Companies are accelerating the adoption of artificial intelligence in IT operations. Yet as systems evolve from analytical tools into autonomous actors, data quality, traceability and tightly controlled access rights become increasingly critical. A NETSCOUT survey shows that technological momentum is strong, but operational trust has yet to catch up.
Artificial intelligence is evolving from a support tool into an active component of IT operations. AI systems correlate alerts, analyse incidents, prioritise potential root causes and assist with remediation. The emergence of AIOps and so-called AgenticOps is bringing organisations closer to a scenario in which AI systems do not merely recommend actions, but initiate them autonomously within predefined boundaries.
This development changes the central question. What matters is no longer simply whether AI is being used in IT operations, but which information informs its decisions, how its conclusions can be verified and what interventions it is permitted to perform in production environments.
This is precisely where NETSCOUT’s study, “Observability, AI Readiness, and Operational Resilience”, identifies a noticeable gap between technological adoption and operational maturity. Conducted at Cisco Live 2026, the survey covered almost 950 IT professionals and technical decision-makers. Its findings portray an industry that is widely testing and deploying AI while continuing to struggle with fragmented data, limited visibility and concerns about trust.
AI has entered mainstream IT operations
According to the survey, 51.5 per cent of respondents are already working on AI or large language model projects in an IT context, testing relevant applications or planning to introduce them within the next twelve months. Of these, 17.2 per cent already use such solutions in production, 16.0 per cent are in the testing or development phase and 18.3 per cent intend to begin implementation in the near term. A further 31.3 per cent are discussing corresponding initiatives.
These figures demonstrate considerable technological momentum. However, they reveal little about the extent to which organisations are prepared to transfer operational responsibility to AI systems.
There is a substantial difference between an AI application that classifies incidents and a system that independently modifies configurations or restarts services. With each additional level of autonomy, the requirements for data quality, contextual information, access controls and oversight mechanisms increase.
The essential prerequisite for autonomous IT processes is therefore not merely a powerful model. It is a reliable and comprehensive picture of the operational environment.
The question of trust is primarily a question of data
The survey findings show that the obstacles to AIOps and AgenticOps are closely interconnected. Some 27.2 per cent of respondents identify a combination of factors as their greatest challenge. A lack of real-time visibility is cited by 22.3 per cent, while 22.1 per cent point to poor data quality or incomplete context. Another 21.8 per cent report insufficient trust in AI-generated outputs, and 21.6 per cent identify siloed operational tools as a barrier. A further 19.1 per cent consider excessive volumes of low-value telemetry data problematic.
A lack of trust in AI therefore cannot be viewed simply as an issue of user acceptance. If data sources are incomplete, dependencies between cloud services, applications and networks remain undocumented, or monitoring systems provide only isolated views of the infrastructure, any AI-based analysis will inevitably be incomplete as well.
This creates a fundamental requirement for organisations: AI decisions must be traceable to verifiable evidence.
Analysts need to understand which signals informed an assessment, which dependencies were considered and how confident the system is in its conclusion. Particularly where automated interventions are concerned, a plausible answer is not sufficient. What is required is a transparent and auditable chain of reasoning and action.
Long resolution times underline the pressure to act
Another survey finding illustrates the need for better analytical and decision-making processes. Some 56.2 per cent of respondents require several hours to resolve performance problems completely. For 20.8 per cent, resolution takes an entire day or longer. Only 11.0 per cent report that such problems are resolved within minutes.
Mean time to resolution, or MTTR, is therefore not merely a technical metric. It has become a key factor in operational resilience.
The effects of prolonged disruption extend directly into business operations. Employee productivity losses are identified as the most significant consequence by 34.8 per cent of respondents. Another 24.2 per cent report an impact on customer service, while 17.6 per cent see a risk to revenue. Additional consequences may include higher costs, reputational damage and regulatory repercussions.
AI can deliver considerable benefits in this area. It can correlate signals, prioritise possible root causes and identify known remediation paths more rapidly. Nevertheless, speed cannot compensate for an unreliable information base. An automated system working with incomplete data may not only produce an incorrect diagnosis more quickly, but also amplify the consequences by executing inappropriate actions automatically.
Observability is becoming a prerequisite for AI autonomy
Almost two-thirds of respondents consequently attach considerable importance to the continuous monitoring of AI processes and services. Some 38.0 per cent regard it as important, while a further 27.0 per cent consider it very important.
This development is also changing the role of observability. Its purpose is no longer limited to monitoring conventional IT systems. Increasingly, the AI itself must become an object of observation.
For production-grade AI services, traditional indicators such as availability and response time are insufficient. Other relevant factors include infrastructure consumption, data access, technical dependencies, output quality and the consequences of actions initiated by the system. Agentic systems raise additional questions: What decision did the system make? On which data was it based? What change was executed? Can it be reversed? At what point is human intervention mandatory?
Organisations therefore require an end-to-end audit trail extending from the initial signal through analysis and evaluation to the action ultimately performed.
