TL;DR
- Agentic AI costs are driven far more by verification, context management, and orchestration overhead than by the model itself.
- McKinsey reports that nearly 60% of agentic AI operating costs go toward verifying and refining outputs—not generating them.
- Object Edge's real-world 30-day test logged 457 autonomous runs but required 256 governance checks, a ratio that signals how heavy the hidden overhead really is.
- The era of subsidized token pricing is ending; enterprises that don't architect for cost now will feel the shock in production—on the CFO's invoice.
The Bill Nobody Budgeted For
Picture the scene: a well-funded enterprise team spends months selecting the right foundation model, negotiates favorable API pricing, and ships their agentic AI system to production. Everyone applauds. Then the first full month of operating costs arrives, and suddenly the CFO is asking questions nobody prepared answers for.
This is not a hypothetical. It is increasingly the pattern McKinsey documented in its July 2026 report on agentic AI economics, and it is the uncomfortable gap between what gets measured during a proof-of-concept and what actually runs up the tab in production.
The surprise is not the model. It is everything that wraps around the model.
What McKinsey Actually Found
McKinsey's July 2026 analysis draws a line that should stop every IT and operations leader mid-slide-deck: nearly 60% of agentic AI operating costs are spent on verification and refining responses, not on generating them. Let that settle for a moment. The capability everyone is buying—the generative intelligence—accounts for a minority of the cost. The majority is spent on the scaffolding: checking outputs, retrying failed steps, managing context windows, and orchestrating tool calls.
The report goes further. Agentic tasks can consume roughly 1,000 times more tokens than equivalent chat-based AI interactions. The culprit is context. Where a chat interface handles discrete, bounded exchanges, an agent maintains rolling context across extended workflows—every prior action, retrieved document, and tool response potentially feeds back into the next step. That context accumulates fast, and tokens are how you pay for it.
McKinsey also punctures a common planning assumption: per-token pricing is no longer a reliable predictor of actual cost. Real costs vary based on reasoning path complexity, how often tools are called, retry frequency, and even the language of the input. Non-English text fragments into more tokens per unit of meaning than English, quietly inflating costs for global deployments. You cannot budget for an agent by multiplying a token rate by an estimated query volume. The variables multiply too.
The report's most quotable observation cuts to the core of why this matters now:
"The decision to scale an agent is increasingly becoming a complex and fast-changing economics decision, not a technical one."
— McKinsey, July 2026
CFOs and CIOs are already arriving at this conclusion from the other direction—demanding measurable ROI—but most enterprises still lack the mechanisms to measure AI's full operational expenditure. They are flying into cost territory without instruments.
Object Edge's Field Test: The Governance Math
If McKinsey provides the framework, Object Edge's 30-day production benchmark provides the receipts.
From May 19 to June 18, 2026, Object Edge ran its Sayya AI agent autonomously across five operational domains—executive briefing, initiative governance, pipeline intelligence, task hygiene, and systems maintenance—without a single manual prompt from a human operator. The published instrumented results logged 457 autonomous runs, 256 governance checks, and 144 data syncs.
The zero-manual-prompts headline is genuinely impressive. But the governance figure is the one that deserves a second look. At 256 governance checks against 457 task runs, the compliance and auditability layer ran at roughly 56% of total task volume. For every two things the agent did autonomously, it triggered more than one governance check. That is not a flaw in the design—for regulated industries, that auditability-first architecture is exactly right. But it is a cost signature that most procurement and IT teams are not modeling in advance.
The system executed 15 playbooks and 31 scheduled workflows across integrations with Gmail, Google Calendar, Google Chat, Git, HubSpot, and Google Drive. It is a genuinely sophisticated deployment. It is also a concrete illustration of why "how much will it cost to run?" is a harder question than it looks.
Object Edge has been sounding this alarm in parallel. A June 2026 post framed AI tokenomics as an emerging engineering discipline requiring deliberate architectural decisions around semantic infrastructure, context-aware retrieval, and agent-level token budgeting. An April 2026 warning was even more direct: the era of subsidized token pricing from major AI providers is ending, and enterprises should build model portfolio strategies before costs normalize to reflect actual model economics.
