As generative AI moves from experimental pilots into full-scale enterprise deployment, the conversation among business leaders is shifting from "Can AI work?" to "Can we afford how AI actually works?" A new McKinsey report reveals that the economics of AI agents, not the technology itself, will define the next phase of adoption. The findings show that organizations must fundamentally rethink how they measure AI value and manage costs as they scale agentic systems across their operations.

What's Driving the Shift From Technology to Economics?

The first two years of generative AI adoption were largely about access, experimentation, and proof-of-concept deployments. Companies rushed to explore what AI could do, building confidence in the technology and identifying high-impact use cases. But as organizations move beyond pilots and deploy AI agents at scale, financial sustainability and return on investment have become the dominant concern.

This shift reflects a maturation in how enterprises think about AI. Early adopters focused on capability and speed; today's leaders are asking harder questions about the true cost of producing reliable AI outcomes. The economics matter because they determine whether AI investments actually improve the bottom line or simply add new expenses.

Where Is Your AI Budget Actually Going?

The McKinsey research reveals a striking cost breakdown: 60% of agentic AI spending goes to response refinement. This means that the majority of AI agent costs are not tied to running the model itself, but to the work required to make AI outputs reliable enough for business use. Response refinement includes human review, corrections, retries, and rework when AI results fall short of quality standards.

This finding challenges a common assumption in enterprise AI: that cheaper models automatically deliver better return on investment. A lower-cost model may have cheaper tokens, but getting great results may require more attempts, more time, or more human review. A more capable model may have more expensive tokens, but complete the same task in one pass, reducing the total cost of producing a successful outcome.

OpenAI proposes a metric called "Useful Intelligence per Dollar" that answers four critical questions about whether your AI spending actually creates business value:

• Work Completion: Is AI completing work that matters to the business, such as resolving customer issues, shipping code changes, or reviewing contracts?
• Cost Per Outcome: What does each successful task cost, including all human review, retries, and rework, not just token expenses?
• Dependability: Can people depend on the result, or does it require significant correction and escalation?
• Scalability: Does each AI dollar produce more value as usage grows, or do costs rise faster than benefits?

How Are Leading Organizations Measuring AI Success?

The shift in how enterprises define success is dramatic. Organizations are moving away from measuring AI adoption through metrics like seats purchased or users active. Instead, they are tracking business outcomes: shortened cycle times, improved decision-making, freed employee capacity, and increased business resilience.

This change in measurement reflects a deeper organizational shift. Companies are asking how intelligence can become embedded into core business processes, not just applied to isolated use cases. For example, HR Path Brazil, a recruiting and talent management firm, embedded AI into its HR systems and achieved a 7% reduction in standard HR support cases while eliminating two hours of HR support workload each week.

The financial impact of this shift is substantial. Organizations investing in AI expect an average return of 21% this year, increasing to 38% in two years, according to Oxford Economics research. As agentic AI continues scaling, it is projected to deliver $17.6 million in returns, more than quadrupling last year's estimates of $4.3 million.

Steps to Calculate and Optimize Your True AI Cost Per Outcome

• Define Clear Success Metrics: Start with one workflow and define what "done" means in measurable terms. For a support team, "done" might mean a customer issue resolved. For engineering, it might mean a code change that passes tests. For legal, it might mean a contract reviewed accurately and on time. Measure outcomes in the system where the work actually happens.
• Calculate the Full Cost of Outcomes: According to OpenAI's framework, add up all costs to complete work well, including model costs, compute, employee time, human review, retries, and rework. Count only the tasks that met your quality bar, then divide total cost by successful tasks. This reveals the true cost per outcome, not just cost per token.
• Track Dependability Across Three Categories: Monitor results that are "ready to use" without modification, results that "need correction" and require another attempt or human edits, and work that "needs escalation" because a person had to step in and finish it. These measures show whether AI is genuinely reducing work, not just automating it.
• Monitor Economics at Scale: Follow the same workflow over time and track how many tasks met your quality bar, the total cost of completing them, and the cost per successful task. If completed work grows faster than total cost while quality holds or improves, each AI dollar is producing more value.

What Does the Autonomous Enterprise Look Like?

The organizations creating the greatest long-term advantage will be those that use early AI successes to rethink how work gets done across the entire enterprise. This represents a shift from systems of record, which primarily captured transactions and automated routine work, to what experts call the "Autonomous Enterprise." In this model, intelligent systems provide context, surface recommendations, orchestrate work across functions, and help teams execute routine decisions with human oversight.

"The organizations creating the greatest long-term advantage won't just be the ones deploying the largest number of AI use cases. They'll be the ones that use those early successes to rethink how work gets done across the enterprise," noted Jan Gilg, global president of Customer Success and Americas at SAP.

Jan Gilg, Global President of Customer Success and Americas, SAP

This transformation does not happen overnight, and it does not replace the need for strong leadership, governance, or talented people. If anything, those become even more important. The journey mirrors how organizations now view early cloud transformation projects: not as the destination, but as the foundation for a fundamentally new way of operating and innovating continuously.

The key takeaway for enterprise leaders is clear: the next phase of AI adoption will be won not by those deploying the most AI use cases, but by those who understand and optimize the economics of AI agents. As response refinement costs consume the majority of AI spending, organizations that focus on reducing rework, improving dependability, and measuring true business outcomes will see the returns that early AI investments promised.