For autonomous systems, observability consequently becomes a security control in its own right.
Access rights define the scope of action
To create genuine value in IT operations, AI systems require access to operational tools and data. According to the survey, 61.9 per cent of respondents consider secure access to such resources for AI assistants and automated workflows important or very important.
That access, however, must be subject to clearly defined limits.
An AI system that only analyses log data presents a different risk profile from an agent permitted to modify configurations, restart services or alter security rules. Organisations must therefore distinguish between read-only, advisory and write permissions.
From a technical perspective, this requires measures including role-based access controls, the principle of least privilege, comprehensive logging of all actions, defined approval levels, and fallback and recovery mechanisms.
A graduated model of autonomy offers a practical approach. Initially, the AI observes the environment. It then recommends measures, proceeds to execute standardised actions following approval and, only once sufficient evidence and clearly defined boundaries are in place, receives limited autonomous authority.
Autonomy should therefore not be treated as a binary decision, but as a controlled scope of action that can be expanded incrementally.
Context matters more than data volume
The quality of AI-supported decisions ultimately depends on which data is available and how effectively different sources are connected.
Some 39.5 per cent of respondents primarily use metrics, events, logs and traces—often grouped under the acronym MELT—as their observability data. Packet captures are the main source for 21.6 per cent, while 18.8 per cent primarily rely on flow data and 9.7 per cent use deep packet inspection, or DPI.
The strategic assessment of DPI is particularly notable: 41.2 per cent consider the technology important and a further 23.1 per cent regard it as very important. In total, 64.3 per cent therefore attribute a high level of importance to DPI.
This finding highlights a central challenge for modern observability architectures. Individual telemetry sources provide only a partial view of the actual state of a system. Network data can supply additional evidence, reveal communication relationships and help trace faults across multiple technical domains.
The objective should therefore not be to collect the greatest possible volume of data. What matters is its quality, correlation and contextualisation.
Where deep network inspection is used, organisations must also consider data protection requirements, purpose limitation and access control. Greater visibility only improves security if access to the underlying information is itself properly governed.
Six priorities for reliable AI operations
The survey findings point to six central areas of action:
- Establish visibility before granting autonomy: Dependencies between applications, networks, cloud services and infrastructure must be visible before AI is permitted to perform operational interventions.
- Correlate multiple data sources: MELT, flow, packet and other operational data should be combined to create a consistent and comprehensive picture of the environment.
- Monitor the AI itself: In addition to availability and performance, organisations must maintain visibility into data access, output quality, decisions and executed actions.
- Apply strict permission controls: Role-based models, the principle of least privilege, audit trails and fallback mechanisms should form part of the basic architecture of agentic systems.
- Expand autonomy incrementally: From analysis and recommendations to approved actions and limited autonomous execution, each stage should be governed by measurable quality thresholds and clearly defined termination criteria.
- Measure value in terms of resilience: Success should not be determined by the number of AI tools deployed, but by shorter incident resolution times, reduced disruption and more stable business processes.
Survey primarily reflects complex enterprise environments
The findings should be interpreted in the context of the survey sample. Depending on the question, between 937 and 950 responses were included in the analysis. Participants were surveyed at Cisco Live 2026 and included enterprise architects, network operations specialists and IT executives in particular.
Some 55.7 per cent of respondents work for companies with more than 5,000 employees, while 37.6 per cent are employed by organisations with more than 10,000 staff.
The study therefore primarily reflects the perspective of large organisations with technically complex IT environments. It is not a representative survey of every industry or company size. It should also be noted that NETSCOUT is itself a provider of network, observability and security solutions. The results offer a relevant snapshot of opinion among a technically specialised audience, but should not be extrapolated directly to the entire market.
Conclusion: Trust must be built into the architecture
The next stage of AI development in IT operations will not be determined by more powerful models alone. As systems assume greater operational responsibility, traceable decisions, reliable data and clearly defined scopes of action become increasingly important.
Observability is consequently acquiring a new function. It must do more than explain why an application is slow or a service has failed. In future, it will also need to demonstrate why an AI system reached a particular conclusion and why it initiated a specific action.
For organisations, this means that trust in AI cannot be granted in advance. It must be created through architecture—through verifiable evidence, transparency, restricted privileges and auditable decisions.
Only then can AI-supported automation evolve into reliable operational autonomy.
The Study at a Glance
Title: Observability, AI Readiness, and Operational Resilience
Publisher: NETSCOUT Systems, Inc.
Survey setting: Cisco Live 2026
Sample: Between 937 and 950 IT professionals and technical decision-makers, depending on the question
Focus areas: Observability, operational resilience, AI and LLM maturity, AIOps and AgenticOps
Notable characteristic: More than half of respondents work for organisations with over 5,000 employeesSource: NETSCOUT, “Observability, AI Readiness, and Operational Resilience”, Cisco Live 2026 Survey, 2026.
