AI
Apr 30, 2026
4 minutes
Every COO and CFO at a private credit fund is fielding the same question right now: which AI should we use?
GPT-4 or Claude? Build something internally or wait for a vendor to package it? Hire a consultant or let the technology mature? The debate has become a fixture of every LP meeting, every ops offsite, every conversation with a portfolio company that just hired a "Head of AI."
It's the wrong debate.
Not because the choice of model doesn't matter -it does, at the margin. But because the model is, increasingly, the easy part. It's the part you can buy off a shelf, swap out in an afternoon, and replace next year when something better comes along. The hard part, the part that actually determines whether an AI agent transforms your loan servicing operation or becomes a very expensive disappointment, barely gets talked about.
It's called the harness.
Think about the difference between a junior analyst who has been handed access to ChatGPT and a senior ops manager who has spent three years learning your loan book, your lender relationships, your workflows, and your approval chains. Same underlying intelligence. Completely different outcomes.
That gap - the context, the memory, the domain knowledge, the judgment about when to act and when to escalate - is what a harness does for an AI agent.
The model is the engine. The harness is everything that makes it useful in your specific environment.
The equation is simple: Agent = Model + Harness.
The model is a commodity. The harness is the IP.
Here's why this matters more than most people realize: a language model, in its purest form, is stateless. It has no memory of what came before. Every session starts from zero. All of the capabilities we associate with useful AI agents: remembering context, carrying knowledge across sessions, executing multi-step workflows, connecting to your systems, none of that comes from the model. It all comes from the harness built around it.
Context management. The model can only see what's fed to it at any given moment, and that window has limits. Feed it too little and it gets things wrong. Feed it too much and it gets confused. Engineering the right information density, giving the agent exactly what it needs to act correctly, at the right moment, is harder than it sounds. When a rate notice arrives from a lender, does the agent see just the email? Or the full loan record, the syndication structure, the relevant covenants, the history of previous notices on that facility?
Memory. And it's distinct from context. Context is what the agent sees in this session. Memory is what it knows across sessions. A well-built harness allows the agent to retain and retrieve information over time: that this borrower tends to request prepayment at quarter-end, that this lender requires 72 hours for syndication approvals, that the last rate change on this facility had a specific exception. Here's the important distinction: the model can decide what's worth remembering. But the harness is what actually stores and retrieves it.
Tool connectivity. Can the agent read incoming notices from your inbox (Outlook), look up the relevant loan record (LMS), calculate the amended payment schedule, push a notification to the right person on your team(Slack)? MCP (Model Context Protocol) is the current standard for connecting agents to the real systems they need to operate in. Without it, you have a very smart chatbot. With it, you have something that can actually run a workflow.
Human-in-the-loop design. In private credit, almost no decision should be made without a human sign-off somewhere in the chain. The harness determines where those checkpoints live, and critically, what the audit trail looks like when they're used, making agents defensible to your LPs and regulators.
There's a pattern in how AI projects fail in this industry that's worth naming: it's almost never the model that's the problem.
The model is usually fine. What's missing is the harness. Generic AI tools (even very capable ones) don't know what a rate notice is. They don't know that a prepayment request triggers a multi-step workflow across every syndicated entity on a facility. They don't know the difference between a funded tranche and a committed-but-undisbursed revolver, or why that distinction matters for how a payoff figure should be calculated.
This is also why hallucinations are more predictable than most people realize. Most times, hallucinations aren't random malfunctions, they're a context problem. The model wants to give you an answer. When it doesn't have access to the specific data it needs, it reasons from what it does know and produces something plausible. Your private credit portfolio data isn't embedded in any public language model. If the harness doesn't provide it, through tools, through memory, through real-time data access, the model will fill the gap with something that sounds right. Which in a financial operations context is a real problem.
The fix isn't a better model. It's a harness that gives the agent access to the right data, at the right moment, with the right tools to retrieve it when it doesn't already have it.
There's a version of this that a general-purpose AI tool can approximate. Connect Claude or ChatGPT to some of your systems, prompt it carefully, and most times you'll get something something good.
However, although a generalized harness can handle well-defined, self-contained tasks, it struggles when the workflow is multi-step, multi-party, time-sensitive, and touches multiple systems simultaneously. Which describes most of private credit post-close operations.
Specialized harnessing solves for the specific domain: the vocabulary, the workflow logic, the approval chains, the data model, the failure modes. It knows what to do when a lender doesn't respond before a deadline. It knows which approval chain applies to which type of change. It knows when to pause and when to proceed. The generalized vs. specialized harnessing is the gap that determines which funds will ultimately build a real operational advantage from AI.
The questions worth asking when you assess any AI tool for loan servicing:
A purpose-built harness will have answers that are specific, testable, and grounded in your actual workflows.
Private credit is early enough in this cycle that the specialized harness layer is still largely unbuilt. Most funds are using general-purpose AI for ad-hoc analysis or are in a holding pattern. A small number are starting to build their own — and discovering how much harder it is than the first few demos suggested.
The funds that move first on specialized operational infrastructure will have an advantage that compounds. Not because they picked the right model (the model will keep getting better and cheaper regardless), but because they built or adopted the right harness, trained it on the right context, and gave it the controls that make it trustworthy at scale.
The model wars will keep running. That's not where the durable advantage is going to be built.
The harness is.
