The second quarter is the real test of AI for credit research

The first demonstration of an AI research tool often begins with a clean task: take a filing, ask a question, inspect the answer. It is a sensible way to establish basic usefulness. It is a poor way to test what happens over a year of coverage.

By the next quarter, the analyst has already settled definitions, chosen comparisons, rejected some adjustments, and identified issues to watch. A system can produce another excellent answer and still leave the analyst doing much of the same setup again.

This is one reason we find “specialized AI versus a general-purpose model” too narrow a framing. General-purpose models can be useful research tools. The more revealing question is what the surrounding workflow remembers, how it uses that context, and whether the analyst can correct it.

The second quarter exposes the difference between an answer that happens to be good and a research process that gets easier to maintain.

Keep the methodology; reassess the assumptions

In analyst conversations, we have encountered recurring requests to preserve calculation preferences, fit existing research formats, and work with models the team already uses. The underlying frustration is straightforward: once a choice has been made, the analyst should not have to explain it afresh in every interaction.

But not every prior choice should become a permanent instruction.

Take a hypothetical coverage workflow. The analyst uses a particular leverage definition and excludes a specified adjustment. Those are methodological decisions. The analyst also expects working capital to normalize in the second half. That is an expectation about the business. A system should not treat both as equally durable context.

The first may reasonably persist until the analyst changes it. The second should be tested constantly against new information. Remembering the expectation is useful because it tells the next update what to investigate; assuming it remains true defeats the purpose of monitoring.

This distinction is easy to lose when everything is stored as undifferentiated chat history. A long conversation is not, by itself, an orderly record of what the team believes, why it believes it, and which parts are still current.

Carry forward the questions, not just the conclusion

Analysts have described a related gap between having a current thesis and having a usable history of how that thesis developed. The latest company page may state the view. The reasoning behind a change may live in a separate initiation note, a model version, or a discussion that is difficult to retrieve.

The cost becomes apparent when someone asks a simple question: what would make us change our mind?

A useful research record should preserve the conditions attached to the view. Perhaps the company was expected to reduce debt after a transaction, restore cash conversion, or limit further acquisitions. Those conditions give subsequent research something specific to test. Without them, each update risks becoming another description of the business rather than an assessment of the argument.

Return to the working-capital example. If another quarter passes without the expected cash release, the update should put that shortfall next to the original forecast and management’s latest explanation. Repeating that cash conversion is “expected to improve” would simply push the assumption into another quarter. The analyst needs to judge whether the recovery is delayed, whether the business now requires more working capital, or whether the original forecast was wrong. Keeping the earlier reasoning on record makes that reassessment possible.

A change needs a history

Suppose management revises a medium-term target. A useful update should not simply replace the old target wherever it appears. The change itself may matter.

The research record needs the previous statement, the new statement, their dates, and enough context to understand the revision. Otherwise, the system can leave every document looking current while erasing the very development the analyst needed to notice.

The same principle applies to methodology. A change in reported segmentation or disclosure may require a new treatment. Carrying forward the old approach without warning is not consistency; it is an undisclosed assumption. The workflow should surface the break in comparability and let the analyst decide how to handle it.

These are requirements for the research process, not claims that any model can resolve them automatically. They need deliberate choices about what is retained, versioned, reviewed, and superseded.

Test continuity with a handoff

A practical way to assess this is to revisit an issuer after a meaningful update, or ask another analyst to pick up the work.

Can that person identify the standing view, the important definitions, the unresolved questions, and the changes since the last review without reading an entire chat history? Can they distinguish a reported fact from an assumption that a colleague accepted? Can they revise one decision without rebuilding the rest of the analysis?

This is a harder test than asking whether the system remembers a name or reproduces a previous answer. It tests whether the retained context is organized well enough to be useful.

Our conversations have pushed us toward this understanding of credit-specific AI: not merely different vocabulary or a more elaborate prompt, but support for research that continues over time. The first answer should be good. The next update should build on the work while remaining willing to change the view.