adopt it properly, not a neutral utility like Excel. His logic: in actively traded markets, advantage comes from process and from having accurate information at your fingertips, including available liquidity and current pricing. Firms with validated, fast data can act before firms still reconciling spreadsheets.

He noted that private markets are still evolving toward that level of information availability, and that Apollo has been working at the front of that shift. In institutional finance broadly, he judged the adoption curve as very early. Current low-end use cases, such as research summarization and quick company briefings, he called lower value-add, though he expects reduced junior support needs in some functions over time.

A caveat on scope: this is one executive's assessment of his own firm and market, offered at an industry conference. It supports a claim about how one structured credit operation competes, not an industry-wide conclusion that AI deterministically separates winners from losers in asset management.

What Redding Ridge Hopes AI Changes Over the Next Few Years

Over the next few years, McDonald wants two things from AI at Redding Ridge: full confidence in information flowing to the investor base, and a team whose default mode is technology-enabled problem solving. With over 400 global investors, the firm cannot afford errors in reported data, and manual verification consumes hours he would rather eliminate.

He said the team has never lacked technical inclination, but technology use had been an afterthought. The goal is to flip that ordering, so that when staff encounter a process, their first questions are why it exists, whether it should change, and whether new technology improves it. Freed capacity from routine automation funds that thinking.

The two conditions for the technology to stay valuable, in his account, are sustained data accuracy and the discipline to keep humans reviewing outputs before they reach investors.

FAQ: AI Agents in Finance at Apollo and Redding Ridge

• How long did a CLO mock portfolio take before AI, and how long does it take now? Per Jack McDonald, building a mock portfolio of 200 to 400 credit names took a portfolio manager two to three days of manual Excel work. With the agent-based platform, the same build completes to 80–90 percent in minutes, and the manager refines the rest.

• Who is Redding Ridge? Redding Ridge is Apollo's structured credit and CLO manager, a separate operating company captive to Apollo that manages about $50 billion in assets. Jack McDonald is its COO and a Principal at Apollo; he joined Apollo in 2020.

• What is Rowspace's role in the workflow? Rowspace is a cloud platform, described by co-founder Michael Manapat, that uses AI agents to perform data engineering work: routing, reconciling, and unifying fragmented financial data so decision-makers see validated numbers from systems such as trade order management, performance, and Bloomberg pricing.

• Why not just use a general-purpose LLM? General models work from public knowledge and lack access to the firm's internal systems. Where one data point has twenty possible sources and only one correct value, the firm needs a validated reconciliation layer before any model output can support trading or investor reporting.

Turn Your Own Recorded Expertise Into a Written Article

The clearest lesson from this conversation is that durable value in AI comes from doing the unglamorous reconciliation work first, then letting agents repeat it. The same logic applies to your own knowledge: the valuable material in a conference talk, podcast, or webinar usually exists only as a video nobody can search or cite.

Skala Blog turns a YouTube video into a structured, publishable article. You paste the video URL, the tool transcribes it, and you get a written draft you can review and edit before publishing.

If you have interviews or explainers sitting in your video library, paste a YouTube URL at Skala Blog and turn the transcript into an article your audience can actually read and find.

Source video