TL;DR
- McKinsey (July 2026) found that ~60% of agentic AI operating costs come from verifying and refining outputs, not generating them — flipping the cost model enterprises assumed.
- Token-based pricing is effectively obsolete; real costs are driven by reasoning complexity, tool orchestration, and retry logic.
- Nearly 60% of enterprises lack the tools to even measure the ROI of their AI-driven decisions — a dangerous blind spot at scale.
- The implication: agentic AI implementation is no longer an IT decision. It's a CFO-level economics problem that must be solved before production, not after.
The Party Was Great. Then the Bill Arrived.
For the past two years, enterprise AI adoption has felt like an all-inclusive resort — everyone's excited, the demos are stunning, and nobody's quite sure who's picking up the tab. That vibe ended in July 2026.
McKinsey's latest research into agentic AI economics landed like a very expensive wake-up call: the cost structure of autonomous AI systems in production is fundamentally different from what organizations budgeted for in the pilot phase. And in many boardrooms, the CFO is now the person asking the loudest questions.
The core finding is almost counterintuitive: approximately 60% of agentic AI operating costs are consumed not by generating responses, but by verifying and refining them. The work that happens after the model speaks — the checking, the retry loops, the validation chains — is where the money actually goes. If your AI cost model assumed you were mostly paying for inference, you were budgeting for the wrong line item.
The 1,000x Token Problem Nobody Warned You About
Here's where the numbers get genuinely startling. Agentic AI tasks, due to their long-lived context windows and multi-step reasoning loops, consume roughly 1,000 times more tokens than conventional chat-based or code-reasoning applications. Not 10x. Not 100x. One thousand.
To understand why, it helps to understand what agentic AI actually does. Unlike a chatbot that responds to a single prompt, an agentic system operates through a continuous five-step loop: perception, reasoning, planning, action, and learning. Each cycle pulls in fresh context — from APIs, databases, CRM systems, enterprise resource planning platforms — and feeds it back into the next reasoning step. That context accumulates. It compounds. And every token in that expanding context window costs money.
McKinsey put it plainly: "Per-token pricing has stopped being a useful measure for what enterprises actually pay for gen AI." The real cost levers are the reasoning path the agent takes, which tools it orchestrates, and how many times it retries a failed step. A task that requires three tool calls and two retries on a heavy reasoning model can cost orders of magnitude more than a simpler equivalent — and the difference isn't visible in a standard per-token invoice.
Economics Over Technology: A Fundamental Shift
Perhaps the most significant framing in McKinsey's report is this direct quote, worth reading twice:
"The decision to scale an agent is increasingly becoming a complex and fast-changing economics decision, not a technical one."
That sentence should be printed, laminated, and taped to the monitor of every enterprise architect who has ever said "we'll figure out the cost model after we validate the use case."
Because here's the trap: organizations that move from pilot to production without resolving the economics question don't just overspend. They get locked in. Rearchitecting an agentic system after production deployment — ripping out context management approaches, swapping orchestration frameworks, redesigning verification workflows — is genuinely painful and expensive. The design phase is the only moment when cost-aware decisions are still cheap to make.
This is precisely why implementation partners who can translate CFO constraints into actual system design choices — which model tier for which task, when to use lightweight verification vs. heavy reasoning, how to manage context windows efficiently — are becoming the real differentiators in the market. It's not about which LLM you pick. It's about architecting for measurable returns from day one.
The Visibility Gap: Spending Blind
If the cost structure problem weren't enough, McKinsey surfaces a second crisis hiding underneath it: nearly 60% of enterprises still lack robust mechanisms to measure the business impact of AI-driven decisions, even as they scale agentic deployments across their operations.
Think about that for a moment. Organizations are committing to autonomous AI systems that make thousands of micro-decisions per day, but the majority cannot actually tell you whether those decisions are generating positive ROI or quietly eroding it. They're flying without instruments.
Chief financial officers and chief information officers are increasingly demanding measurable business outcomes from AI investments — not just cost reduction narratives, but demonstrable, quantifiable value. The conversation has matured from "AI will save us money" to "show me the number." And right now, most enterprises cannot show the number.
The Scale Trajectory Makes This Urgent
Before anyone concludes this is a problem for next year's planning cycle, consider the adoption curve. Gartner projects that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. By that same year, an estimated 15% of day-to-day work decisions at large organizations will be made autonomously by AI agents.
That's not a gradual wave. That's a wall. And it's approaching while most enterprises are still figuring out how to measure what their current agentic systems are actually doing — let alone costing.
The cost variability problem compounds at scale in ways that are genuinely unpredictable. Because expenses depend on reasoning complexity, tool selection, and retry counts per task, a single process change — a new data source, a modified approval workflow, a model update — can shift your monthly AI operating costs by a meaningful percentage without any visible trigger. This is not a billing anomaly. It's a structural feature of autonomous systems that IT alone cannot manage without CFO-level visibility into what "acceptable cost variance" even means for that organization.
The Community Is Noticing
The conversation isn't staying inside boardrooms. Practitioners across the industry are picking up on these dynamics, and the discourse on social platforms reflects a growing awareness that the "just ship it" era of agentic AI is giving way to something more rigorous:
The signal from practitioners is consistent: the questions getting asked in 2026 are fundamentally different from the ones being asked in 2024. The technology curiosity phase is closing. The accountability phase is opening.
What Good Implementation Actually Looks Like Now
So what does cost-aware agentic AI architecture look like in practice? A few principles are emerging as non-negotiable for enterprises serious about sustainable deployment:
1. Right-size your reasoning model to the task. Using an advanced reasoning model for a routine classification task is the AI equivalent of hiring a neurosurgeon to take your blood pressure. The output is fine; the cost is absurd. Matching model tier to task complexity is one of the highest-leverage cost levers available.
2. Design verification workflows before you design generation workflows. Since verification accounts for the majority of operating costs, the question "how will we validate this output?" should precede "how will we generate this output?" in every system design conversation.
3. Treat context as inventory, not infrastructure. Every token of context sitting in an agent's working memory costs money to maintain and process. Pruning, compressing, and managing context windows deliberately — rather than letting them grow unbounded — is a genuine cost engineering discipline, not a nice-to-have.
4. Build measurement before you build scale. If you cannot currently measure the business impact of your AI decisions, adding more autonomous agents doesn't compound your value — it compounds your uncertainty. Instrument first, scale second.
5. Bring CFO constraints into the architecture room. This is the meta-lesson. If the only people designing your agentic system are engineers optimizing for capability, you will build something impressive that your finance team cannot sustain. The economics conversation belongs in the design phase, not the post-mortem.
The Bottom Line
Agentic AI is not going away. The capability is real, the adoption trajectory is steep, and the use cases are genuinely transformative. But the economic model that enterprises carried into 2025 — built on chatbot-era assumptions about token costs and verification overhead — is no longer fit for purpose.
The organizations that will win the next phase of AI deployment are not the ones with the most ambitious pilots. They're the ones that designed their systems with an honest answer to a question that now belongs as much to the CFO as to the CTO: What does it actually cost to get this right, at scale, every day?
The inflection point is here. The question is whether your architecture was designed to meet it — or whether you're about to find out the hard way what 1,000x token overhead feels like on a quarterly budget review.
Published in Stream · Dispatch #457 · July 19, 2026 · 8 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.