The hidden cost of verifying AI credit research

An AI-generated credit note can be quick to read and slow to use.

The summary says leverage improved, liquidity is adequate, and management remains committed to debt reduction. Before relying on it, the analyst still has questions. Which leverage measure? What supports the liquidity assessment? Is the statement about debt reduction a fresh commitment, an old target, or the LLM’s interpretation?

If answering those questions requires re-collecting the documents and rebuilding the analysis, the AI may impede the workflow rather than aid in speeding it up.

This has been one of the clearest lessons from our conversations with credit analysts while building Passu. Most want more help assembling the record rather than a single piece of output purporting to be a polished conclusion. That is not a request for less capable analysis; it is a request for work they can inspect, adapt, and stand behind.

Before using an AI-generated comparison in a credit memo, an analyst needs to check the periods, definitions, and assumptions behind it. If that requires redoing the work that precedes the output, introducing AI has saved less time than it appears to.

Finding the source is only the first check

Consider an illustrative comparison of two borrowers. An answer lists a leverage ratio for each and concludes that one has more room to absorb a weaker quarter. Both figures appear in the cited disclosures. But one is a reported ratio at the last quarter-end; the other is management’s target after a planned asset sale. The answer has compared a financial position with an intention.

There is no fabricated number to catch. The error is in what the answer assumed the numbers mean. A working citation will not reveal that unless the analyst can see the period, definition, and other surrounding context attached to each figure.

This is why checking whether a source exists is a weaker test than checking whether it supports the conclusion. Before comparing the borrowers, the reviewer needs to know what has happened, what is expected to happen, and whether the measures are comparable. A superficial link to the right document does not settle those questions by itself.

We heard a related point in another interview: being able to open the whole source document (with the exact relevant line highlighted) made an answer easier to assess than seeing an isolated excerpt. The surrounding table, heading, reconciliation, or other context could change how the passage should be read.

For AI-powered credit research, source access is therefore part of the work, not decoration at the end of an answer. It should help the analyst test whether the evidence supports this particular use of the information.

Make the judgment visible

Even after the source is correct, interpretation remains.

Consider the statement that management has become more committed to deleveraging. Supporting it requires choices about what counts as evidence. Is the assessment based on a revised target, a change in stated capital-allocation priorities, actual debt repayment, or simply stronger language on the latest call?

When we discussed tracking management credibility with practicing credit analysts, the more useful output was not a single rating. It was a history of what management had said and how it had changed: which priorities persisted, which targets moved, and when the language shifted. The analyst wanted to examine the pattern and form a view.

That feedback changes the design of the analysis. A credibility score compresses the evidence and the interpretation into one result. A dated comparison exposes the events on which that result would depend. It also gives the reviewer somewhere precise to disagree.

The same principle applies to a calculated credit metric. Choosing an EBITDA adjustment is not merely copying a fact. If the system applies that adjustment, the choice should be visible alongside the reported figure and the calculation. The analyst can then accept the source while rejecting the treatment.

Making these choices explicit is more useful than describing a specific output as objective. Research inevitably involves selection and interpretation. What matters is whether the analyst can see where those decisions enter the work.

Review should not require reconstruction

None of this means that analysts only want tables of facts. We have also heard clear demand for first-pass credit views when several new issues arrive at once, and for summaries that help decide where to spend time. An initial interpretation can be valuable precisely because the analyst has not yet completed the work.

The requirement is that the interpretation remains easy to challenge.

Take the opening example: leverage has improved and management appears focused on debt reduction. A reviewable answer would identify the leverage definition and periods, explain the movement using the relevant inputs, and show the dated statements supporting the assessment of management’s priorities. Any assumption about future cash generation would be identified as an assumption, not blended into the reported results.

The analyst still has substantive work to do. They may reject an adjustment, question whether the improvement is repeatable, or disagree about management’s willingness to prioritize creditors. But those are the decisions the research is supposed to support. They should not first have to work out which numbers and statements the system silently combined.

This is the standard our conversations have pushed us toward at Passu. The objective is not to remove review or to imply that a source-linked answer cannot be wrong. It is to make review more focused: expose the evidence, explain the assumptions, and identify the choices that deserve attention.

Faster generation is useful. The more consequential question is what happens after the answer arrives: can the analyst use the work, or do they have to do it again?

Explore the Passu platform.