What building an AI-native finance function taught me

Key Takeaways

• Give people access, but anchor experimentation in real work. OpenAI found that broad access becomes more useful when employees closest to a problem can build and test solutions, while leadership concentrates resources on the highest-value opportunities.
• Redesign the workflow around the decision, not around existing tasks. Instead of merely accelerating spreadsheets, reconciliation and slide preparation, start with the decision that matters and work backward through the data, tools, approvals and handoffs needed to support it.
• Domain experts can increasingly become builders. Finance professionals at OpenAI are using ChatGPT Work and Codex to create live dashboards and tools themselves, allowing the people who understand the business problem to shape and continually improve the solution.
• More automation requires clearer human ownership, not less. AI can prepare explanations, reconcile information and draft responses, but outputs should remain traceable to reliable sources, important changes should require approval, and people remain accountable for the final result.
• Measure completed useful work, not AI consumption. Seats and token usage reveal little about value; finance should track outcomes such as cycle time, reconciliation rates, forecast accuracy, review effort and decision quality—including whether a stronger model reduces total cost by requiring fewer retries and less human review.