Emerging tech & Deep tech • 2 days ago • Shruti Das
For the first two years of the generative AI boom, enterprise strategy was largely defined by experimentation. Companies launched copilots, built internal chatbots and tested large language models across functions ranging from customer service to software development. The underlying question was straightforward: Where can AI be useful?
That question is now evolving.
As AI agents move from demonstrations and pilots into real business workflows, enterprise leaders are confronting a more demanding challenge: Can these systems create measurable value at an economically sustainable cost? Recent research and industry developments suggest that this is becoming one of the defining issues of the next phase of enterprise AI adoption. McKinsey, for example, argues that agentic workflows are introducing complex and sometimes unexpectedly expensive economics, even as the unit cost of AI models continues to fall.
The shift matters because an AI agent is not simply another software licence. It is an active system that consumes compute, calls models, retrieves information, uses tools and may take different paths to complete the same task. As enterprises begin deploying these systems at scale, the conversation is moving beyond technological capability and towards something far more familiar to business leaders: return on investment.
The era of AI experimentation is giving way to accountability
The first wave of enterprise generative AI rewarded speed. Organisations were under pressure to demonstrate that they were not being left behind, and experimentation became a strategic priority. In many cases, success was measured by relatively simple indicators: the number of pilots launched, employees given access to AI tools or use cases identified across the business.
Those measures are becoming less meaningful.
An enterprise can deploy dozens of AI agents and still struggle to demonstrate material business value. In fact, recent research points to a growing distinction between organisations that are actively measuring the impact of agentic AI and those that are simply deploying it. Contentstack’s 2026 research found that fewer than half of surveyed enterprises had clearly defined KPIs and were actively measuring their agentic AI programmes.
That distinction could become increasingly important. Technology adoption creates activity; it does not automatically create outcomes. An AI agent answering more customer queries, generating more code or producing more reports may look productive, but those outputs only matter if they translate into something the business actually values—lower costs, faster operations, higher revenue, improved customer experience or reduced risk.
This is why the next phase of enterprise AI will require a more disciplined mindset. The question is no longer simply whether an organisation can deploy an agent. It is whether the organisation can clearly explain what that agent is doing, what it costs and what business outcome it is improving.
AI agents have a fundamentally different cost structure
Traditional enterprise software generally comes with relatively predictable economics. A company might pay per user, per licence, per month or according to a negotiated infrastructure contract. Finance teams can usually estimate expenditure before widespread deployment.
Agentic AI is different.
The cost of completing a task can vary depending on the model selected, the amount of information retrieved, the number of reasoning steps involved, the tools an agent invokes and whether it needs to retry part of a workflow. McKinsey notes that agentic systems can follow different paths to complete similar tasks, meaning costs can behave less like a fixed unit price and more like a variable distribution.
This creates an important paradox for enterprises. AI models may become cheaper on a per-token basis, while the overall cost of using AI continues to rise. As organisations give agents more autonomy and ask them to handle increasingly sophisticated workflows, the systems may consume more intelligence rather than less.
The economics, therefore, cannot be understood by looking at the price of a single model call.
A customer-service agent, for example, may need to interpret a request, search a knowledge base, access a CRM system, compare information across multiple sources, call another specialised agent and generate a final response. Each of those actions may introduce additional computational and infrastructure costs. By the time the interaction is complete, the relevant financial question is not the cost of one token.
It is the cost of successfully resolving the customer’s problem.
From cost per token to cost per outcome
This may be the most important measurement shift facing enterprise AI. Technical metrics such as token consumption, inference cost and API usage remain useful for engineering teams. But they are not sufficient for executives deciding whether an AI initiative deserves further investment. A CIO cannot take a token-cost report to the board and reasonably expect it to answer whether the organisation’s AI strategy is working.
Business outcomes are what matter.
For an enterprise deploying AI agents, the more meaningful questions might include:
- What is the cost per successfully completed transaction?
- How much time is being saved compared with the previous process?
- How often does a human need to intervene?
- Has the workflow improved customer satisfaction or conversion?
- What is the cost of an AI-assisted process compared with its human alternative?
- Does the system become more or less economical as usage scales?
Research into agentic AI measurement increasingly supports this broader approach. Among enterprises actively measuring agentic programmes, common KPIs include productivity, operational cost reduction, time saved, customer satisfaction, conversion improvements, reliability and revenue influence.
The implication is significant: AI ROI cannot be reduced to AI cost.
A highly capable agent may be relatively expensive to operate but still deliver exceptional value if it eliminates a major operational bottleneck. Conversely, a very cheap agent may create little value if employees spend substantial time reviewing and correcting its work.
The objective should not be to build the cheapest possible AI system. It should be to build the most economically productive system.
Why AI FinOps is emerging as a strategic capability
Cloud computing created FinOps because organisations eventually realised that cloud flexibility could also create uncontrolled spending. Teams could provision infrastructure quickly, but without visibility and accountability, costs could escalate just as quickly.
AI may now be heading towards a similar moment.
The emerging discipline is broader than simply tracking model invoices. Enterprises need visibility across the full AI operating stack, including model usage, orchestration, monitoring, security and the surrounding services required to run production-grade agentic systems. McKinsey has highlighted that some rapidly growing AI costs can sit outside the model itself, including gateways, orchestration platforms and monitoring layers.
This is where a form of AI FinOps becomes strategically important.
The goal is to connect technical consumption with business accountability. Finance needs to understand where money is being spent. Technology teams need to understand which systems are consuming resources. Business leaders need to know whether that expenditure is generating measurable outcomes.
