ChatGPT for Financial Services Changes the Data Contract, Not the Review Duty

The demo ends with a clean valuation model, a sourced earnings note, and a client deck in the firm template. The room starts discussing seat price. That is the wrong negotiation. ChatGPT for Financial Services is not simply another assistant license; it joins licensed data, retrieval, reasoning, templates, and distribution inside one operating surface. The buyer is choosing a data and review architecture.

What happened

OpenAI introduced ChatGPT for Financial Services on September 10. The product includes built-in datasets from providers such as Daloopa, PitchBook, LSEG News, and Crunchbase; entitlement connections for existing subscriptions; more than 50 connectors; firm-managed Office templates; and granular citations back to source tables and passages. It was shaped with Morgan Stanley and Evercore and is available to eligible financial institutions through sales.

OpenAI says business data is not used to train its models by default, data is encrypted at rest and in transit, retention can be configured, supported workspace logs can be exported, app actions can be controlled by role, and separate workspaces can enforce information barriers. Those are meaningful enterprise controls. They are also the beginning of diligence, not the conclusion.

What is real

The integration burden is a real source of cost and failure. Teams lose time negotiating separate access, maintaining connectors, normalizing identifiers, and proving which source supported which figure. Putting entitled data and citations in the same place as the analysis can shorten that route. The useful unit is not “answer from GPT.” It is a reviewable work package: number, definition, period, source passage, transformation, and final judgment.

The product also makes a strategic trade: less plumbing for the buyer, more dependence on the platform’s indexing, entitlement handling, retrieval behavior, templates, and audit export. That can be the right bargain. It should be priced and governed as a workflow dependency rather than compared only with a general-purpose chat seat.

The skeptical read

The launch page contains no public price, no customer outcome study, and no published error rate for the named workflows. Design partners and data providers are participants in the offer, not independent evaluators. Citations reduce the distance to evidence; they do not prove the evidence was selected correctly, interpreted consistently, or carried into a model without an accounting mistake.

The SEC’s chief economist made the counterpoint days later: AI lowers the cost of processing information, but stronger analytical tools make structured, consistent, high-quality inputs more important. Machine-readable abundance is not a substitute for materiality or verification. In finance, a confidently sourced non-GAAP adjustment can still be the wrong adjustment.

The buyer’s six-part data contract

• Rights: which built-in and connected datasets may each role use, retain, quote, and place in client materials?
• Lineage: can a reviewer move from every material figure to the exact source, period, unit, and transformation?
• Freshness: how quickly does each dataset update, and how does the product mark stale or conflicting values?
• Boundaries: where are material non-public information, research, banking, and client information separated?
• Evidence: which logs can compliance export, how long are they available, and do they include source and action history?
• Exit: can the firm preserve prompts, templates, citations, logs, and work products if it changes models or platforms?

A pilot worth running

Use one recurring, reviewable job such as an earnings update for 20 covered companies. Run the current process and the new process side by side for two cycles. Measure analyst preparation time, reviewer time, citation completeness, number of corrected figures, material rework, and time to approved output. Predefine the stop rule: one information-barrier breach, missing lineage on a material claim, or a correction rate above the current process pauses the pilot.

If the tool wins, negotiate around the whole workflow: data entitlements, connector support, retention, audit exports, template administration, incident response, and exit. A low seat price can hide an expensive dependency. A higher price can be rational if it removes enough integration and review cost. The evidence comes from accepted work, not the launch demo.

Primary sources

• OpenAI: Introducing ChatGPT for Financial Services — Product scope, data providers, controls, and availability.
• SEC: Information in the Age of AI — September 17 remarks on structured data, materiality, and verification.