The American capital market consulting firm Chatham Financial recently provided a more concrete figure than “all employees are using AI”: in an early measurement of transaction verification, the manual review that originally took about 30 minutes was reduced to less than 4 minutes with the help of an application developed by Codex. This result does not mean that financial transactions are now automatically approved by models. The company also emphasized that the application is comparing real transactions with the results of senior reviewers on a case-by-case basis, and expanding the scope of automation is still the next step in their plan. For financial institutions, this limitation is even more important than the speed itself.
Transaction verification may seem like a back-end process, but in reality, it involves answering three key questions: what authorization has the customer granted, what actually happened during the transaction, and what has the system recorded. If there is any misalignment in any of these aspects, subsequent settlements, risk exposures, and customer reports can all be affected by errors. Traditional processes rely on professionals to repeatedly search for evidence between orders, confirmations, and internal records. The approach adopted by Chatham is not to let models “judge right or wrong” based on a piece of text; instead, it breaks down the tasks of collecting evidence, comparing key terms, and identifying discrepancies into verifiable steps, leaving any anomalies to be handled by humans.
What is truly saved is the time before evidence is found.
In the cases disclosed on October 2nd by OpenAI and Chatham, what is most noteworthy is not which generation of model was used, but rather the redesign of the work process. Chatham refers to this approach as Process Zero: first determine what outcome the work should produce, then list the minimum evidence required for that outcome, and finally clarify which judgments must be undertaken by professionals. AI is assigned to steps that can be structurally compared and retrospectively verified, while employees are responsible for dealing with exceptions, explanations, and ultimate responsibilities. Such a division of labor is not the same as simply throwing an entire transaction into a chat window and asking if it is correct.
The company stated that the application will gather materials supporting the transactions, compare key terms, and highlight any inconsistencies. If there are differences in the maturity dates, notional principal, or cash flow arrangements of an interest rate hedging transaction across different systems, reviewers need to first identify where these discrepancies occur and which original documents they are based on, rather than simply relying on a model conclusion that reads "risks exist." The increased speed is only valuable when the chain of evidence is complete; otherwise, the review time may merely be shifted from the operators to subsequent audit and error correction processes. Chatham has not disclosed the sample size, the error rates for various types of transactions, or the stability across products, therefore it cannot be generalized that "less than 4 minutes" represents the common performance for all transactions.
The company also integrates its employees' self-built applications into the internal platform Chatham Vibes. Disclosure materials show that the AI function in these applications defaults to using GPT-5.6 and Terra, with specific projects having the option to choose from GPT-5.6 and Sol. Employees have used it to handle tasks such as expiring interest rate cap transactions, fixed-income quotes, hedging dashboards, and confirmation document reviews. The common factor here is that the materials and rules are relatively clear, yet there is a significant amount of mechanical organization work involved. Just because employees can set up applications more quickly does not necessarily mean they can bypass data permissions, model validation, and customer delivery reviews; in fact, the more people can build their own tools, the more the enterprise needs to establish unified regulations regarding sources, versions, access, and change records.
Moving from single-point tools to systems, the challenges lie in traceability and accountability.
The Onyx platform, which is currently under development by Chatham, aims to place assets, debts, and derivatives in a governed data environment. The company states that Codex is involved in planning, development, testing, documentation, and code review, while the platform selects between different models based on the complexity of the tasks. Simple analyses and non-production tests tend to use less expensive models, with more powerful models being invoked only for complex tasks. This approach differs from the practice of using the most expensive model for all scenarios; the ultimate cost of a transaction includes not only the model invocation fees but also data reading and writing, personnel review, log retention, and exception handling.
The ChatFIN in the platform can help users search for historical market data, locate relevant legal documents regarding debts or derivatives, and provide corresponding links. Such links are not for decoration; they represent the most fundamental elements of trust in financial scenarios. If a model answers that "a certain clause allows for early termination," users must be able to return to the original contract to verify the wording, applicable dates, and parties involved in the transaction. If the document version has been updated, what previously seemed to be correct summaries may no longer be valid. Placing data governance behind the user interface is precisely what distinguishes enterprise applications from one-time demonstrations.
Of course, the case study is still a practical material jointly released by the company and the technology provider, not an independent random experiment. It indicates that there is room for improvement in certain processes, but it has not yet been proven that all institutions can achieve the same level of efficiency. For transactions that are more complex, have incomplete data, or involve more exceptions, the time savings may be significantly different. Presenting the early results as the industry average is neither accurate nor does it highlight those situations where human judgment is truly necessary.
For the financial industry, a more realistic measure of success might not be "how many tasks AI have been completed," but rather how much time reviewers spend on difficult cases: whether they can find evidence more quickly, whether they can discern the differences pointed out by the models, and whether they can stop when responsibilities are unclear. Chatham plans to expand the types of transactions covered while maintaining professional oversight. Only if this approach stands up to larger samples, more products, and external audits can AI be considered to have moved from improving individual efficiency to a reliable financial production process. As far as what has been disclosed so far, it is a valuable early case study, not a pass that allows one to skip reviews.
Source: OpenAI and Chatham Financial, case disclosed on October 2, 2026. https :// openai.com / index / chatham-financial /
