The next big enterprise AI question may not be “What can we automate?”
It may be: What does every successful AI outcome actually cost?
That question is becoming more urgent as organizations move from experimentation into production.
The amount of money flowing into AI certainly is not slowing down. Gartner forecasts worldwide AI spending will reach $2.59 trillion in 2026, up 47% year over year, with AI infrastructure expected to account for more than 45% of spending over the next several years.
But rising investment does not automatically translate into rising value.
Recent research reported by TechCrunch illustrates the gap. 74% of 150 enterprise IT professionals surveyed by Madrona plan to increase their AI budgets over the next 12 months, while the remainder expect spending to stay level. Yet those same organizations report that fewer than half of their AI pilots currently make it into full production.
That creates an increasingly important question for business leaders.
If AI investment continues to rise while many initiatives struggle to reach production, how can organizations identify which investments are genuinely worth scaling?
The answer increasingly comes down to economics.
At Gartner's 2026 Application Innovation & Business Solutions Summit in London, the 'hidden cost of AI' has become part of the agenda, reflecting growing attention on the economics behind copilots, agents and increasingly complex AI workflows.
A simple chatbot interaction may involve one prompt and one response. An agentic workflow can involve repeated reasoning, multiple tool calls, access to enterprise data and several intermediate steps before a task is complete.
Gartner predicts that inference costs per agentic workflow will increase more than fivefold through 2028. Routing a task to an agentic reasoning model can already increase provider inference costs by at least five times compared with a basic chatbot interaction, with costs potentially increasing further as task complexity grows.
That changes the unit economics of AI.
Scaling successfully now requires organizations to understand not simply whether a use case works, but whether the outcome it produces justifies the cost of producing it.
AI economics can look manageable during a pilot.
A small group of employees may test a copilot. A development team may trial an agent. A business function may connect an LLM to a limited dataset.
The real cost profile becomes clearer once that system is used repeatedly:
The scale of the change is already becoming visible in infrastructure spending. Gartner expects worldwide spending on AI-optimized infrastructure as a service to reach $42.3 billion in 2026, an increase of 96.4% from 2025. More importantly, spending on inference workloads is forecast to reach $23.3 billion, overtaking the $19 billion expected to be spent on training workloads.
That crossover matters.
Training happens periodically. Inference happens every time an employee, customer, application or agent uses the model.
As AI becomes embedded in everyday operations, inference increasingly becomes a recurring operating expense.
Gartner describes another complication as the 'Inference Paradox': although the economics of individual tokens are improving, increasingly sophisticated AI capabilities consume more tokens and involve more complex workflows. Better unit economics can therefore coexist with higher overall AI costs.
And there are already signs of businesses encountering this problem in practice.
TechCrunch reported in June that some organizations were facing unexpectedly rapid AI consumption, including a routine AI software contract renewal at Priceline reportedly coming back four to five times more expensive. The report highlighted how falling per-token prices have been accompanied by rapidly increasing consumption as businesses expand AI adoption and deploy more autonomous systems.
The lesson is not that organizations should reduce AI adoption.
It is that consumption needs to become visible before it becomes material.
Tokens are useful for measuring consumption.
They are less useful for measuring business value.
A board is unlikely to care that an AI system consumed 500 million tokens if nobody can explain what the business received in return.
The more useful question is: What did those tokens achieve?
This also reflects a broader shift in how enterprises want AI to be priced.
TechCrunch reported research from Andreessen Horowitz involving 50 technical AI buyers which found that more than half wanted AI pricing tied to the work produced or outcomes achieved rather than usage measures such as token consumption.
That is an important signal.
The enterprise AI conversation is beginning to move from ‘how much AI are we using?’ toward ‘What are we getting for what we spend?’
That requires organizations to connect technical telemetry with business performance.
Tenth Revolution Group helps businesses establish those foundations across data, AI, cloud and transformation, connecting technical delivery with the operating models and capabilities required to turn investment into measurable outcomes.
One of the most immediate opportunities to improve AI economics is to stop treating every task as if it requires the most capable model available.
Some business problems genuinely need frontier-level reasoning.
Others do not.
A straightforward classification task, information extraction workflow or internal knowledge query may be handled effectively by a smaller model at significantly lower cost.
