Most IT leaders in large organizations claim to be ready to adopt AI in their operations. The IT service management tool is up to date, the data exists, and processes are at least partially documented. But the pilot project stalls, and root cause analysis always reveals the same weak points.
An organization’s level of AI operations readiness depends on concrete, verifiable conditions relating to the data, processes, and rules already in use. In these areas, AI does not replace the foundations of ITSM — if anything, it puts them to the test like never before. An AI system reproduces and amplifies the quality of the information and processes to which it is applied: where data and procedures are well-organized, the value produced increases; where they are inconsistent, it multiplies errors and unnecessary alerts.
This article lists ten recurring problems that specifically concern AI operations readiness, organized into four categories. Many of the issues described also emerge from the EasyVista report The Reality of ITSM in 2026, conducted with 1,100 IT professionals.
Data
The problems in this category weigh more heavily on AI operations readiness than others, because every additional automation built over time inherits the flaws of the data it works with.
1. Fragmented data sources
AI feeds on historical tickets, knowledge base articles, logs, asset information, CMDB data, and monitoring readings. When these sources reside in separate systems, the AI system works each time on limited portions of the context. The EasyVista report The Reality of ITSM in 2026 finds that 58% of organizations use dedicated ITSM platforms, while 25% use basic ticketing systems and 7% still rely on manual processes.
Solution: to identify a potential problem, list the systems that contain operational data and verify which of them exchange information automatically. Build a catalog of sources, indicating origin, owner, and actual usage.
2. Non-uniform ticket quality and management
If every team fills in tickets in its own way, the AI learns that inconsistency and reproduces it. Poorly designed mandatory fields, categories assigned out of habit, descriptions so brief as to be obscure and therefore unusable: starting from conditions like these, the recommendations produced may appear sensible but could in fact be flawed, incorrect, useless, or even harmful.
Solution: it is better to opt for a quick check. Extract a sample of one hundred recent tickets and verify how many are classified consistently. The fix requires shared naming conventions, deduplication, and continuous maintenance of the knowledge base.
3. Absence of a single authoritative reference system
When the same asset appears under different names in three systems, no automation can determine which version is correct. The most obvious symptom appears when two teams, asked about the same service, provide conflicting counts because each draws from a different system.
Solution: your intervention here will consist of identifying, for each information domain, which system holds the authoritative data, and ensuring all others derive from it.
Operations
The Operations category concerns the way activities are handed off from one person to another and from one system to another. An AI applied to poorly defined processes inherits the uncertainty that characterizes them, rather than correcting it.
4. Undocumented workflows
The accuracy of recommendations produced by an AI system depends on the degree of formalization of the processes on which the AI operates. When incidents follow a defined flow, requests are standardized, and escalation criteria are shared, the system has stable references on which to base its proposals. In the EasyVista report cited above, 44% of organizations describe their approach as well-structured and equipped with defined processes, while over 40% still operate in only partially structured or ad hoc environments.
Solution: carry out a simple test. Ask two team leaders who are involved in the same process to describe its steps, then compare the answers. The solution to the problem of undocumented workflows always comes through mapping the flows actually in use, exceptions included.
5. Unclear responsibilities
When a model suppresses an alert or lowers the priority of an incident, someone must be accountable for that decision. In the absence of defined ownership, every error generates a discussion instead of a correction.
Solution: the most direct way to verify and correctly assign the various responsibilities is to ask the people involved who actually has the authority to expand the level of automation: conflicting answers indicate that ownership has not been assigned. A clear definition of data owners, quality contacts, and a formalized escalation chain is required.
6. Manual handoffs between teams or systems
Every time information is manually copied from one tool to another, delays and opportunities for error are introduced. The prevalence of these manual handoffs remains high: the EasyVista survey cited multiple times above finds that interconnected workflows between ITSM and SecOps tools have been implemented by 55% of organizations, and that 62% use shared dashboards. The remainder of the sample still relies on human operators to transfer data between systems that do not communicate.
Solution: quantify the manual handoffs present in your organization’s IT environment by reconstructing the complete journey of a high-priority alert, from detection to closure, noting every point at which a simple transfer of information takes place. These handoffs are the first candidates for automation.
Governance
The third category concerns the rules that establish how far the system can act autonomously. The problems in this category are the most insidious, because they remain invisible as long as everything is working and only manifest themselves during the highest-impact incidents.
