Your enterprise is deploying AI agents at twice the rate it was six months ago — and you probably have no idea what it's costing you.
That's not a hypothesis. It's the central finding from KPMG's Q2 2026 AI Pulse Survey, which captured perspectives from 204 U.S. C-suite and business leaders at organizations with annual revenue of $1 billion or more. The numbers reveal a dangerous gap forming at the center of enterprise AI strategy: the gap between how fast we're deploying agents and how well we understand what we're spending.
Multi-agent deployments — where coordinated AI agents work across workflows, share goals, and automate cross-functional decisions — doubled from 9% to 18% of organizations between Q1 and Q2 2026. Meanwhile, only 26% of those same organizations have real-time visibility into the cost of running AI at scale. That's not a minor oversight. It's a structural problem that's about to become someone's very expensive quarterly surprise.
The Double Without the Discipline
The pace of enterprise AI adoption has been relentless. But the move from isolated AI tools to coordinated, enterprise-wide agent systems changes the financial equation entirely.
When you deploy a single AI assistant to help your marketing team draft copy, the economics are simple: a flat monthly seat fee, predictable usage, manageable cost. When you deploy three agents that autonomously coordinate across your advertising stack, inventory system, and campaign analytics — each passing context to the others, triggering new tasks, and making decisions without human intervention at each step — you're no longer dealing with fixed costs. You're dealing with variable compute: token consumption, inference calls, API chains, and orchestration overhead that can spike in ways your finance team has no framework to anticipate.
EY calls this the "agentic FinOps" problem. In a May 2026 analysis, EY noted that agentic AI is fundamentally shifting enterprise AI from fixed software and labor costs to variable compute use — and that most organizations haven't built the financial literacy or governance infrastructure to manage that shift.
The KPMG data confirms this in sharp terms. While 66% of organizations have monitoring dashboards and 61% have approval processes in place, only 36% have implemented direct token or usage controls. You can monitor what's happening. You can get approval before you start. But once the agents are running, most enterprises have no mechanism to actually constrain what they spend.
The Agentic Multiplier Problem
Here's what makes this particularly acute: a single agent task doesn't just consume one round of tokens. It consumes a chain.
An agent tasked with preparing a competitive analysis doesn't just call one model once. It calls a search tool, parses the results, calls the model with context from those results, evaluates the output, decides to search again, refines the query, calls the model with updated context, and produces a final synthesis — each step consuming tokens. Multiply that across hundreds of concurrent agents running across your enterprise, and you have what Portal26 describes as a task that can "spiral into millions of tokens in minutes."
The math isn't intuitive. Enterprise finance teams are accustomed to thinking about AI costs the way they think about SaaS: a contract, a seat count, a fixed monthly line item. Agentic AI doesn't work that way. The cost is a function of the complexity of tasks assigned, the number of agents running, the models you route to, and the efficiency of your prompts — none of which are static.
According to KPMG, 35% of enterprise leaders explicitly identify AI cost management and economic literacy — including understanding usage-based pricing models like token and inference costs — as barriers to their AI strategy. That's more than a third of the leaders driving these deployments who admit they don't fully understand what they're buying.
What CFOs Are About to Do
CFOs have noticed.
Enterprise AI spending is entering a phase of financial scrutiny that mirrors what happened to cloud infrastructure spending around 2019: the initial wave of open-ended investment is giving way to demands for measurable returns and tighter budget controls.
Forbes contributors covering the enterprise technology space note that CFOs across industries are moving aggressively to impose budget controls on AI projects. Organizations that cannot demonstrate clear productivity gains, cost reductions, or revenue impact from their AI investments are facing project cancellations and budget freezes.
The average enterprise plans to invest $202 million in AI over the next 12 months, according to KPMG — roughly flat with the prior quarter's $207 million average. That plateau is itself a signal: the era of writing blank checks for AI transformation is ending. CFOs want to understand what they're getting before they approve more.
The convergence of accelerating multi-agent deployments and tightening financial scrutiny creates a specific risk for enterprises that haven't closed the cost visibility gap: you may be doubling your agent usage at precisely the moment your CFO starts asking where all the money went.
Three Things Enterprises With Cost Visibility Do Differently
The 26% of organizations with real-time cost visibility aren't just lucky. They've built specific capabilities that the other 74% haven't.
First, they have token-level attribution. Not just "AI cost this month was $2.3M" but "Agent X, running in the Marketing department, consumed 47M tokens on Campaign Planning workflows, with 63% of that going to search queries that could be cached or batched." Attribution at that level of granularity is what allows you to optimize intelligently rather than just cap spending arbitrarily.
Second, they've implemented circuit breakers. These are automated cost controls that pause or reroute an agent when it's consuming tokens at an anomalous rate. Think of it as a breaker panel for compute: when one circuit gets overloaded, it trips rather than burning down the building. TrueFoundry and similar platforms describe these as essential infrastructure for production agents — token budgets, circuit breakers, model routing, and real-time cost attribution are no longer optional in an environment where agents can run continuously.
