On September 10th, OpenAI launched ChatGPT, for, and Financial Services. The goal is clear: to bring together research, modeling, and customer material creation in one controlled workspace. This version utilizes GPT-6 and Astra, and comes built-in with financial data such as Daloopa, PitchBook, and LSEG News. Analysts can extract figures from publicly disclosed and authorized data to generate models with formulas, or create Word, Excel, and PowerPoint documents according to institutional templates. The product was designed in collaboration between OpenAI and Morgan Stanley, Evercore. It is currently available only to qualified financial institutions and requires contact with sales; it is not a service open to all individual users.
The financial industry is not short of tools that can write summaries. What is truly expensive is the process of linking together sources, models, formats, and approvals: researchers extract data from financial reports, analysts input the numbers into models, and senior staff check the assumptions before presenting the conclusions in a format that clients are familiar with. If any number lacks a source, subsequent reviews will result in the need to start over. This time, there is an emphasis on fine-grained citation of tables and paragraphs, and administrators are also allowed to publish their own templates. This indicates that the focus of competition has shifted from “generating a text that resembles a report” to “whether one can be integrated into the formal workflow.” OpenAI
Financial data is placed within its context; what is reduced is the need for manual transfer, not the ability to make judgments.
Daloopa is adept at structuring company disclosures, PitchBook covers the private market and company information, while LSEG News provides financial news. By integrating these features directly into the workspace, it theoretically reduces the need for copying and pasting between terminals, browsers, spreadsheets, and chat tools. Users can let the system compare company indicators, track industry changes, create comparable groups for transactions, or extract historical data required for models from the materials. Officials state that the relevant data is indexed and managed by OpenAI, so institutions no longer need to configure separate connections for these built-in sources.
However, "built-in" does not equate to "all data being complete," nor does it mean that "it can be used for any purpose." The scope of authorization, regional coverage, historical depth, and redistribution rights may vary among different organizations. Whether raw data can be displayed in customer materials also depends on contracts and internal rules. Before launching a product, legal and data procurement teams should still verify each permission boundary individually; one should not assume that data can be copied for external reports just because a number is displayed on the product interface.
Fine-grained references can shorten the review process, but they cannot replace accounting judgment. Companies may adjust the scope of their segments across different quarters, and non-GAAP (Generally Accepted Accounting Principles) indicators may exclude various items. Data from private enterprises may also come from estimates or earlier rounds of reporting. If a model simply combines indicators with similar names, even if the formulas are completely correct, the comparison may still be distorted. Analysts must check the period, currency, unit of measurement, whether the data has been audited, and whether the figures come from original disclosures or third-party calculations.
Generating financial models requires particular caution. A model is not merely a table of historical data; it also includes revenue drivers, profit margins, capital expenditures, discount rates, and various scenario assumptions. AI can be used to quickly set up formulas, check cell relationships, and generate sensitivity analyses, but the key assumptions should be clearly entered and signed by the person responsible for the project. Institutions must also prevent the system from replacing formulas with hardcoded numbers or causing reference discrepancies after inserting new rows. The most reliable approach is to write rules for testing the model: ensuring that assets and liabilities are balanced, cash flows are closed, that key outputs can be traced back to their sources, and that outliers trigger manual reviews.
The real barriers lie in permission management, record-keeping, and model risk management.
The enterprise controls listed for this product include SAML single sign-on, SCIM user management, role-based access control, transmission and storage encryption, and configurable data retention. Officials also stated that business data is not used for model training by default, and logs from supported workspaces can be exported through Compliance Platform. These capabilities provide a foundation for institutions to connect, but compliance results are not automatically generated by the list of functions. Enterprises still need to decide which teams can use which data, which tasks are allowed to be networked, where generated files should be saved, and how to revoke access after employees leave the company.
Financial institutions are also required to incorporate the output of AI into their existing model risk management and record-keeping systems. The responsible parties for research opinions, valuations, trading recommendations, and customer communications must be identified individuals, rather than "system-generated" outputs. Prompt words, references, model versions, manual modifications, and approval records should all be traceable. If the same request results in different answers after a model upgrade, institutions need to know whether the changes are due to new data, different reasoning processes, or alterations to the templates.
External materials also involve issues of appropriateness, fair disclosure, and conflicts of interest. AI may present public news alongside internal institutional information in the same context; therefore, work area isolation, project permissions, and information barriers should not be compromised for the sake of convenience. Even if the system does not intentionally disclose such information, it may indirectly expose restricted transactions, customer identities, or unpublicized forecasts through summaries or comparisons. High-risk projects should limit the tools that can be used and require manual approval before sending, downloading, or sharing any data.
The most reasonable early use for this product is to start with low-controversy, easily verifiable tasks, such as organizing public financial reports, checking model formulas, and fitting approved content into standard templates. Once the organization has accumulated data on accuracy rates, rework rates, and types of exceptions, it can then expand to more complex research and client deliveries. Measuring success should not only focus on "how many hours are saved," but also consider whether there has been a reduction in errors in references, unauthorized access, model rework, and the burden of review.
ChatGPT for Financial Services It is evident that general chat products are evolving into industry workstations. Built-in data and office file capabilities can significantly reduce manual tasks, allowing analysts to devote their time to hypothesis formation and judgment; at the same time, these tools also lead to errors being more quickly incorporated into formal documents. What the financial industry truly needs is not a writing machine that is always confident in its output, but a production system that can leave evidence, comply with authority, and be subject to review. Models can be drafted, and data can be automatically entered into tables, but the numbers that are ultimately presented to clients and the market must still be understood, verified, and for which responsibility can be assumed by humans.
