The problem with enterprise AI agents is no longer just what they cost. It’s that nobody can predict the invoice.
That is the central finding of the Splunk State of Observability 2026, released today by Cisco. The sixth edition of the report surveyed 2,050 ITOps and engineering respondents across 10 countries and describes a widening gap: the urgency to deploy AI agents is rapidly outpacing the ability to monitor them. For CIOs who have spent the past two years green-lighting agent pilots, the data reads like an audit finding.
Priority without proof
Monitoring AI has become a boardroom-level expectation. According to the report, 87% of respondents say it is a priority for their organisation. Yet 54% have no clear business outcomes tied to that priority, which means a majority are funding oversight they cannot define success for.
The measurement gap is sharper still. Some 56% say they cannot measure AI workload metrics such as LLM response quality, a core indicator of whether an agent is doing its job. And while 85% call visibility into hallucinations important, only 58% say they are effective at it.
The anxiety behind these numbers is widespread: 83% of respondents are concerned about the unintended consequences of AI agents. As agents move into software development and customer experience workflows, the report warns, one incorrect action can escalate into a critical business failure. Monitoring them, it argues, requires a new class of behavioral metrics tied to business outcomes and agent decision-making, not just uptime and latency.
The surprise is the problem
If quality is the first crisis, cost is the second, and the report frames it carefully. The issue is not only the size of the bill but its volatility.
• 55% say they cannot control AI costs, as consumption grows and usage stays hard to predict.
• 60% say unexpected billing spikes cause more operational strain than high baseline costs.
• 57% say their telemetry costs have eclipsed the cost of the infrastructure being monitored.
That last figure deserves a CIO’s attention. When organisations run dozens of agents across multiple models, the data generated to watch them can become a budget line of its own. AI is exacerbating an existing data volume problem, making it nearly impossible to sift through the mountain of signals.
Kamal Hathi, SVP and GM of Splunk at Cisco, put it bluntly: “the excitement of AI is clashing with the reality of the bill.” His prescription is to fold observability into an AI tokenomics framework so that spend becomes measurable and tied to business value rather than an unpredictable expense.
There is a practical architecture lesson here as well. The report notes that 67% of respondents use a single observability platform, which gives a unified view for pinpointing exactly where telemetry costs originate. Tool sprawl, in other words, is a cost-attribution problem in disguise.
Security and observability are merging
Perhaps the most consequential organisational shift is happening between teams that historically kept their distance. As agents gain access across the entire tech stack, 81% of respondents say monitoring AI workloads has increased collaboration between observability and security teams.
The operational evidence is tangible:
• 73% can accurately trace performance issues to security root causes, up from 65% last year.
• 77% of observability teams now troubleshoot with security teams, up from 62% in 2025.
• Data sharing between the two functions has climbed from 74% to 84%.
Jeetu Patel, Cisco’s President and Chief Product Officer, argued that questions like “have our agents been compromised?” and “are they doing their job well?” are now inseparable, and that they move at machine speed. His conclusion: common data platforms that detect and respond in real time are the key to deploying agents that can be trusted at scale. Note the vendor perspective, but the survey numbers point the same direction.
What the leaders do differently
The report identifies an advanced cohort of organisations achieving concrete business outcomes. Their results are the strongest argument in the study:
• Cost attribution: 91% of leaders can accurately attribute AI costs to specific agents, models and workloads, versus 69% of non-leaders.
• ROI: Leaders are 2.4 times more likely to say their observability platform’s return significantly exceeds expectations.
• Revenue impact: 54% say observability has yielded at least $5 million in annual business revenue, compared with 28% of non-leaders.
• Security alignment: 94% say monitoring AI has increased security collaboration (versus 80%), and 86% can trace performance issues to security root causes (versus 71%).
Three habits set them apart. They rely on AI agents to fix and prevent issues frequently, using them to execute key remediation tasks. They closely monitor the factors that keep AI in check: model quality, cost, usage, security and privacy. And they connect signals to business impact, treating telemetry as an input to planning and strategy rather than an operational afterthought.
A global view
The themes held across all 10 countries, with notable regional differences:
• India: 51% describe their practice as very effective at monitoring AI model quality, versus 35% globally.
• United Kingdom: 32% use AI agents all the time for root cause diagnosis (20% globally), and 55% strongly worry about unintended consequences.
• Saudi Arabia: 98% say AI lets them focus more on innovation, well above the 84% global average.
• Germany: 74% rate visibility into AI privacy as very important, leading on data sensitivity.
What CIOs should do now
The report’s findings translate into a short agenda for technology leaders:
1. Tie agent monitoring to named business outcomes. If 54% of peers cannot do this, doing it is an immediate differentiator.
2. Instrument for quality, not just availability. Response quality and hallucination detection are the metrics most organisations admit they lack.
3. Make AI spend attributable. Know which agents, models and workloads drive cost before the invoice arrives, and treat spikes as the primary risk.
4. Audit your telemetry bill. If monitoring costs rival the infrastructure under watch, consolidation and data-volume discipline belong on the roadmap.
5. Share data with security by default. The leaders’ results suggest convergence is paying off in faster, more accurate root-cause tracing.
The bottom line
Agentic AI promised productivity. The Splunk research suggests it will also deliver something less welcome to finance chiefs: variance. The organisations pulling ahead are not spending less on AI so much as understanding what they spend, why, and what it returns. In the agent era, observability is no longer an IT hygiene function. It is how the CIO earns the right to scale.