Third, they've separated model routing decisions from application logic. Rather than hardcoding "use the flagship model for everything," they route tasks dynamically based on complexity, cost, and latency requirements. A simple summarization task routes to a cheaper, faster model. A complex multi-step reasoning task routes to a more capable but more expensive one. The routing layer pays for itself quickly when you're running agents at scale.
For Technical Leaders: Build This Now
If you're a CTO, CIO, or Head of AI, the cost visibility problem is squarely in your domain — even if it feels like a finance issue.
The infrastructure choices you make in the next two quarters will determine whether your enterprise can scale agentic AI responsibly or whether you'll spend 2027 doing post-mortems on runaway compute budgets.
Three things to prioritize:
Instrument everything at the agent level. Before you deploy the next wave of agents, build logging that captures token consumption, model calls, and execution time per agent, per task, per workflow. This is boring operational work, but it's the foundation of every useful cost conversation you'll have with your CFO.
Create a weekly AI cost review cadence. Don't wait for end-of-month finance reports to learn that AI compute costs spiked. Weekly reviews of AI cost trends — broken down by agent type, department, and use case — let you course-correct before surprises become crises.
Treat prompt efficiency as an engineering priority. The gap between a well-crafted prompt and a poorly-crafted one can be an order of magnitude in token consumption for equivalent outputs. Teams that invest in prompt engineering as an operational discipline — not just an initial setup task — see consistent cost advantages as they scale.
For Business Leaders: The Right Questions to Ask
If you're a CFO, COO, or business unit leader trying to understand your AI cost exposure, KPMG's data suggests you're not alone in the dark. But the questions you ask now will determine how quickly your organization closes the gap.
Start with these: Can we see our AI costs in real time, broken down by department and use case? Do we have controls that prevent an agent from consuming unlimited resources? Do we understand the difference between token costs and total agent operating costs — including the people time, integration costs, and quality assurance overhead that surrounds each automated workflow?
If your team can't answer those questions with data rather than estimates, you've just identified your cost visibility gap in concrete form.
KPMG's data also flags a concerning trend worth watching: "token-maxxing," a practice where organizations gamify token consumption through incentives and leaderboards to drive AI adoption. While 41% of leaders say they'd consider it, KPMG's Edwige Sacco, Head of Workforce Innovation, was direct in her assessment: "Token-maxxing presents a classic risk of incentivizing activity over outcomes. What starts as a productivity metric can quickly become a distraction. In the short term, it deteriorates value; in the long term, it undermines culture."
This warning applies broadly to any leader considering aggressive adoption incentives: measuring AI usage by volume — tokens consumed, queries submitted, agents deployed — creates the wrong incentives at exactly the wrong moment in enterprise AI maturity.
The Competitive Advantage in the Visibility Gap
There's a contrarian framing worth considering here.
The 74% of enterprises that don't have real-time AI cost visibility aren't necessarily in crisis right now. Many of them are still in a phase where agent deployments are small enough that the cost visibility gap hasn't created a material business problem. The KPMG data shows overall AI investment holding steady at $202M, which suggests the industry isn't panicking.
But the doubling of multi-agent deployments is the leading indicator. Today's 18% of organizations running coordinated multi-agent systems will be 40% or 60% within 18 months. The complexity and cost trajectory of agentic AI only goes one direction — up — as models improve, use cases expand, and organizational appetite for automation deepens.
At FinOps X in 2026, conversations with enterprise teams revealed that AI is now a core operating cost for 98% of organizations — up from a small fraction just two years ago. The infrastructure for managing that cost is still catching up.
The enterprises that close the visibility gap now will have a structural advantage: they'll be able to scale more confidently, optimize more effectively, and report more credibly to their boards. The ones that wait will face the same governance scramble that cloud teams faced in the early 2020s, except compressed — because agentic AI is scaling faster than cloud did.
KPMG's Rahsaan Shears, AI enterprise transformation leader, framed it directly: "AI agents are changing both the operating model and the economics. As organizations shift from isolated deployments to coordinated, enterprise-wide use, good governance is what ties scale, performance and value together."
The agents have doubled. The costs haven't been explained. That's the gap to close — and the organizations that close it first will have a real competitive advantage heading into 2027.
The bottom line: Multi-agent AI is no longer a pilot. It's running in production, at scale, with variable costs most enterprises cannot see. The 26% with real-time visibility aren't ahead because they're richer or more technically sophisticated — they're ahead because they treated cost governance as infrastructure, not afterthought. That decision is still available to the other 74%. It's just getting more expensive to delay.
Sources: KPMG Q2 2026 AI Quarterly Pulse Survey (June 25, 2026); EY Agentic AI Token Costs analysis (May 2026); CFO Dive; MarketScale; Portal26.
