Financial AI needs source evidence

Financial AI is becoming a product made up of a model, a data licence and an evidence trail. On 10 September, OpenAI introduced ChatGPT for Financial Services: a ChatGPT Work environment aimed at investment banking and equity research, with GPT-6 Astra and integrated financial data. For banks, this is less a model release than a new procurement and control question.

Verify provenance, not only answers

According to OpenAI, the offering includes data from providers such as Daloopa, PitchBook, LSEG News and Crunchbase. It is intended to trace figures and claims back to their sources through granular citations. Existing subscriptions are to be connected through shared sign-in and entitlements; OpenAI also cites central management of access and data connections.

This matters because financial analysis is not judged by the appearance of a presentation. The crucial questions are whether a figure comes from an authorised data set, whether the source used is current and whether the path to the source can be reconstructed in review. CNBC additionally reports features for citations, chart verification and administrative controls for sensitive deal materials.

The procurement object gets wider

For banks, insurers and corporate-finance teams in the DACH region, the implication is clear: do not buy access to a model alone. Buy and assess a verifiable work chain. Before a pilot, business teams, procurement, privacy and information security should document at least these points:

• Data rights: Which built-in data may be used for which purpose and user group?
• Entitlement: How is it verified that only already licensed sources and internal content can be reached?
• Evidence: Can every decision-relevant figure be traced to a document, table row or data record?
• Approval: Who reviews analyses, charts and generated material before they enter credit, investment or transaction processes?

OpenAI describes the offering as available to eligible financial institutions. Whether it becomes a productive deployment does not depend on the model alone. A reliable pilot primarily needs a test set with real but controlled cases: correct sources, deliberately outdated data, missing permissions and conflicting figures. Only when the system visibly separates those cases is it a robust tool for demanding financial work.