patrick.belmans@servinomy.be
en

Agentic AI 

Silver bullet or maturing technology?

Servinomy | July 2026

Introduction: Between promise and reality

Agentic AI is one of the most discussed technologies right now. In boardrooms, at technology conferences and across social media, the message is often the same: AI agents will make decisions independently, execute complex processes and automate substantial parts of knowledge work.

That expectation is not wholly unrealistic. AI systems can already summarise information, surface knowledge, classify tickets, trigger workflows, call software tools and carry out a sequence of steps in tightly defined circumstances. Yet a substantial difference exists between a system that supports tasks and one that can reliably, safely, and independently achieve complex business goals.

For IT Service Management, that distinction matters greatly. Incidents, changes, access rights and customer communications affect continuity, security, compliance and trust. The relevant question, then, is not whether agentic AI has value. It does. The question is when autonomy adds value and when human oversight remains essential.

What is agentic AI?

The term agentic AI is not used consistently across the market. In practical terms, it usually refers to an AI system that operates within a defined environment and can:

  • interpret an objective;
  • develop a plan or sequence of actions;
  • use tools or software systems;
  • evaluate the results;
  • select follow-up actions;
  • and escalate to a human when needed.

This is different from a conventional chatbot that only generates text. It is also different from traditional RPA, where fixed rules and predefined process steps are central.

In practice, however, the boundaries are blurred. Many suppliers now label chatbots, copilots, RPA workflows and basic automation as agents. Gartner calls this agent washing: relabeling existing products as agentic AI without substantial autonomous capability. Gartner estimated in 2025 that only about 130 of the thousands of vendors offering agentic AI met its definition of a genuinely agentic product. That is a Gartner estimate, not a universal count of the entire market.

Source

The hard reality behind the interest

In June 2025, Gartner predicted that more than 40% of agentic-AI projects would be cancelled by the end of 2027. It cited escalating costs, unclear business value and inadequate risk controls. This statement must be read accurately: it is a forecast, not a finding that more than 40% of projects have already failed.

In a Gartner webinar poll of 3,412 participants in January 2025, 19% said their organisations were making significant investments in agentic AI. A further 42% were investing conservatively, 31% were taking a wait-and-see approach or were unsure, and 8% had not invested. These figures show strong interest, but they are not a representative global survey of all enterprises.

A separate Gartner survey of 360 IT application leaders from organisations with at least 250 employees showed a similar pattern. Only 15% were considering, piloting or deploying fully autonomous AI agents. At the same time, 74% regarded AI agents as a new attack vector, and only 19% had high or complete trust in their vendors' ability to provide adequate hallucination protection.

This does not mean that organisations are broadly rejecting AI agents. Gartner reported that 75% were piloting, deploying or had already deployed some form of AI agent. The gap lies between limited or supervised automation on the one hand, and fully autonomous systems without human oversight on the other.

Source-1 | Source-2

Autonomy is not a binary choice.

The debate around agentic AI is often framed incorrectly as a choice between full human execution and full autonomous replacement. In reality, autonomy is a spectrum. A reliable ITSM model may operate across several levels:

Level | ITSM example | Human role

  • Assistance | Summarising incident history | Human decides and acts
  • Recommendation | Suggesting a classification or root cause | Human validates
  • Bounded automation | Password reset or standard access request | Human sets rules and exceptions
  • Approval-led execution | Preparing a change for approval | Human authorises the action
  • Autonomous execution within guardrails | Restarting a non-critical service | Human defines boundaries, monitoring and escalation
  • High-risk autonomy | Security changes, production recovery or customer compensation | Appropriate only in exceptional, demonstrably controlled circumstances

Most immediate value is usually not found at the far end of this spectrum. Organisations can gain significantly from classification, routing, knowledge retrieval, summaries, draft responses and controlled automation without giving a model unrestricted decision-making power.

The Klarna example: a relevant lesson, not universal proof

Klarna is often cited as evidence that AI cannot replace human customer service. The case deserves attention, but it needs careful reading.

Klarna promoted its AI customer-service capability very ambitiously and said it could perform work equivalent to the capacity of around 700 customer-service staff. Later, CEO Sebastian Siemiatkowski acknowledged that cost had become too dominant a factor in assessing the solution and that this could reduce quality.

The lesson is not that AI customer service necessarily fails. The lesson is that cost is an insufficient success metric. A serious evaluation should also consider:

  • resolution quality;
  • escalation rates;
  • customer satisfaction;
  • rework;
  • the risk of inaccurate information;
  • brand impact;
  • accessibility for complex or vulnerable customer cases.