That requires a common operating model.
Without one, enterprises risk creating a familiar problem: individual teams deploy agents independently, budgets become fragmented and the organisation eventually discovers that it has significant AI expenditure without a clear picture of which investments are delivering value.
Governance is becoming an economic issue
AI governance is often framed primarily around security, privacy and compliance. Those concerns remain critical, particularly as agents gain access to enterprise systems and become capable of taking autonomous actions.
But governance also has an economic dimension.
Poorly governed AI systems can create unnecessary costs through duplicated workflows, excessive retries, inappropriate model selection and uncontrolled agent proliferation. Gartner has warned that enterprises are likely to face growing agent sprawl as autonomous systems spread across organisations, creating management complexity alongside security and operational risks.
The challenge is not simply to restrict every agent.
Overly restrictive governance can prevent useful systems from delivering value, while insufficient controls can create unacceptable operational and financial risk. Gartner has argued that governance should reflect differences in an agent’s autonomy and access rather than applying a single uniform approach to every system.
For enterprises, this means governance increasingly becomes part of ROI management.
An organisation that understands what its agents can access, how they behave, how frequently they act and what outcomes they produce is in a much better position to optimise both risk and cost. Observability is therefore no longer just an engineering concern. It becomes an essential management capability.
The biggest problem may be measuring value, not creating it
One of the more interesting questions emerging from enterprise AI is whether some organisations genuinely have poor ROI—or simply poor measurement.
AI value can be difficult to capture in traditional financial reporting. An agent might reduce the time required to complete a process without immediately reducing headcount. It might improve decision quality, prevent an operational failure or enable employees to focus on higher-value work.
Those benefits are real, but they can disappear into the organisation if nobody has established a framework for measuring them.
This is particularly important because the greatest economic gains from AI may not always come from straightforward labour savings. McKinsey’s recent work on the economics of AI argues that value can also emerge through faster decisions, better use of existing assets and opportunities that organisations might otherwise miss.
That means enterprises need to establish measurement frameworks before scaling an AI initiative—not after.
Every significant agent deployment should ideally begin with a baseline. What does the workflow cost today? How long does it take? What error rate does it produce? How much revenue does it influence? Without those reference points, proving improvement later becomes considerably more difficult.
The winners will not necessarily deploy the most agents
There is a temptation to treat the number of AI agents inside an enterprise as a measure of maturity. The more autonomous systems an organisation has, the more advanced it must be. That assumption is likely to prove misleading.
The next generation of successful AI-native enterprises may be distinguished not by the number of agents they operate, but by the quality of their decisions about where agents should exist in the first place. The most valuable opportunities are likely to be workflows with meaningful scale, measurable outcomes, accessible data and a clear connection to a business objective. This requires a more selective approach to deployment.
Some processes may benefit enormously from autonomous systems. Others may be better served by a copilot that keeps humans in control. Still others may simply not justify the cost or complexity of AI at all.
That kind of discipline may sound less exciting than an enterprise-wide race to deploy agents everywhere. But it could ultimately create a more sustainable competitive advantage. The companies that win the AI race may not be those with the largest collection of agents.
They may be the companies that become exceptionally good at deciding which work should be automated, how that automation should be measured and when an AI system is no longer worth paying for.
The new enterprise AI maturity model
Enterprise AI is beginning to move through a familiar technology lifecycle. First comes experimentation, then comes adoption. After that comes the harder work: operational discipline.
The organisations entering this phase will need to develop capabilities that extend well beyond model selection and prompt engineering. They will need stronger measurement systems, AI-specific cost management, governance frameworks, observability and clear ownership of business outcomes. This is ultimately the heart of the AI agent ROI reckoning.
Intelligence is becoming easier to access. Models will continue to improve, and agents will become increasingly capable of navigating complex enterprise workflows. But abundant intelligence does not automatically create abundant value.
Value has to be designed into the operating model.
The next great enterprise AI challenge, therefore, is not simply building smarter agents. It is building organisations capable of managing them intelligently.
Conclusion
The enterprise conversation around AI agents is changing from possibility to proof.
For technology leaders, this represents an important turning point. The success of the next phase of AI adoption will not be determined solely by model performance or the speed of deployment. It will increasingly depend on whether organisations can connect AI activity to measurable business outcomes while maintaining control over cost, risk and complexity.
That is why the emerging discipline around agentic economics deserves attention. Enterprises are discovering that an AI agent is not a one-time technology purchase. It is an ongoing economic system whose value depends on how it is designed, governed and measured.
The question facing CIOs and business leaders is becoming increasingly simple:
Not “What can our AI agents do?”
But “What are they worth?”
Key Takeaways
- Enterprise AI is moving beyond experimentation. The focus is shifting from deploying agents to proving measurable business value.
- Agent economics are fundamentally more complex than traditional software economics. Costs can vary based on workflow paths, model usage, tool calls and retries.
- Cost per token is becoming an insufficient business metric. Enterprises need to measure cost per successful outcome and connect AI spending to operational or commercial results.
- AI FinOps is emerging as an important capability. Organisations will need stronger visibility into AI consumption across models, infrastructure, orchestration and supporting services.
- Governance is also an economic issue. Poor visibility and agent sprawl can create unnecessary operational complexity and uncontrolled spending.
- The winners will be selective, not simply aggressive. The strongest enterprises may be those that identify the highest-value workflows and scale AI with discipline.
- The next phase of AI maturity is accountability. Deploying agents is becoming easier; proving that they are worth the investment is becoming the real competitive challenge.