From Conversational AI to Execution-First Finance Agents
Finance AI agents are software systems that combine language understanding with direct connections into business applications so they can follow defined procedures and execute accounting tasks with human oversight instead of only answering questions. For early finance adopters, this marks a clear break from the chatbot era. As Stephanie Montelius of Nominal argues, traditional conversational tools explain, summarize, and provide context, but they do not remove work from overloaded ERP users. Finance teams now want agentic AI execution that can post entries, prepare reconciliations, and manage intercompany steps alongside the ERP system. At events such as Sage Future, the discussion has shifted from “Can AI explain my numbers?” to “Can it complete the workflow for me without losing control?” This change is driving interest in autonomous accounting systems that operate near, but not inside, systems of record, and that embed enterprise AI oversight from the start.
Why Execution Demands Auditability, Determinism, and Oversight
As finance AI agents move from pilots to production, the main concern is no longer model accuracy alone but governed execution. Work like period-end close, billing cycles, and reconciliations spans thousands of steps across ERP, cloud services, and legacy tools, all bound by timings, approvals, and dependencies. Charles Crouchman of Redwood warns that probabilistic AI can introduce subtle inconsistencies into these chains that “do not fail loudly” but accumulate as processes continue. That is why enterprise AI oversight is now central to adoption. Systems must keep clear logs of every agent decision, apply deterministic guardrails before data touches systems of record, and make every automated action auditable. For CFOs, the promise of agentic AI execution only holds if they can answer a simple governance question at any time: who (or what) did what, when, and according to which standard operating procedure?
ERP-Adjacent Agents: Automating Work Without Giving Up Control
Execution-focused finance AI agents tend to sit ERP-adjacent rather than inside core ledgers. Nominal’s approach is one example: an agentic performance management layer that follows standard operating procedures, integrates across multiple ERP environments, and handles recurring accounting workflows such as intercompany transactions and reconciliation. In parallel, SAP is building an intelligence layer around Joule Work, AI Agent Hub, and Company Memory so agents can understand process context and coordinate across HR, procurement, supply chain, and finance. Redwood positions its RunMyJobs platform at the execution handoff, giving agents governed access to existing workflows, jobs, and enterprise connectors as constrained tools. This architecture keeps human accountability on the finance side while enabling autonomous accounting systems to run large volumes of work. Agentic AI can act, but deterministic workflows, access controls, and clear escalation paths define how far it is allowed to go.
Implementation Complexity and the Autonomous Enterprise Readiness Gap
The move toward an autonomous enterprise makes finance leaders confront implementation complexity that chatbots mostly avoided. SAP’s more than 200 agents reveal how quickly agentic environments expand, increasing the risk surface across multi-entity, multi-ERP estates. Crouchman calls the hard part “governed execution”: coordinating many agents without losing determinism or auditability. One practical barrier is the cold-start problem. Many organizations already have years of automation in job schedules, finance workflows, and integration scripts. Re-teaching that logic to new agents via prompts would be slow and error-prone. Redwood’s answer is to treat this automation library as reusable execution logic, so agents can call proven workflows instead of inventing new ones. Nominal, meanwhile, focuses on ERP-agnostic integrations and standard operating procedures. For finance teams, readiness now means more than buying models; it means cataloging existing automation, clarifying ownership, and defining which tasks can be delegated to finance AI agents under strict oversight.
