The Cost Curve Enterprises Did Not Model
Agentic AI's cost structure breaks the assumptions that enterprise procurement teams used to approve it. Traditional automation — RPA, workflow orchestration — has predictable compute costs tied to task volume. An agentic workflow that reasons across a long context, calls external tools, and spawns subagents generates token consumption tied to task complexity instead. A simple task costs little; an ambiguous or multi-step task can consume orders of magnitude more. Enterprises that approved agentic pilots under software licensing logic are now running what functions as a metered utility, and the per-task cost is only visible in retrospect.
EY's breakdown of agentic AI enterprise token costs names infrastructure, governance, risk, and Agent FinOps as distinct compounding layers — categories that did not exist in the prior automation procurement vocabulary. The closest historical parallel is early cloud adoption, where organizations budgeted compute as a fixed cost until AWS bills rewrote the model and cloud FinOps became a mandatory discipline. Agent FinOps is following the same arc, with one complication: the consumption variable is reasoning depth, not compute hours, and reasoning depth cannot be capped without degrading what the agent can do. The organizations absorbing this cost surprise are not outliers; they are the early majority of enterprise agentic adopters, and the correction will show up in budget reviews, not product announcements.
What Lloyds Proved — and What Regulators Noticed
The financial sector is the sharpest test case for agentic AI at scale, and the results are genuinely mixed in instructive ways. Lloyds' deployment of real-time AI fraud detection — which has prevented over £1bn in fraud — demonstrates that agentic systems can deliver measurable operational outcomes at institutional scale. That result is real and the deployment is live.
Regulators drew a different conclusion from the same landscape. A global financial watchdog's call for tighter controls on agentic AI in finance arrived while Lloyds-scale deployments were already running. The regulatory response confirms what the deployment proves: agentic AI in finance is no longer a pilot-stage conversation. The question regulators are now asking — about controls, accountability, and systemic risk — is the question enterprises skipped when they approved the pilots. The ones that answered it before deploying, rather than after, are the ones that will avoid the governance correction.
The Subsidy Argument No Vendor Has Answered
The sharpest line in the current public conversation about agentic AI is not about capability — it is about price. The concern circulating in practitioner communities is that agentic AI is currently priced at venture capital subsidy rates, and that customers exposed to real operational costs would abandon the technology . This is not a fringe position; it is the structural critique that enterprise vendors have declined to engage directly.
The ACM TechBrief on agentic AI policy identified the same asymmetry from a different angle: agents that act on behalf of users — browsing, executing code, sending messages — create accountability gaps that neither vendors nor enterprises have closed . The financial version of that accountability gap is the cost model. Practitioners being moved into agentic AI roles at employer direction are inheriting both the capability and the liability of systems whose true economics remain obscured by the current subsidy environment. When the subsidy compresses, the enterprises that built agentic workflows around subsidized pricing will face a forced renegotiation of what they actually need the technology to do — and the vendors that priced below sustainable cost will not be the ones writing the renegotiation terms.
Apple's Restraint as a Revealed Preference
The conversation about agentic AI's costs does not happen in isolation from competitive positioning. Apple's deliberate restraint on agentic AI at WWDC — prioritizing a privacy-focused, utility-oriented Siri over broad agentic autonomy — reads differently against the cost and accountability concerns now surfacing in enterprise deployments. Where the dominant vendor narrative treats agentic capability as an unambiguous accelerant, Apple's posture treats current models' hallucination rates and trustworthiness gaps as disqualifying for the use cases that would generate the most liability .
That restraint may prove to be the more commercially durable position. The enterprises discovering that agentic AI costs scale unpredictably with task ambiguity are also discovering that the tasks generating the most cost are often the tasks where agent reliability is lowest. Apple is, in effect, refusing to sell the product whose failure mode is already documented — and the enterprise buyers who ignored that failure mode are the ones now writing the corrective budget memos.
Who Absorbs the Correction
The enterprises most exposed to agentic AI's cost correction are identifiable by a specific pattern: they scaled from pilot to production before building the FinOps discipline to monitor agent token consumption in real time, and they approved budgets under licensing assumptions that do not match metered utility pricing. The correction, when it arrives at the budget review level, will not be labeled a technology failure — it will be labeled a forecasting error, and the accountability will fall on whoever signed the procurement approval.
The practitioners now being asked to move into agentic AI roles — often without choosing to — are the ones who will be asked to explain the variance. The vendors shipping agentic tooling will not absorb it. The data infrastructure layer is quietly becoming more strategic as a result: CTERA's integration with automation platform n8n signals that the organizations solving AI's data access problem securely are positioning themselves as the mandatory foundation layer beneath whatever agentic tooling enterprises choose. The enterprises that build cost discipline around that foundation layer — rather than around the agents themselves — are the ones that will emerge from this correction with a viable long-term model.