The Knowledge Fragmentation Tax
There is a third cost driver that tends to slip under the radar in discussions about model pricing and orchestration overhead: enterprise knowledge fragmentation.
Object Edge has documented this extensively. When an AI agent operates across a business where content lives in Salesforce, SAP, Confluence, email archives, and shared drives—with no unified semantic layer—the agent does not magically bridge those silos. It burns tokens trying to. Fragmented knowledge inputs produce fragmented, lower-quality outputs that require more verification cycles, more retries, and more governance checks. The knowledge problem compounds the orchestration problem, which compounds the context problem. The costs are not additive; they interact.
The fix is not glamorous: it requires treating knowledge orchestration as a prerequisite infrastructure layer before agent deployment, not an afterthought. But it is one of the most significant cost-reduction levers available, and one that McKinsey also flags—alongside orchestration design and prompt engineering—as a place where significant savings are possible without compromising outcomes.
"You Are Still Responsible for Your Software"
There is a cultural dimension to this cost problem worth naming. The enthusiasm around agentic AI has, in some corners, drifted into a kind of magical thinking—a sense that if the agent can do it autonomously, the human is off the hook. Andrej Karpathy drew the line clearly, and it applies here:
"You are still responsible for your software, just as before. You are not allowed to introduce vulnerabilities because of vibe coding."
— @akshay_pachaar (citing Karpathy)
Substitute "unexpected operating costs" for "vulnerabilities" and the point holds just as well. Autonomous execution does not eliminate the cost of verification—it shifts when and where that verification happens. Designing it out of the picture during the build phase does not make it disappear; it just moves the bill to production, where it is harder and more expensive to fix.
This is precisely where the open-weight AI investment thesis also intersects the economics conversation. As one investor framing noted:
"#4 K3 strengthens the investment case for open-weight AI. Open models may struggle to charge premium licensing fees directly, but they can monetize through: API usage, enterprise private deployment, coding-agent subscriptions, cloud and inference services, fine-tuning and developer tools..."
— @gate_ventures
The model monetization landscape is shifting precisely because the cost structure of production AI is shifting. Enterprises hunting for cost relief will increasingly pressure-test open-weight alternatives, private deployments, and tiered model portfolio strategies—exactly what Object Edge's April guidance anticipated.
The Question Has Flipped
For years, the dominant enterprise AI question was "Which model should we use?" That question has not disappeared, but it has been joined by a harder one: "How much will verification, context management, and governance cost at scale—and who is modeling that before we deploy?"
This is where implementation partners, platform architects, and operations-focused teams create genuine value. Not by picking the best benchmark score on a leaderboard, but by building in cost visibility, context strategy, and orchestration efficiency before the system hits production. The enterprises that did this work in the design phase will not be shocked by their first full operating bill. The ones that skipped it will be having a very different conversation with their CFO.
McKinsey is right that this has become an economics decision. But economics decisions require economic visibility—and right now, most enterprises are scaling agents without the instrumentation to see what they are actually spending, or why.
The production shock is not inevitable. It is just what happens when the architecture conversation stays technical and the cost conversation gets left for later.
What to Do Before the Invoice Arrives
For operations and IT leaders currently evaluating or scaling agentic AI, a few practical starting points:
- Model your governance overhead now. Object Edge's 56% governance-to-task ratio is a real-world reference point. Build that into your operating cost projections, not just your capability assessments.
- Audit your knowledge layer. Fragmented content in siloed systems is a token-burn multiplier. Unifying semantic infrastructure before deployment reduces downstream costs across every dimension.
- Stop relying on per-token pricing as your budget line. Model costs by workflow type, tool usage, retry rates, and context length. McKinsey's framework is the right starting frame.
- Build a model portfolio strategy. Subsidized pricing is a window, not a wall. Plan for the cost structure you will face in 12 months, not the one you are paying today.
- Instrument for cost visibility in production. If you cannot measure it, you cannot manage it—and right now, most enterprises cannot measure it.
The agents are getting better. The economics are getting real. The gap between those two facts is where the next wave of enterprise AI decisions will be made.
Published in Stream · Dispatch #459 · July 21, 2026 · 8 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.