We can stop asking whether artificial intelligence works in municipal credit. It works.
I watched the question change in real time at The Bond Buyer Tech Forum. No one was debating if AI belongs in credit research anymore, only how to use it well. That is the more useful conversation, and the harder one, because deploying AI in credit is less a technology problem than a management problem.
The best mental model is also the most familiar one in our business. Treat AI like the best junior analyst you have ever hired, and manage it like one. Start with what that junior analyst already does today, because it is concrete and it is not hype.
The biggest change is the death of swiveling. Analysts used to spend most of the day finding and formatting information rather than judging it: Control+F across hundreds of PDFs.
Picture the haystack: The market produces roughly 9,000 new issues a year and on the order of 140,000 continuing disclosures, most of them financial filings, and that's before tapping into a single news story or alternative data like climate risk and demographic analysis.
No human team reads all of that. A model can, and the use cases are real and in production: pulling structured fields out of official statements and audited financials, triaging a 300-page document down to what matters, scanning a portfolio for material events, and drafting a first-pass credit memo an analyst then sharpens. This is the unglamorous, highest-return work.
The frontier is moving from answering a question to running a task, which is what people mean by agentic AI, and it is where the junior-analyst metaphor fits best. An agent can ingest a new filing, extract the figures, compare them to the prior year, flag a covenant or fund-balance change, draft the surveillance note, and route exceptions to a person.
Some tech-forward firms in credit are already piloting exactly this.
The opportunity is real: coverage per analyst several times higher, and a book is watched continuously rather than at the next scheduled review.
Here is where the metaphor earns its keep: You would never let a brilliant new hire sign off on a credit unsupervised, and agentic is not the same as autonomous. An agent in credit is a junior analyst that never sleeps. It does the leg work and escalates at defined checkpoints. The design question is not whether it can run end to end, but where you place the human checkpoints so you get the speed without owning a hallucination. The stakes climb as the work climbs.
An extraction error is a typo you catch on the next read; an agentic error is a decision already made and acted on. And the validation problem underneath is the one almost no one talks about: in credit, you do not learn whether a model's read was right for years, until a default or rating migration actually happens. So the answer cannot be blind trust in an output. It has to be designed checkpoints, source citations, and a human who owns the call.
That is why governance belongs here, and why we get it wrong when we treat it as a wall. In a regulated market, responsible AI is a product requirement, not a compliance afterthought. It is use-case inventories, risk tiers, human-in-the-loop thresholds, logging, and a rollback path. But the goal is to enable responsible experimentation, not to block it.
Governance should be a speed bump that forces thoughtfulness, not a roadblock that kills the experiment before it proves value. Risk-aware innovation is the whole game, and firms that treat those two words as opposites tend to do neither well.
A good manager also knows you do not deploy the same hire the same way on every desk. There is no one-size-fits-all here, and a market this variable punishes the pretense that there is.
Where credits are data-rich and homogeneous, AI's win is scale: monitor and triage a huge book continuously and surface the rare deterioration. Where credits are idiosyncratic and data-thin, the win is augmentation, not automation in which we surface the buried covenant and
the local-news signal but let the human do far more of the judging.
The rule of thumb is simple: The thinner the data and the higher the stakes, the higher you set the human checkpoint. Choosing how to deploy AI in credit is, in the end, an act of credit judgment itself and the discipline is the one we already practice.
None of this comes at the expense of the craft; it enhances it. Roughly 70% of credit work is repeatable and standardizable, and that is the part the technology is rapidly absorbing. The other 30% is the art of credit: the political context, the willingness to pay as distinct from the ability to pay, the novel security structure, the call where the data is thin and the model is confidently wrong.
AI can draft the first version of credit work. It cannot be accountable for being wrong about it. The analyst who learns to direct AI covers many times the credits and grows more valuable, not less. We are not automating judgment out of this market.
Managed well, the new hire finally gives our best people the time to do all they want and need to do.