7. Absence of defined confidence thresholds
A recommendation formulated with minimal uncertainty and one formulated on weak grounds require different treatment. Without explicit thresholds, all proposals from the system are treated the same way, with two possible outcomes: an excess of manual reviews, or automations applied to cases that do not offer sufficient guarantees.
Solution: determine what level of confidence a system must reach before one of its proposals is executed without review. Each use case requires its own threshold, expressed as a precise value and agreed upon before activation.
8. Lack of human-in-the-loop controls
A good level of AI operations readiness includes human review points placed where the risk is highest. Their absence is felt above all in the fact that no one can say how many automated decisions have been reviewed within a defined period of time — for example, in the last quarter.
Solution: establish which categories of decision require confirmation, who must provide it, and with what level of traceability. Concerns related to hallucinations and data accuracy are marginal among IT leaders in large companies, cited as the main obstacle by only 4%: the perceived risk is low, which makes explicit controls all the more necessary.
Automation
The final category deals with the infrastructure that supports automations. Even with well-organized data and defined processes, an environment in which tools do not communicate with each other limits AI to isolated interventions.
9. Technology silos that prevent integration
The presence of too many separate tools generates operational costs and slowdowns. On the security front, EasyVista finds that 45% of leaders report excessive use of separate tools, 63% report significant operational overload in managing multiple solutions, and 47% report that this fragmentation slows incident response. The most recognizable symptom of the problem is the presence of multiple tools flagging the same event, forcing teams to determine each time which alert is the one to handle.
Solution: the direction to work toward is consolidation. 75% of organizations report that they are already actively reducing the number of solutions in use, consolidating tools to reduce complexity.
10. Absence of an orchestration layer
Many organizations introduce automations one at a time, each confined to the application in which it was configured, and the benefits remain limited to that perimeter. An orchestration layer is the software component that coordinates these separate automations: it establishes the order of execution, manages dependencies between one activity and another, and makes it possible to suspend an entire sequence when a step produces anomalous results.
Solution: ask yourself which automations are active at any given moment and who is responsible for them. You can make improvements by centralizing the definition and monitoring of automations, making them observable and reversible.
How to use this list
An assessment of AI operations readiness produces useful results when conducted with the necessary critical distance. The ten problems described should be examined one by one, assigning each a status: absent, partial, resolved. The result indicates where to intervene before launching any pilot project.
The order in which the 10 problems are resolved also matters. Data gaps come before automation gaps, because an orchestration layer built on inconsistent sources amplifies the disorder. Governance gaps must be addressed before increasing the system’s autonomy, for reasons that become clear in the transition from pilot to production, where volumes, variety of cases, and integration complexity all increase simultaneously.
It is also worth resisting the temptation to purchase additional features to eliminate these critical issues. The risk, well documented, is paying for modules that the organization will not be able to use because it lacks the necessary level of maturity, with the result of adding complexity to the very problem it intended to solve.
The most prepared organizations are those that have cultivated continuous improvement over time as a discipline, not those with the greatest number of tools or ongoing experiments.
FAQs
- Which of the 10 problems is the best starting point?
The data ones. Fragmented sources, inconsistently completed tickets, and the absence of a single authoritative reference system affect all subsequent interventions, because every automation inherits the flaws of the information it works with. - How long does it take to address these 10 gaps?
It depends on the starting point, but we are talking months, not weeks. Naming conventions and flow mapping can be tackled in parallel, while defining confidence thresholds requires sufficient historical data to recalibrate them. - Is a new platform needed to achieve a good level of readiness?
In most cases, no. The ten problems concern data, processes, responsibilities, and coordination — areas that depend on organizational choices more than on tools. Purchasing additional modules before addressing them adds costs and complexity. - What does human-in-the-loop mean in practice?
It means establishing which automated decisions require human confirmation before taking effect, who must provide it, and how it is recorded. The control point should be placed where the risk is highest — for example, on production changes and priority adjustments for critical incidents.
A guide to AI in ITSM
Discover how to integrate artificial intelligence into your ITSM, redesign your processes, and take your company’s efficiency to the next level.
A guide to AI in ITSM
Discover how to integrate artificial intelligence into your ITSM, redesign your processes, and take your company’s efficiency to the next level.