Who Owns Your AI Agents
AI agents are entering finance workflows as systems that execute work semi-autonomously: processing invoices, monitoring contract compliance, flagging anomalies, routing exceptions.
The architecture question is settling. ERP remains the foundation layer—your system of record for transactions, balances, and audit trails. AI agents operate in the intelligence layer above it, drawing from multiple systems for pattern recognition and cross-system analysis. A flexibility layer handles ad hoc queries and executive dashboards.
Agents depend on this foundation. When an agent flags a compliance issue or recommends a hedge, the underlying data still lives in ERP. That's where you reconcile. That's what auditors trace to.
The technical architecture is becoming clear. What's missing is the governance architecture to match it.
The Complication
Srinivasan and Wei, writing in Harvard Business Review, document the emergence of dedicated roles to govern agent performance.
Salesforce's Agentforce platform now autonomously resolves 74% of inbound customer support cases. Their sales development team went from 150 meetings in 30 days to over 350 meetings in a single week after deploying AI agents, generating $60 million in annualized pipeline within four months.
Order-of-magnitude changes that only happen when AI agents operate semi-autonomously at scale.
Salesforce created dedicated roles to govern these systems. Zach Stauber, an agent manager in their organisation, describes his daily routine: "Data, Data, Data. I start and end my day in dashboards, scorecards, and agent observability monitoring."
Governance embedded in operations.
The Ownership Question
Srinivasan and Wei make a claim with direct implications for CFOs: "In the pre-agentic world, AI deployment lived within IT or the data science organisation. In the agentic era, business units need to take control."
If AI agents are performing real work for a business unit, that unit must own their performance.
The deterministic-probabilistic boundary matters here. ERP processes—general ledger posting, tax calculations, regulatory reporting—require rule-based certainty where "99% accurate" means compliance failure. AI excels at probabilistic tasks: pattern recognition, anomaly detection, forecasting.
The governance challenge is maintaining clear separation between these domains. AI provides recommendations and flags exceptions. Humans approve decisions that impact financial records. Deterministic code executes approved transactions.
This separation requires someone accountable for agent performance. Someone who assesses whether outputs remain reliable and knows when to escalate.
For finance, this accountability extends beyond agents deployed directly within your processes. If an AI agent processes supplier data before it reaches your AP workflow, someone must own that agent's performance—and finance needs visibility into output quality and clear escalation paths when issues arise. The owner may sit in procurement or operations, but the governance gap affects finance regardless.
The capabilities Srinivasan and Wei identify for effective agent managers:
AI operational literacy: Operational judgment about probabilistic systems. Understanding how agents operate and how to diagnose performance issues.
Functional depth: Deep knowledge of the business process the agent supports. Someone who knows what "good" looks like in invoice processing or variance analysis.
Systems thinking: Visualising how agents interact across workflows and departments, recognising that finance processes connect to upstream and downstream systems.
Work design across humans and machines: Creating hybrid workflows, assessing machine capabilities, designing escalation routines.
Change resilience: Adapting to shifting models through continuous monitoring. Salesforce's agent managers work in weekly "test-deploy-learn" cycles.
The Metric Shift
The HBR research documents a shift in how performance should be measured when agents enter the workforce. In sales and support contexts, Salesforce moved from activity metrics to orchestration outcomes—measuring how well people govern the human-agent system.
For finance, the question is whether productivity gains actually convert to value.
Shah and Course's Six Levels of Benefits framework offers a structure. Productivity gains—time saved through automation—sit in the middle of a progression. They deliver no financial value unless explicitly converted: either downward to cost reduction through actual headcount or spend decreases, or upward to strategic capability through redeployed capacity that influences decisions.
Measuring time saved without tracking conversion perpetuates the measurement gap that has plagued finance transformation for a decade.
Anders Liu-Lindberg operationalises this through the Value Log: finance employees document specific actions where they influenced business performance. Ørsted implemented this across 600+ finance employees. Their Impact Log now contains close to 1,000 documented cases. A global pharmaceutical company sets annual targets of $150 million in documented value contributions.
Liu-Lindberg emphasises that the Value Log drives behaviour. Finance employees become conscious of how they contribute to value creation across the drivers that matter: revenue growth, margin improvement, working capital optimisation, risk mitigation.
When agents handle volume, the human role shifts to business partnering. The outcome is whether finance analysis influenced a better capital allocation decision, pricing strategy, or risk position. That's what gets logged.
The Resolution
The three-layer architecture clarifies what's required. ERP remains your foundation—the system of record that agents depend on. The intelligence layer adds capability. That capability demands answers to four questions:
Who owns agent performance in the intelligence layer, including for upstream processes that feed finance data?
What monitoring enables finance to assess output reliability?
How do you convert productivity gains to documented value creation?
Who governs the continuous iteration cycles that the intelligence layer requires?
The organisations getting value from AI agents are building the operational infrastructure to govern them. That infrastructure includes clear ownership extending across process boundaries, monitoring that finance can interpret, and measurement systems that track value conversion.
The foundation stays stable. What you build on top requires ongoing governance.
