Is That AI Agent Worth It? McKinsey's New Discipline of Agentic Economics

Inference got radically cheaper, and enterprise AI bills went up anyway. That's the paradox at the center of a new McKinsey Quarterly framework on what it calls agentic economics — and the numbers behind it are stark enough to reset how CEOs think about deploying agents.

According to Stanford HAI, the cost of GPT-3.5-level capability fell from $20 to $0.07 per million tokens through 2024 — a 280x drop. Over that same window, Menlo Ventures found enterprise LLM spending tripled in the twelve months to end of 2025. 93% of surveyed enterprises report exceeding their AI budgets, one-fifth have already throttled AI use because of operating costs, and AI spend is on track to reach roughly 25% of enterprise IT budgets. As Pay-i CEO David Tepper puts it: "Tokens are not value; tokens are the bill."

Why Cheaper Tokens Didn't Mean Cheaper AI

McKinsey's own infrastructure shows the scale of the shift: the firm now processes roughly 5 trillion tokens per month as of May 2026, with about 10% of users driving 65% of consumption. The old assumption — that falling per-token prices would make AI cheap by default — only held for chat. Agents broke it, for six specific reasons:

1. Long-lived context. Agents resend accumulated context at every step of a task, so agentic workloads can consume roughly 1,000x more tokens than an equivalent chat interaction.
2. Refinement is the real sink. Roughly 60% of an agentic task's total cost goes to checking, repairing, and reverifying its own work — not producing the first answer.
3. Autonomy creates variance. The identical task can cost 30x more in one run than another. Agent cost is a distribution, not a fixed price.
4. Expensive reasoning on cheap tasks. Extended thinking modes are pure overhead when applied to simple work, which makes routing decisions a direct cost lever.
5. Orchestration compounds cost. How a task is decomposed and handed between agents changes total spend dramatically — a poorly designed handoff chain can multiply cost with no gain in quality.
6. Information structure matters. Prompt formatting, verbosity, and even language choice (non-English text tokenizes less efficiently) all move consumption. McKinsey found that concise, direct "caveman language" prompting cut some workflows' token use by 30–40% with no quality loss.

Governing Machine Work Like Capital

McKinsey's prescription for CEOs treats agent deployment less like a software rollout and more like a capital allocation problem:

• Allocate intelligence like capital — not every task gets the most capable, most expensive model.
• No autonomous system runs without a defined mandate, budget, and stopping rule. Agents that can act indefinitely need an explicit cap, not an implicit one.
• Measure cost per completed business outcome, not per token. Token counts are an input metric; they don't tell you if the work was worth doing.
• Treat contextual data capture as a strategic asset. The context an agent has access to increasingly determines how well — and how cheaply — it performs.
• Rethink outsourcing. Agents perform best with proprietary context kept in-house, which cuts against reflexively offshoring workflows that used to be commodity labor.

The Real Competitive Axis

This is the agent hype cycle landing on the CFO's desk. Falling token prices were supposed to make AI adoption a formality; instead, agents absorbed the savings and then some, because they consume tokens in loops, retries, and self-verification rather than single clean requests. As frontier and mid-tier models converge in raw capability — see Fable 5 at the expensive end and Sonnet 5 commoditizing the cheap end — the model itself stops being the differentiator. Governance of machine work becomes the moat: routing the right task to the right model, capping runaway agents, and measuring outcomes instead of tokens.

That governance question has a data half too — an enterprise can have perfect routing discipline and still lose the plot if agents can't reliably and safely reach the content they need to do the work in the first place. That's exactly the gap Box's new report on enterprise content quantifies, and it pairs directly with McKinsey's own 20,000-agent internal deployment as a real-world stress test of this framework.

The winners of the next phase of enterprise AI won't be whoever has access to the best model — that access is becoming a commodity. They'll be whoever builds the discipline to govern what machines do with it.