The same caution applies when broader claims are made about Duolingo, Starbucks, McDonald's, or a general trend of "AI lay-off reversals". Individual company cases can be valuable warnings, but they do not by themselves prove a general rule.

Hallucinations: a risk that requires context

Hallucinations are a genuine problem: language models can produce inaccurate, fabricated or insufficiently supported information. But general error percentages are misleading when presented without context.

The performance of an AI system varies substantially according to:

  • the model and model version;
  • the type of task;
  • the domain;
  • the quality of available data;
  • the use of retrieval and trusted sources;
  • tool use;
  • prompt and workflow design;
  • the evaluation method;
  • human validation.

A hallucination rate from a particular benchmark cannot therefore be translated directly into the probability that an ITSM agent will misdiagnose an incident. That requires separate testing in the organisation's own environment, with its own data, processes and realistic edge cases.

The correct conclusion is not that autonomous AI is impossible. The correct conclusion is that organisations should permit autonomous action only where they understand the failure modes, limit the consequences and maintain reliable fallback mechanisms.

Where AI in ITSM is already delivering value

AI is already delivering value in ITSM, particularly when it strengthens people and processes.

A 2026 TeamDynamix market study of 392 IT professionals reported that organisations with broader AI-ITSM adoption experienced ticket deflection, faster ticket resolution and improved customer satisfaction. The study identified knowledge-article suggestions and generation, virtual agents, intelligent ticket routing and AI-assisted responses as important applications.

These findings are relevant, but must be interpreted appropriately. TeamDynamix is itself an ITSM and automation supplier. Its figures are useful indicators of reported respondent experience, not independent proof that every organisation will achieve the same results.

The most realistic ITSM applications today include:

  • finding, updating and drafting knowledge articles;
  • summarising and classifying incidents;
  • routing tickets;
  • handling recurring enquiries;
  • generating draft responses;
  • combining signals for faster diagnosis;
  • carrying out standard procedures within pre-approved limits.

These are not marginal applications. They can improve response times, reduce avoidable manual work and raise the quality of service processes.

Source

The complexity paradox

AI can improve ITSM operations while making their governance more complex. SolarWinds reported in its 2026 IT Trends Report that many IT professionals experienced additional responsibility around AI-supported incident response, validation and governance. Associated reporting noted, among other findings, that 44% experienced a new or increased responsibility for cross-team incident response, while 71% continued to check AI outputs manually.

This is not an argument against AI. It shows that AI does not merely add automation; it also creates new work:

  • quality assurance;
  • governance;
  • logging and observability;
  • access control;
  • assessment of model and workflow performance;
  • incident response for AI behaviour;
  • exception handling;
  • staff training.

An organisation that looks only at time saved in one workflow may miss this additional management burden. An organisation that treats AI as a socio-technical system can deliberately design, measure, and reduce those burdens.

Source-1 | Source-2

From demonstration to production

Many agentic-AI initiatives do not fail because the model cannot produce an impressive demonstration. They fail because production environments impose different requirements.

A robust implementation requires at least:

  • a clearly defined business objective;
  • explicit decision boundaries;
  • reliable, current and accessible data;
  • permissions based on least privilege;
  • logging of prompts, tool calls, actions and outcomes;
  • monitoring for errors, anomalies, cost and security risks;
  • test scenarios involving exceptions and poor data;
  • human escalation for uncertainty or high impact;
  • rollback and recovery procedures;
  • measures of real business value.

An agent allowed only to classify and route tickets has a fundamentally different risk profile than an agent that modifies production configurations, manages credentials, or provides external customers with binding information. Governance must therefore align with the action, not just the model.

The real lesson for ITSM

Agentic AI is not a miracle cure, but it is not an empty promise either. It is a collection of rapidly developing capabilities that can already deliver value in specific, well-designed situations.

The most successful organisations will likely not be the first to roll out the most autonomous agents. They will be organisations that:

  • choose a narrow use case with measurable value;
  • organise reliable data and knowledge;
  • increase autonomy incrementally;
  • design risks in advance rather than remediating them afterwards;
  • retain human expertise where context, judgment, and responsibility are needed;
  • measure performance based on quality, safety, customer value, and total workload, not just cost savings.

The right question, then, is not: Can agents replace people?

The better question is:

Which tasks can we automate safely, measurably, and responsibly, so that people can make better decisions and perform more complex work better?

Agentic AI only becomes valuable when technology, process design, data, governance, and human responsibility mature together.