In analyses published on August 25, 2026, McKinsey released "The State of AI in 2026: On the road to ROI" and related commentary that delivers a sobering message to companies: as AI evolves from simple chat models into agents capable of executing multi‑step tasks, corporate AI spending is rising rapidly and cost volatility is far greater than expected. The same task can cost up to 30 times more when handled by different agents or via different execution paths. About one‑third of organizations report dedicating more than 10% of their technology and communications budgets to AI; 60% of respondents plan to increase spending next year, and nearly one‑fifth say AI operational costs have already begun to constrain usage. Businesses are facing the reality of "more use, higher costs, and not necessarily proportional returns."
Details
McKinsey's survey shows that large enterprises are moving faster on agent deployments: among organizations with annual revenue over $1 billion, 40% report scaling use of AI agents (up from 27% last year), while the share among smaller organizations remained roughly flat. Code‑generation agents are becoming common: about one in five respondents (31% among large enterprises) are scaling their use. Nearly one‑third of organizations have even chosen to develop software in‑house using available agent coding tools, foregoing purchase of some software products or features.
The core of the cost pressure lies in the fundamental differences between agents and chat‑style AI. Chat models primarily generate text and have relatively controllable token consumption; agents must plan, call tools, observe results, and iteratively adjust, often undergoing multiple rounds of reasoning and external interaction. McKinsey notes that different agents or different execution paths of the same agent can produce up to a 30× cost variance for the same task. McKinsey Senior Partner Lari Hämäläinen likened it to "running a business where daily costs can vary by a factor of 30."
This volatility stems from uncertainty in task paths, model choice, reasoning depth, multi‑agent collaboration patterns, and the degree of human review. Code‑focused agents are particularly token‑hungry and can be hotspots for billing sensitivity. Meanwhile, about 20% of respondents say AI‑related operational costs (including tokens) have already limited usage, even though most companies still plan to increase investment.
The report also acknowledges agents' efficiency potential: for some transformation office tasks, agents can reduce human time by 35%–40%, and sometimes by more than 70%. The key is whether companies can accurately measure whether "time saved and output value" offsets the highly variable operating costs. Today, only about 6% of respondents are "AI high performers" (AI contributing materially to EBIT, at 5% or more), and the corporate‑level financial impact is roughly unchanged from last year, indicating that value realization at scale still lags investment enthusiasm.
Potential impacts
1. From "can we use it" to "can we afford it and account for it"
Companies can no longer select solutions solely on capability; they must establish task‑level cost accounting and ROI evaluation. If the same objective can have a 30× cost difference, system architecture (single agent vs. multi‑agent), model routing, caching strategies, and thresholds for human intervention become central governance issues. Traditional IT budgeting models may not apply directly; AI FinOps (cost operations) may become a new essential.
2. Coding and automation scenarios hit first
Software development teams use coding agents heavily, consuming many tokens with variable execution paths, making bills easiest to lose control of. Some companies have already set usage limits. Firms need to distinguish tasks suitable for full agent autonomy from those suitable only for assistance, avoiding excessive compute consumption on low‑value tasks.
3. Investment enthusiasm versus return gaps may widen
Sixty percent of companies plan to increase AI investment, but the share achieving meaningful EBIT contribution remains low. If costs continue to climb without validated value, boards and finance teams may tighten budgets, causing some pilots to be halted or scaled back. The gap between high‑performing organizations and the rest may widen further.
4. Vendor and architecture choices will be affected
More affordable and predictable models and platforms will be favored; vendors that provide clear cost visibility, routing optimization, and caching mechanisms will have an advantage. Internally, firms may increasingly adopt hybrid architectures of "model routing + rules engines," delegating simple decisions to low‑cost models or deterministic code and reserving complex tasks for high‑capability agents.
5. Chain reactions across the AI industry
The agent boom drives explosive usage, but cost uncertainty may curb some large‑scale deployments and, conversely, push vendors to accelerate efficiency improvements (lower token prices, better caching, more precise effort control). In the long run, solutions that demonstrate controllable unit‑cost‑per‑value will prevail.
Conclusion
McKinsey's "The State of AI in 2026" does not deny the value of agents, but it warns companies that the leap from chat to agents brings not only capability gains but a fundamental change in cost structure. A 30× cost variance means that "enthusiastic deployments" must yield to rigorous economic accounting. Only by measuring agents' time savings and business value on the same ledger as their highly variable operating costs can companies avoid being consumed by their own AI enthusiasm and genuinely move toward ROI.
▶️ Enjoyed this piece? Creating content is hard!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com