What should change when a credit team adopts AI?

Imagine an analyst testing an AI platform during earnings season. The platform prepares a useful update. The analyst checks it, makes a few corrections, then returns to the workbook and research note they were already maintaining.

The analyst has gained a useful second view, but still has the same work to finish.

Running both processes is reasonable while the team establishes trust. But if the analyst will always have to complete the original work and review a separate AI output, even capable software may struggle to justify its place.

That is why adopting AI requires a willingness to reconsider the process around it. The firm and the provider should agree on which work could move into the platform, what would need to connect, and how they will judge the result.

Give the pilot something it can change

Take that earnings update. The work begins with obtaining the reporting package and ends when the analyst has updated the model and research record and briefed the portfolio manager. Generating a note is only one step.

A meaningful pilot follows the assignment through. Can the platform use the team’s metric definitions, prepare changes in the workbook analysts actually maintain, and carry unresolved questions into the next review? If it can, the team should be able to stop assembling some of that material separately.

This requires cooperation beyond the analyst trying the product. Documents need to be accessible, permissions need to be clear, and someone must be able to approve changes to the workflow. Asking analysts to improvise around missing connections can consume the time the platform was supposed to save.

The team does not need to rebuild its entire infrastructure for a trial. It does need an owner with the authority to test a different way of completing one recurring assignment. Agree on the quality and review standards that would allow the old steps to stop. If the provider cannot meet them, the team has learned something useful without committing to a wider rollout.

What does the team do with the time?

Suppose the earnings workflow works. Analysts spend less time assembling updates and can examine credits they previously left outside active coverage. The next question is whether that additional attention produces research worth acting on.

For a credit team, one useful measure is the number of qualified investment ideas reaching analyst or PM review before and after adoption. Agree beforehand what qualifies: a distinct opportunity that fits the mandate and has enough supporting work to warrant consideration. A longer list of AI-generated names tells you little.

Look at the coverage behind those ideas, too. Are more issuers being followed with current models and documented views? Can the team revisit an opportunity quickly when pricing changes? Record useful risk discoveries and well-supported rejections as well as prospective investments. Compare similar coverage universes and account for changes in staffing and market activity.

This takes more than one quarter to judge. One busy issuance period cannot establish that the team has sustainably expanded its research capacity, much less improved returns.

Cost savings may arrive later. An existing subscription may need to run until renewal, and the team may need both systems during the transition. Identify which expenses the platform will replace and when they can realistically be removed.

Cash savings may take six to twelve months to materialize where implementation and contract renewals delay them. Include platform fees, integration, support, and review in the calculation. A longer measurement period should follow an initial pilot that meets its standards, with scheduled decisions about continuing—not postpone the decision on software that still does not work.

Choose a provider that can support the new process

Once the platform becomes part of daily coverage, the firm depends on more than its answers. It depends on the sources it can use, the connections it maintains, and the cost of running the workload.

That makes optionality a practical concern. The team may want to keep its preferred data providers, use licensed expert-network material where AI processing is permitted, or change approved models later. Test those choices during the pilot. For a missing integration, establish whether the provider will build it, by when, who pays, and who maintains it. The firm’s own research and methodology should also be exportable in a usable form, subject to underlying data rights.

Pricing should support the same freedom to use the product properly. A small token allowance can make a trial inexpensive while leaving the cost of full coverage unclear. Ask for an all-in fixed price for the agreed workflows and volume, including earnings-season peaks.

An all-in fixed price gives the provider a reason to complete the work accurately and efficiently. Corrections illustrate why: under token-based pricing, repeated attempts can generate more revenue; under a fixed price, the provider bears that cost. The provider should absorb the cost of correcting its own failures. Check for throttling, model restrictions, and separate data or integration charges before treating a quote as all-in. If pricing is outcome-based, define the deliverable and quality standard; investment returns are not an outcome the vendor can control.

Return to the analyst testing an earnings update. By the end of a useful pilot, they should know which work they can hand over, how they will review it, and what they can now cover that they could not before. Before committing, the firm should know what it will stop paying for, what the new workflow will cost, and whether it can keep its preferred data sources and change models later. Set a review date to check whether the expected coverage gains and savings have materialized.