Inside Apollo, the credit manager that runs the firm’s structured credit and CLO business has brought AI agents, software that carries out multistep tasks on its own, into two core jobs. A portfolio model that once took a portfolio manager two to three days to build in Excel now arrives as an 80% to 90% complete draft within minutes. A weekly trade reconciliation that tied up two to three staff members for hours now produces a report in about one minute. The speed is the headline, but the more useful lesson sits underneath it: in a business where one figure can be recorded in 20 places and only one of them is right, the automation appears to have worked because the team first settled which data to trust.
Why credit data resists automation
A CLO, or collateralized loan obligation, is a security built from a pool of corporate loans. Apollo’s in-house CLO manager oversees roughly $50 billion in assets. It runs dozens of fund entities and thousands of named loan pools in the broadly syndicated loan market, the market for large corporate loans shared among many lenders, and it takes in thousands of new data points every day.
Stocks and publicly traded bonds come with standardized feeds from exchanges. Private credit and structured loans do not. Their data sits in privately negotiated documents and custom agreements: indentures, the legal contracts that govern each deal, along with trade logs and bank statements. CLO management is also highly active, so evaluating risk and executing trades safely depends on real-time data that the team can rely on.
To close that gap, the manager partnered with a startup that builds an agentic AI platform for financial firms. The startup’s agents automate data engineering, pull figures out of documents, and reconcile numbers across systems. The goal was to extract the critical calculations buried in complex loan contracts and establish a single source of truth, one authoritative version of each figure that the whole business works from.
Two workflows, before and after
Building a CLO portfolio
Constructing a CLO portfolio is a heavily rule-bound job. Every portfolio has to satisfy its legal indenture and a predefined cash-flow waterfall, the order in which incoming cash is paid out. Portfolio managers used to read through legal documents running to hundreds of pages, then spend two to three days building complex Excel models covering 200 to 400 individual credit names.
Now the agents gather real-time credit data and assemble an initial portfolio model within minutes, reaching 80% to 90% completion. Portfolio managers apply their own judgment and expertise to the remaining 10%. Because the first draft no longer takes days, each round of revising and rechecking a portfolio becomes far shorter.
Reconciling warehouse trades
Before loans are packaged into a CLO and sold as securities, the manager buys and accumulates them using bank credit lines known as warehouse facilities. Each week, someone has to confirm that those purchases settled correctly.
That check used to take two to three staff members several hours. They logged into bank accounts one by one, matched trade settlements against bank statements, and wrote up email summaries for management. Today the system ingests the bank records and produces a structured visual report in about one minute. The staff’s job has shifted to daily quality assurance and exception checks.
| Workflow | Before | With AI agents | What people still do |
|---|---|---|---|
| CLO portfolio construction | Two to three days of manual Excel modeling across 200 to 400 credit names | An 80% to 90% complete draft in minutes | Judge and refine the final 10% |
| Weekly warehouse trade reconciliation | Two to three staff spending hours checking bank records and writing emails | A structured visual report in about one minute | Daily quality checks and exception reviews |
In both cases the agents do not take over the job outright. They produce most of the work as a draft or a report, and the people involved move into the roles of judge and checker.
▲ AI draft with a human finishing touch
Why off-the-shelf AI models fall short
An obvious question is why a firm cannot simply drop its documents into a public large language model (LLM) such as ChatGPT or Claude. The credit manager’s chief operating officer draws a clear line. General models trained on internet data are useful for broad research and for exploring concepts. They cannot run proprietary credit workflows, because they have no access to the firm’s internal books, its custom indentures, or its validated internal balances. Trading and portfolio decisions rest on verified, proprietary calculations, so there is no room to trade away precision.
The data itself adds a second problem. In private credit, a single figure may be recorded in 20 different sources inside and outside the firm, and only one of them is the true, definitive number. Picking the right one requires systematic rules. The startup’s platform harmonizes data from order management systems, performance engines, and outside market pricing providers such as Bloomberg into one automated system.
Accuracy matters this much because a wrong number can cascade through an organization. Apollo serves more than 400 institutional investors worldwide, so spotless data is essential to managing risk and protecting capital. A top priority for the COO over the coming months and quarters is investor reporting the team can trust without hours of manual reconciliation. That means systems that guarantee zero discrepancies and zero hallucinations, the plausible but false output AI models sometimes produce, before any figure reaches an investor. When the data pipeline is reliable, investment committees can act quickly with full confidence in the numbers underneath.
Trust gets built on the operations floor
Who decides which source wins and how each figure is calculated? In this partnership, the startup sent its engineers directly onto the trading and operations floor to work alongside portfolio managers, finance staff, and risk teams. The COO credits the collaboration’s success to the startup’s CEO and engineers sitting with Apollo staff day in and day out.
His advice to anyone building AI for finance is to put boots on the ground. To outside engineers, financial workflows often look strange, unconventional, or redundant. Usually, though, those steps exist for subtle legal, regulatory, or structural reasons. Rather than assuming an odd process is inefficient, builders need to watch front-office professionals do their daily work and learn why each step is there.
▲ Engineers embedded with finance teams
Finance is still in the early innings
Software development, law, and customer service have already changed sharply because of AI. Institutional finance, in the COO’s assessment, is still early. Large financial firms run huge middle- and back-office operations full of procedural friction and structural complexity, and the penalty for a data error is severe.
Many current uses in finance sit at the front end of the work: drafting financial models, compiling background research, and profiling companies. He sees these summarizing and aggregating tasks as relatively low-value next to the long-term potential of transforming entire workflows from start to finish. Over time, he expects routine support work traditionally given to junior analysts to shrink, with people moving toward oversight and creative deal structuring.
Apollo set this direction early. Back when AI coding tools such as Cursor were still at the tab-completion stage, the firm concluded that it needed to rebuild itself on AI-native enterprise platforms, systems designed for AI from the ground up, instead of stitching together disconnected single-purpose tools. Its leadership chose to embrace the technology rather than discount or resist it, and the firm’s flat, open culture pushes employees to keep upskilling. A key goal now is making investment teams tech-enabled by instinct, so that people routinely ask why a legacy workflow exists and whether it can be automated.
Will AI in finance end up as a baseline tool that everyone has, the way Microsoft Excel did? The COO expects firms with a well-designed, tailored AI strategy to build significant and lasting competitive advantages. He sees the biggest opening in private markets, which remain opaque and complex, while public markets have already automated much of their pricing and liquidity data.
What to take into your own high-stakes work
The pattern behind this case seems to be that AI agents earn a place in accuracy-critical work when they produce fast drafts, while people keep control of which numbers count and of the final call. That lesson likely applies well beyond credit markets. If an error in your work carries real cost, a few principles follow from the case:
- Start with the repetitive heavy lifting. Document extraction, first-draft models, report drafting, and other back-office chores are the natural first targets.
- Name the authoritative source. When the same figure lives in several places, decide with the people who use it which source counts and how the number is calculated.
- Validate before you trust. Use general-purpose LLMs for research and exploration, and route proprietary calculations and any figure that leaves the company through a validation step that catches discrepancies and hallucinations.
- Keep humans on the last stretch. Even when AI drafts 80% to 90% of the work, people should own the final 10% to 20% of decisions and the sign-off.
- Turn data entry into oversight. Move staff from typing and report writing toward daily quality assurance and exception handling.
- Build next to the practitioners. If you are building the tool, sit with the people who do the work and find out why the strange steps exist before removing them.
The broader takeaway appears to be that a platform with verified data and clear checks creates a more durable edge than adding one more single-purpose tool.