Gartner argues that there is no economical one-size-fits-all model for advanced AI. Instead, increasingly sophisticated AI products are likely to require multimodel ecosystems that route work according to the level of intelligence required.
That can mean combining:
The principle is relatively simple.
Use enough intelligence for the problem, rather than the most intelligence available.
This can make AI architecture an important financial lever.
Model choice, routing, caching, quantization and infrastructure utilization can all change the cost of producing the same business outcome.
There is a risk that focusing on AI economics turns into a simple cost-cutting exercise.
That would miss the point.
The goal is not necessarily to make every AI system cheaper. It is to understand whether spending more produces proportionately more value.
Deloitte's 2026 State of AI in the Enterprise research provides useful context here. 66% of organizations report productivity and efficiency gains from enterprise AI, while 40% report cost reductions. Yet only 20% currently report increased revenue, despite 74% hoping AI will contribute to revenue growth in the future.
That gap matters.
Organizations are already seeing operational benefits from AI, but converting those benefits into broader commercial outcomes remains a work in progress.
A more capable model that costs twice as much may be the right decision if it produces substantially greater business value.
A cheaper model is not economical if its outputs create more manual review, more errors or poorer customer outcomes.
AI economics therefore needs to consider both sides of the equation: cost and value.
This changes the role of technology architecture.
Infrastructure choices have traditionally been assessed using measures such as performance, reliability, security and availability.
Those measures remain critical.
But AI introduces another dimension: the economics of every interaction.
Technology leaders increasingly need to understand how decisions around models, GPUs, cloud platforms, data architecture and orchestration affect the cost of delivering a business result.
This is particularly important as agentic AI grows.
An agent that completes a task through 15 reasoning steps may produce an excellent result. But if another architecture produces an equivalent result in five, that difference becomes commercially significant when the workflow is executed millions of times.
Architecture therefore becomes part of the business case.
The organizations best positioned to manage this will be those that connect engineering, data, cloud, finance and product expertise rather than optimizing each discipline independently.
There is another cost that is easier to overlook.
Capability.
Organizations can buy access to models and infrastructure relatively quickly. Developing the expertise required to use them effectively takes longer.
Teams need to understand:
And increasingly, they need to understand how those areas interact.
The cost of poor architecture is not simply an expensive cloud bill. It can mean building the wrong solution, scaling an uneconomical workflow or investing heavily in a use case that never delivers sufficient value.
That is why Tenth Revolution Group approaches AI transformation through Talent, Training and Transformation.
Organizations may need specialist expertise to solve an immediate challenge, workforce development to build sustainable internal capability or practical project delivery to take AI from strategy into production.
The right answer will often involve all three.
The wrong AI metric can create the wrong behavior.
If teams are rewarded for adoption alone, they may increase usage without increasing value.
If they are judged only on lower token consumption, they may reduce costs while damaging performance.
If they focus only on model accuracy, they may ignore the economics of delivering that accuracy at scale.
The better question is whether the system delivers a valuable outcome at a sustainable cost.
For each production AI use case, organizations should understand:
Define the business result before measuring consumption.
Include model usage, infrastructure, data, orchestration and operational support.
Avoid sending straightforward tasks to expensive models when a lower-cost alternative can deliver the required result.
The TechCrunch findings are an important reminder that moving from pilot to production remains difficult. Economics should therefore be tested at realistic volumes, not extrapolated from limited experiments.
Routing, caching, quantization, model selection, infrastructure utilization and stronger data context can all change the economics significantly.
Enterprise AI investment is growing rapidly.
So is the pressure to demonstrate what that investment actually delivers.
With Gartner forecasting $2.59 trillion in worldwide AI spending this year, organizations have moved beyond the point where AI economics can be treated as an infrastructure detail.
The organizations that scale successfully will not necessarily be those with the biggest budgets or the most AI tools.
They will be the ones that understand the economics of their AI systems at a granular level:
AI is becoming more capable. The next stage of maturity is making sure its economics become more intelligent too.
Tenth Revolution Group helps organizations combine Talent, Training and Transformation to build cloud, data and AI capabilities that are scalable, measurable and commercially sustainable.