Most enterprises can answer exactly one question about their AI rollout: what did we spend?
Very few can answer the questions that actually matter:
- What did this agent cost us per customer served?
- Which AI features are expanding revenue or improving margin?
- Where is AI making the business more efficient?
Agentic systems spend through behavior: recursive model calls, tool fanout, retries, delegation chains, and sometimes endless replanning loops, but unlike human effort, that spend is variable and hard to predict – a single workflow can burn through real money in minutes, autonomously, without judgement.
Oracle recently published an architecturally serious treatment of runtime budget guardrails: closed control loops that track execution state, decide if a run still deserves more budget, and circuit-break to prevent obvious waste.
Applying judgement that governs agents in the moment matters. But judgement is only as smart as the signals that inform it.
Guardrails that kill a run at $50 in token spend are useful in exactly the same way a guardrail on a highway is useful – obvious disasters are prevented. The smart traffic light that turns green for an ambulance and red for cross traffic (or a circuit breaker that knows whether that $50 generated $500 in customer value, or zero) is doing something else entirely.
The control plane pattern isn’t new
Every mature infrastructure category eventually produces a control plane: a single authoritative layer that issues decisions for distributed systems to enforce.
For example: Okta is an identity control plane. It issues identity decisions that runtime systems consume. Okta doesn’t run your applications. It provides the authoritative signal that determines what they’re allowed to do.
Open Policy Agent works the same way for authorization. OPA evaluates policies that services must enforce. The runtime doesn’t need the decision logic; it just acts in alignment with the signals sent from the control plane.
These layers showed up on their own schedule. Identity got a control plane when the number of applications outgrew any team’s capacity to manage access by hand. Authorization got one when policy stopped fitting cleanly inside individual services. The trigger was the same in both cases: scale. The problem outgrew what ad hoc, distributed decisions could absorb.
AI economics is there now, and the right time to build its layer is before AI cost becomes the reason AI projects get cancelled.
That layer is the financial control plane. Outcome-level information (unit economics, contract margins, customer profitability) are consumed by downstream systems. The financial control plane isn’t another orchestration layer. Its job is narrower and more important: producing the economic signals other systems act on, converting runtime enforcement and operational governance from crash barrier to P&L guardian.
Three layers, three jobs
The enterprise AI governance stack is settling into three tiers, and money gets lost when these tiers are conflated.
Runtime control planes (Oracle’s budget guardrails, gateway-level token metering, AgentCore-style execution control) apply judgement at the moment of execution.
Keep running. Switch models. Stop. These have to be infrastructure-speed decisions because an agentic workflow that spawns sub-agents and retries tool calls is too fast for human-in-the-loop. This layer implements policy, it doesn’t create it.
Operational control towers (ServiceNow’s AI Control Tower, Salesforce Agentforce, Microsoft Agent 365) govern what AI is doing across the enterprise. They discover agents, manage lifecycle, and report on AI activity at the org level.
ServiceNow’s tower now integrates across 30+ enterprise systems. Agent 365 puts a unified governance surface over agents regardless of where they were built. This is where executive-level AI governance is going to live, and where customer ROI claims get made.
The structural limit on these products is the signals they can consume. They govern behavior and policy from the application tier. Some supply dashboards filled with activity metrics, but consumption reported from the application tier doesn’t offer the visibility to allocate spend to outcomes. Cost and value need to come from a different vantage point.
The financial control plane sits underneath both. It produces unit cost per user session, cost per customer outcome, margin impact by AI feature, allocated development cost by product release. Those are the signals that let the other tiers act intelligently rather than blindly.
It also has to be provider-neutral in a way the other two don’t. Unit economics have to be calculated across shared costs, across providers, and you can’t get there from a single-cloud governance product. Unit economics requires a view of spending no single-provider source can get.
Without the financial control plane, runtime guardrails are blunt instruments enforcing arbitrary thresholds, and control towers report on AI activity but can’t call which bets deserve more investment.
Why this matters to the CFO’s office
Agentic AI spending will hit $201.9 billion in 2026, up 141% year-over-year per Gartner. The same analysts expect over 40% of agentic AI projects to be canceled by end of 2027, with runaway cost, unclear value, and inadequate controls cited as the reasons.
That cancellation rate isn’t a cost-management problem. Cost management is a solved discipline at this point. The harder problem is telling the AI bets that produce ROI per customer, per feature, per process apart from the ones that don’t. Infrastructure billing and operational dashboards are structurally unable to answer these questions, because they can’t allocate spend to outcomes.
That’s what the financial control plane does. It connects what AI costs to what it produces, expressed in terms that line up with a P&L.
The architecture requires all three
Token-level guardrails govern infrastructure spend and control towers govern enterprise AI activity. Outcome-level governance (unit economics, contract margins, customer profitability) requires the financial control plane upstream of both. None of these layers substitutes for the others.
Treating AI cost governance as a one-layer problem is most of what will deliver that 40% cancellation rate.
Companies that build all three layers, with the financial control plane producing the signal that gives the other two something to work with, won’t be flying blind through what comes next.
FinOps In The AI Era: A Critical Recalibration
What 475 executives told us about AI and cloud efficiency.
