When the Model Says No: Fiduciary Governance for AI Inside Leveraged Credit

The first time I watched a credit committee defer to a machine-learning output without naming who owned that decision, I knew the governance work hadn’t kept up with the technology.

The model flagged a covenant deterioration on a unitranche position six weeks before the analyst would have caught it. The committee accepted the flag. The risk officer initiated the workout conversation. Inside the room, this looked like progress. Outside the room — in the institution’s fiduciary record — there was no documentation of who had authority over the underlying override threshold, no recorded rationale for accepting versus contesting the model’s read, no escalation path defined for the next divergence. The capability was real. The posture around it was not.

That gap is the actual story of AI inside leveraged credit right now.

Models Concentrate Judgment. They Do Not Eliminate It.

Machine learning systems are now inside leveraged loan underwriting, covenant monitoring, distress flagging, and portfolio surveillance at scale — at the largest BDCs, at credit-focused private funds, inside CLO managers, and at the bank lenders that warehouse for them. The capability has matured faster than the institutional governance built to oversee it.

The most common mistake is the assumption that model deployment reduces fiduciary load. The opposite is true.

When underwriting was manual, judgment was distributed across dozens of small calls. Which comparable set to use. How to weight customer concentration. Whether to haircut an EBITDA bridge for a one-time working-capital release. Whether the management team’s revenue guidance was credible. Each of those decisions was recorded — sometimes as a model footnote, sometimes as a credit memo paragraph, sometimes as committee minutes — and each was reviewable after the fact.

When an ML system performs those same functions, the human decision migrates to a smaller number of higher-consequence choices. Do we accept the model’s output? Do we override it? How do we weight model conviction against the factors the model cannot see — a CEO transition the model doesn’t know is happening, a customer concentration the model is weighting from stale data, a regime shift the model has not been retrained against?

The decision surface narrows. The weight on each remaining decision grows. The fiduciary obligation does not move. It concentrates.

Most credit committees and boards are still structured for the old decision geometry — for distributed analyst judgment, periodic memo review, and exception-only escalation. They have not been re-architected for the regime where five percent of the decisions carry ninety percent of the fiduciary weight.

The Four Questions Every Credit Platform Should Answer in Writing

A credit platform using ML at scale should be able to put four answers on paper for its board, its LPs, and — when the next downturn produces the next dispute — its counsel.

What is the model predicting, and over what horizon. Default probability inside twelve months is a different model than covenant breach inside six months, and neither is the same as a recovery-rate prediction at workout. The institution that cannot distinguish between them in writing is not governing the model; it is consuming it.

What is the training data, and is the underlying regime still intact. A model trained on 2013 to 2022 behavior is operating against a different rate environment, a different covenant package, a different sponsor-friendly market, and — for leveraged credit specifically — a materially different distribution of loan-only structures versus traditional first-lien-second-lien stacks. Regime change does not invalidate the model. It does invalidate the assumption that yesterday’s calibration is still valid. The platform should know when the calibration last got challenged and by whom.

Who has authority to override the model, and under what documented rationale. If overrides are never recorded, the institution has surrendered part of its fiduciary record. If overrides are recorded but never reviewed, the institution has a paper trail with no governance function. The right posture is named override authority — typically the chief credit officer or a designated portfolio manager — paired with a documented rationale taxonomy and a periodic challenge cycle.

What is the escalation path when model output and human read diverge materially. That path should be defined before the divergence, not after. “Materially” is itself a definition that belongs in writing — at what conviction-delta does a divergence trigger a portfolio-level review, a workout-team consult, an LP disclosure consideration. The institutions that work this out in advance are the institutions that do not lose three weeks in the middle of a stress event arguing about whose call it is.

“The decision surface narrows. The weight on each remaining decision grows. The fiduciary obligation does not move. It concentrates.”

What Durable Governance Looks Like

The institutions that will navigate this period well are not the ones with the most sophisticated models. They are the ones whose boards, chief risk officers, and chief credit officers treat model governance as a first-class oversight function — with named ownership, documented authority, periodic challenge, and a clear written record of where judgment was exercised and why.

In practice, that looks like a quarterly model risk committee that reviews calibration drift, override frequency, and divergence incidents — not as a compliance report but as a substantive credit conversation. It looks like a board credit committee that has, in its charter, explicit oversight responsibility for AI and ML systems used in underwriting and surveillance. It looks like an investment committee memo template that includes a “model concurrence and override rationale” field for every flagged transaction. It looks like an LP report that says, in plain language, where models are deployed, what they govern, and how their outputs are challenged.

None of this is technically difficult. All of it is institutionally rare.

Why This Is Fiduciary Work, Not Technology Work

Models are useful. Fiduciary duty is non-delegable. The two statements belong in the same sentence in every credit platform’s governance documentation.

The reason this matters now — not in five years — is that the leveraged credit market has tightened around private credit at the same moment the AI tooling has matured. The fastest-growing pools of institutional capital in middle-market lending are also the pools with the least board-tested governance precedent around ML decisioning. Regulators are still calibrating their posture. LPs are still developing their diligence questions. The interval before that catches up is short.

At Pluribus Capital, model governance is a structured-finance question, a board-oversight question, and a special-situations question — not a separate technology workstream. The work of capital advisory across cycles is unchanged: name who owns the decision, document why, build the escalation path before the stress, and structure the close so the fiduciary record holds up to the second look.

Models will keep getting better. The institutions that get the governance right alongside them are the ones whose track records will compound. The institutions that don’t will find out the hard way which judgment they delegated and to whom.

That distinction is, and will remain, the work.