Enterprise AI agents do not fail only because a model is wrong; they fail when ownership, access, context, review, and evidence are unclear. This session shares practitioner patterns from building production AI workflows for finance 
and enterprise operations at OpenAI: governed tool contracts, MCP-style integrations, Apps SDK UI surfaces, Temporal workflows, evals, citations, access controls, and human-in-the-loop review. Attendees will see how to design agent workflows that preserve audit trails, support reviewer trust, enforce permission boundaries, and give governance, risk, and operations teams evidence they can actually use. 

Key takeaways: 

• Define governed tool and data contracts for agents. 
• Separate model reasoning from permissioned actions. 
• Use citations, provenance, and workflow state to create an audit trail. 
• Design reviewer experiences for approvals, exceptions, and escalation. 
• Measure agent quality through evals, incident review, and workflow outcomes. 

Speaker: Siddharth Jain, AI Engineering Manager, Finance AI, OpenAI 

Siddharth Jain is an AI Engineering Manager working on finance and enterprise AI systems at OpenAI. He builds production AI workflows for internal finance and operations, including contract intelligence, procurement support, revenue commentary, vendor risk review, approvals, MCP integrations and UIs, Apps SDK surfaces, Temporal workflows, evals, access controls, citations, and human-in-the-loop review. His work focuses on turning AI prototypes into reliable systems that preserve auditability, governance, and controls in real business workflows.