As organizations accelerate adoption of agentic AI, they are finding the cost picture to be complex and sometimes surprisingly large. Although per-token prices have trended down, aggregate enterprise spending on AI is rising.
This piece was prepared by Chandana Asif, Dieter Kiewell, Lari Hämäläinen, Tom Kolaja, and Tunde Olanrewaju of QuantumBlack, AI by McKinsey. Their analysis highlights that pricing structures and the broader discipline of “tokenomics” are evolving rapidly, creating uncertainty for leaders who must decide which workflows to agentify.
Pricing today and why it matters
As of early July 2026, a frontier model such as OpenAI’s GPT-5.5 costs about $5 per one million input tokens and $30 per one million output tokens. By contrast, an earlier lightweight model cost roughly $0.20 and $1.25 per one million input and output tokens, respectively. These differences show that models vary markedly in capability and price, and that not every task requires a top-tier model.
Many organizations, however, do not yet make deployment decisions based on those trade-offs, and that can produce substantial cost consequences.
Concrete cost examples
Operational costs can be material. The authors’ analysis indicates that customer-facing AI agents in some banks can cost $20,000–$30,000 to run a single-agent workflow, while running a multiagent team might cost $100,000–$200,000. When those figures are multiplied by typical run counts and by the number of potentially agentifiable workflows — and combined with shifting AI pricing — it’s easy to lose control of costs and jeopardize value creation.
A practical illustrative case: using a conversational agent to onboard 2,500 new customers per year can cost about $10,000–$15,000; doubling the onboarding volume increased costs only to about $15,000–$20,000. This demonstrates how higher volumes can amortize fixed costs.
Using standard benchmarks, the total cost to complete an onboarding workflow per customer can decline from roughly $50–$150 to about $10–$30 when agents, deterministic systems, and human oversight are combined effectively.
How to evaluate agentic economics: three starter questions
Leaders who navigate agentic trade-offs well tend to ask three practical questions when planning deployments:
- Which workflows are the best candidates for agentic automation (considering volume, reuse, and value)?
- How do we calculate and reduce the total cost of ownership (TCO), including fixed and variable costs?
- What governance, operating model, and review cadence are needed to keep economics under control?
Total cost of ownership and levers to reduce spend
Managing agents requires a TCO mindset that accounts for visible model and token costs as well as less visible expenses: development, AgentOps, monitoring, compliance checks, model approvals, remediation, and ongoing human oversight.
While many teams optimize on model choice and token pricing because those are most visible, the more productive approach is often to redesign workflows to reduce exception rates and simplify human review steps that drive up costs.
Practical levers to lower token consumption and related costs include:
- model routing and rightsizing to match capability to task,
- caching and reuse of outputs,
- redesigning workflows to reduce inferences,
- batching requests and enforcing token budgets,
- committed-use pricing agreements.
Three main factors determine token costs: the chosen model’s quality (lightweight general model vs. specialist reasoning model), required latency, and task frequency (one-step lookup vs. open-ended looping job).
Which workflows favor agent economics?
Agent economics favor two characteristics: high volumes and reuse. Once fixed costs are in place, increasing agent run volumes helps amortize spending. Reuse—building agents once and applying them across multiple high-value workflows—also improves returns but requires disciplined identification of common tasks across priorities.
A single agent’s price tag alone does not determine whether a workflow pays off. What matters is tracking impact: the fully loaded cost to finish the job (for example, a completed onboarding, a resolved claim, a closed sale) combining humans, agents, and deterministic systems relative to the value delivered.
A completed customer onboarding may require five to seven agents, multiple deterministic systems, and two to four oversight teams. This simplified view underscores the support needs and cost drivers leaders must include when forming an accurate ROI picture.
Governance, operations and continuous reassessment
Agent economics continue to evolve: a workflow that is uneconomic today may become attractive months later, and vice versa. Leaders should therefore establish a regular review process to reassess workflows as model capabilities, prices, and operating practices change.
Agentic systems typically require ongoing adjustments—production agents often need tweaks every few days as new foundation models, context protocols, enterprise integrations, and business requirements appear. In some cases it will be appropriate to retire an agent and build a replacement.
Organizations must also budget for AI gateways, AgentOps, monitoring, compliance checks, model approvals, and remediation to maintain security and regulatory compliance. Changes in regulation or internal risk appetite (for example, restrictions on which frontier models are permitted) can force reevaluation, modification, or pause of agents.
Central teams, domain capabilities and financial disciplines
Business leaders should avoid building everything from scratch. Many AI capabilities are proven and easy to use. Leaders deciding what to build, partner with, or buy should keep strong ties to a central team that can:
- continuously evaluate models, tools, and providers; negotiate commercial terms; identify lower-cost alternatives; and oversee provider risk, and
- maintain reusable capabilities and enforce guardrails so domain teams reduce tool and agent fragmentation.
Just as organizations built FinOps to manage cloud spend, business units now need an equivalent discipline to manage agentic spend. That requires a cross-functional capability to continuously manage spend and reallocate tasks.
To manage agentic programs over time, leading organizations should incorporate AI unit economics into Quarterly Business Reviews (QBRs). QBRs give leaders transparency into where AI creates the most value, which workflows justify spending, and which should be redesigned, consolidated, or shifted to lower-cost models or enterprise capabilities.
Looking ahead
We are still in the early stages of the AI economics journey. Frontline leaders should focus on today’s trade-offs when implementing agentic workflows, but also prepare to answer second-order questions about market dynamics, organizational impact, and long-term business-model effects.
Those business-line leaders and domain owners who master unit economics—measuring costs, prioritizing reuse, and deciding when to build versus buy—will be best positioned to capture the significant prize that agentic AI can offer.
Acknowledgments
The authors thank Brandon Shi, Keisha Rastogi, and Varsha Seshadri for their contributions. The article was edited by Barr Seitz, editorial director in the New York office.
