Illustration of a spending gauge with the needle near the limit, next to a usage line that climbs in steps, like the cost of an AI agent growing with each step.

AI agents that cost more than they deliver: nearly half of companies have already reined them in

Nearly half of companies (49%) have already scaled back, delayed or paused their AI agents because they cost more than they delivered. That is the finding of KPMG’s Global AI Pulse for the second quarter of 2026, a survey of 2,145 executives in 20 countries (KPMG, 2026). The underlying problem is visibility: only 35% fully see what it costs them to run their AI.

Launching an agent is no longer the hard part. The hard part is knowing what it costs to keep it running and whether that spending pays off.

What the data says

IndicatorFigureSource
Companies that scaled back, delayed or paused agents when their cost outweighed the value they generated49% (24% scaled back or narrowed deployment and 25% delayed or paused it)KPMG, 2026
Companies that questioned the decision but have not made changes yet22%KPMG, 2026
Companies for which cost and value remain aligned24%KPMG, 2026
Companies that fully see and monitor their AI costs35%KPMG, 2026
Companies with established ROI: those that see all their costs versus those that do not15% versus 3%KPMG, 2026

Source: KPMG, 2026. Online survey conducted from April 28 to May 25, 2026, of executives at organizations with revenues of at least $50 million, in countries including Brazil, Mexico, the United States, Spain and Germany.

According to KPMG, this does not reflect less confidence in AI. Companies are becoming more selective and concentrating investment where the expected return is highest (KPMG, 2026).

Many companies do not know how much they spend on AI

Cost visibility is the weak spot (KPMG, 2026):

  • 35% fully see their AI costs and actively monitor them.
  • 42% see them only in part.
  • 13% find out only when the bill arrives.
  • 8% barely see them at all.

The difference shows in the results: companies with full visibility are five times more likely to report established ROI on their AI investment (15% versus 3%). In addition, a third of executives say they lack skills in AI costs and economics to deploy agents, and 29% struggle to understand and control operating costs as systems scale (KPMG, 2026).

Why does the cost of an AI agent grow?

Unlike a fixed-price license, most AI models are billed by consumption, that is, by the amount of text they process and generate, measured in tokens. KPMG notes that this usage-based pricing is becoming more common and that many companies still cannot forecast, monitor or manage that spending (KPMG, 2026). An agent multiplies consumption for several reasons:

  • It makes many calls per task. To resolve a single request, an agent may plan, use tools, review its result and try again. Each step is a call to the model.
  • It carries a lot of context. If you send it the full history, documents and instructions at every step, you pay for that text again and again.
  • It uses the most expensive model for everything. Many simple tasks, such as classifying an email, do not need the most powerful model.
  • It can get stuck in a loop. Without a limit on steps, an agent that cannot find the answer keeps trying over and over.
  • Usage grows with success. If the agent works, more people use it and the cost rises with them.

That is why knowing the model’s price is not enough. Falling prices for the most advanced models help (Wasivia, 2026), but total spending depends on how the agent is designed.

Which AI cost controls companies use

ControlCompanies that have itSource
Cost review when approving AI projects54%KPMG, 2026
AI cost-monitoring dashboards53%KPMG, 2026
Usage or token budgets40%KPMG, 2026
Architecture or prompt design standards39%KPMG, 2026
None of the above5%KPMG, 2026

Source: KPMG, 2026.

KPMG also finds that access to lower-cost, high-quality models is the fastest-rising influence on AI strategy: up 7 points from the previous quarter (KPMG, 2026).

Seeing the cost is not enough: you have to weigh it against value

The next edition of the study, with 2,131 executives surveyed from July 23 to August 26, 2026, shows that cost control is advancing quickly: 61% review AI costs when approving a project and 59% monitor them once the system is running (KPMG, 2026b). Because these are not the same companies as in the previous quarter, the figures are not directly comparable.

The missing step is a different one: only 12% consistently weigh the value of AI against its cost across the organization, and another 26% do so for most of their initiatives. Among companies that already report established ROI, the figure rises to 48%, four times the average (KPMG, 2026b). The gap between the most mature companies and those just starting to experiment also shows in their controls:

ControlCompanies still experimentingCompanies with established ROISource
Cost-monitoring dashboards43%77%KPMG, 2026b
Cost review when approving projects50%73%KPMG, 2026b
Usage or token budgets31%46%KPMG, 2026b

Source: KPMG, 2026b.

In KPMG’s words, most companies can now “see the meter running.” It is the same gap that Genpact’s study of agents in production shows, where only 15% of companies have fully tied their AI investment to the income statement (Wasivia, 2026). What sets apart the companies that get a return is that they also know what that spending bought them (KPMG, 2026b).

What it means for companies in Ecuador

The survey does not include Ecuador, but the problem does not depend on the country; it depends on how AI is billed. For a mid-sized company, the risk is not a million-dollar bill but an agent that costs more than it saves without anyone noticing for months. And because billing is based on consumption, a pilot that works well can become more expensive just as it is rolled out to the whole team.

How to control the cost of an AI agent

  1. Calculate the cost per task, not per month. How much does it cost for the agent to answer a query or process an invoice? Compare it with what it costs to do it today without the agent.
  2. Set a cap from day one. Define a monthly budget, a limit on steps per task and alerts for when spending approaches the cap (Wasivia, 2026).
  3. Use the right model for each step. Reserve the most powerful model for what really needs it, and use cheaper models to classify, extract or summarize.
  4. Reduce the context. Send the model only the information required for each step, not the entire history.
  5. Review cost and value every month. If the agent does not pay for itself, adjust it, narrow it or stop it, as nearly half of the companies surveyed have already done. And confirm that value with your finance team (Wasivia, 2026).

At Wasivia, we design AI agents and multi-agent systems with cost per task, spending limits and value measurement built in from the start.

References

Similar Posts