A capital request arrives with a spreadsheet, a confident executive summary and a sponsor who would really like it approved. Finance checks the numbers, then asks an AI assistant to review the proposal. In this hypothetical, the assistant has everything it needs to calculate the return, but it also has several pages explaining why management should love the investment. The resulting recommendation may look wonderfully independent, right down to the sensible headings and restrained use of adjectives. Finance still needs to establish how much of that recommendation came from the evidence and how much came from the pitch.

In its September 10 report, CFA Institute tested nine OpenAI models across ten hypothetical investment scenarios, presenting equivalent information in positive and negative terms. The evaluations changed with the wording, particularly for drawdown performance and client retention. Those were investment-management exercises using specific model versions, so the findings don't establish how today's systems would handle a company's capital budget. They do suggest a useful test before finance starts treating AI-generated investment reviews as an independent second opinion.

The mitigation results were narrower still. In two categories tested with an August 2025 version of GPT-5 nano, manually presenting both perspectives reduced framing sensitivity more than the tested bias-detection workflow. Adding a definition of framing bias didn't significantly improve the result. The full report gives finance a reason to examine how information reaches the model, alongside what the model returns. A reminder to be objective doesn't establish that the review is objective, and the experiment didn't measure realized returns or losses from corporate investments.

Consider a PE-backed manufacturer evaluating a hypothetical $4 million equipment purchase. The sponsor expects lower scrap, greater throughput and fewer emergency repairs. Finance can put approved cash flows, implementation costs, ramp-up assumptions and residual value into a controlled calculation, then have the assistant investigate supporting evidence and challenge the proposal. Before that review, the analyst documents the source for each assumption and distinguishes confirmed orders from hoped-for demand. The assistant should apply the same review criteria to the expansion everyone likes and the maintenance project that has never once been described as transformational.

Then give the assistant equivalent versions of the same case in separate sessions, using the same model settings and review instructions. A vendor's stated 90% availability can also be described as 10% unavailability over the same period; both descriptions must preserve the same operating assumptions. Record whether the recommendation, risk assessment or requested evidence changes, and repeat the comparison enough times to distinguish a consistent reaction from ordinary output variation. Keep calculations and source documents fixed. Removing an actual guarantee or a relevant execution constraint would change the case itself. This is a proposed capital-review test, not a result demonstrated by the CFA study.

An analyst should examine the differences before the recommendation reaches the investment committee. If one version flags production disruption and another skips it, the reviewer can inspect the cited evidence, check whether the assistant misread a condition and require the unresolved question to be answered. NIST's generative AI risk profile discusses automation bias and cautions against extrapolating narrow evaluations to actual deployment. In this workflow, human review has a specific job: checking whether a recommendation follows from the approved facts. The investment decision still goes through the company's existing approval process, with the assumptions and unresolved concerns visible.

For recurring reviews, retain the input versions, model identifier, calculation outputs, cited evidence and the analyst's resolution. Test the assistant against past cases whose disputed assumptions finance understands, including projects that were rejected or revised. A second model can offer another review, but agreement between two assistants doesn't establish that either used the right facts. Track how often reviewers find unsupported assumptions, how much work it takes to resolve them and whether the recommendation changes when new evidence arrives. Include that effort in the cost of an accepted finance output, alongside model usage and repeat runs.

Next quarter's capital review could start with a small set of real proposals, using authorized data and a documented comparison before anyone relies on the assistant's recommendations. Across a portfolio, each company would retain its own permissions, hurdle rates and approval authority. The review team can use the exercise to improve proposal templates, tighten evidence requirements and identify which judgments the assistant handles inconsistently. Those findings belong in the maintained finance AI governance policy. As the model and workflow change, the team can rerun the cases to check whether rewriting the introduction changes the investment recommendation.