Top 5 Risks of Using AI Agents in ITSM and How to Mitigate Them

AI agent risks shown as a city split between a fiery red sky and a pale domed building, with a figure in a three-cornered hat and horse-drawn carriages between them

Summary

AI agent risks in ITSM fall into five areas: whom the agents are built for, what data and systems they can reach, which model they depend on, how much of your environment they understand, and whether the AI comes from the ITSM platform or a third party. Richard Mendis recommends designing agents around specific personas, treating them as fully privileged participants whose every action is logged and auditable, and requiring architectures that let a model be swapped without rewriting workflows. He adds that AI agents need deliberate onboarding in your institutional knowledge, much as a new hire does, and that a hybrid of native and third-party AI is the route that scales with the least risk and spend. He backs this with SailPoint’s finding that 96% of tech leaders see AI agents as a growing security threat while fewer than half have policies to manage them, and MIT’s finding that 95% of generative AI pilots fail to deliver measurable impact.

“It was the best of times, it was the worst of times.” A Tale of Two Cities may be centuries old, but the opening line from Charles Dickens’s novel reflects the current reality of artificial intelligence (AI) use in IT service management (ITSM). Adopting AI promises unprecedented transformations like self-healing ITSM, but it can also expose organizations to new vulnerabilities and AI agent risks. This undermines the very gains AI agents were meant to deliver and increases the load on IT teams. 

This article considers these most immediate risks of deploying AI agents in your ITSM strategy and how to mitigate them.

1. Misaligned use cases: not knowing your target users

The fastest way to derail an AI initiative is to deploy AI agents without defining whom they’re meant to help. McKinsey’s State of AI in 2025 Report found that more than two-thirds of respondents say their organizations use AI in more than one function, but only half report using AI in three or more functions, leaving a significant gap in terms of AI agent specialization. 

Optimizing your ITSM platform requires thinking about all of its personas, who have very different needs:

  • End users (often employees) want quick, accurate answers to common questions. They don’t care about your architecture; they care about resetting their password in 30 seconds.
  • ITSM administrators want relief from repetitive maintenance tasks, such as user‑role cleanup, configuration management database (CMDB) hygiene, and publishing knowledge base articles. 
  • ITSM tool “developers” want help building catalog items, forms, flows, and portal components faster and with fewer errors.

If you don’t design AI agents around these personas, you end up with technology nobody adopts. Overgeneralizing an AI agent risks making it “a master of none,” where it doesn’t provide any true benefit to a given audience. 

Although generalist AI agents appear more cost-effective because they can take on many types of tasks across domains, there may be greater long-term value in developing AI agents that are specifically equipped to handle pain points like repetitive “ITSM builder tasks” because it will help solve larger business challenges like turnover, overworked IT teams, and reactive ITSM platform use. 

2. AI Agent Risks: Data exposure and security gaps

AI agents are effective because they can actively participate in your ITSM environment, but this also creates vulnerabilities if not governed correctly. Although 96% of tech leaders agree that AI agents are a growing security threat, fewer than half have policies to manage them, according to a SailPoint study (sign-up is needed). This played out in July when OpenAI lost control of an AI model and cyberattacked another company, Hugging Face. 

A common misconception is that AI agents inherit the security posture of the ITSM platform they connect to. They don’t. They introduce new data and security risks that must be governed independently.

Without properly designed controls, an AI agent can unintentionally:

  • Access data outside its intended scope
  • Bypass approval workflows
  • Execute actions without human approval 
  • Expose information through prompts or logs.

Many organizations deploy AI agents without solid foundational data privacy policies and practices, which compounds risk through:

  • Data exposure
  • Regulatory consequences
  • Revenue loss
  • Reputational damage. 

The irony is that AI initiatives typically aim to improve operations and reduce long-term costs. However, without proper security measures, it is more likely to incur remediation costs that could exceed millions of dollars. Perhaps the greatest risk is that leaks and attacks can fly under the radar, as in the July OpenAI and Hugging Face incident. 

CIOs should treat AI agents as fully privileged participants in the ITSM ecosystem, not lightweight add-ons. That means ensuring AI agents:

  • Authenticate securely into the ITSM platform
  • Operate strictly within established role-based access controls
  • Execute only within approved infrastructure environments. 

Just as importantly, every action an AI agent takes must be logged, traceable, and auditable. Without this level of visibility, the environment isn’t enterprise-ready.

The challenge is that this is neither quick nor straightforward. Even the best protections could be outsmarted by recursive AI. For most organizations, it quickly becomes an effort that surpasses the bandwidth of their internal IT, security, risk, and legal teams. Bringing in AI consultants to guide these integrations can reduce risk by ensuring the right foundation is in place to scale AI safely and credibly across the organization. 

3. Model lock-in and lack of LLM flexibility

Choosing a single large language model (LLM) or vendor locks you into today’s capabilities and costs. The AI landscape is evolving faster than any technology cycle in the last 20 years. New models outperform previous generations, costs fluctuate, and capabilities evolve every few months. Even if one service, like Claude or ChatGPT or whatever native model your ITSM vendor uses, clearly leads the other in a given capability, that can quickly change. So it makes sense to factor flexibility into your strategy. 

Locking your ITSM strategy to a single model or vendor is a strategic risk with long‑term consequences – including betting your business success on a nascent technology that is very far from its final form. Nobody can predict which AI model is the best investment because the “best” model today will not be the best in 12 months. Not to mention, some competitors will inevitably become obsolete within the next few years as new models surpass the abilities of LLMs altogether. Plus, different tasks require different AI model strengths, so you may want to experiment with a range of AI agents depending on the functions and user. 

Organizations that lack LLM flexibility face three predictable risks:

  1. Escalating costs as usage‑based pricing increases
  2. Lower accuracy as competitors adopt better‑performing models
  3. Reduced agility when compliance or business needs change.

CIOs must demand architectures that allow for AI model substitution without rewriting workflows. The ability to choose the right AI model for the right task and to change that choice over time will significantly reduce risk. 

4. AI Agent Risks: AI agents without context or institutional knowledge

You would never drop a new employee into your ITSM environment and expect them to instantly understand your naming conventions, configuration standards, governance rules, or the expectations of your environment. Yet many organizations make that exact mistake with AI agents. MIT reports that 95% of generative AI pilots fail to deliver measurable impact, but it’s likely the lack of data readiness and subsequent lack of AI model training that is causing this. 

Without deliberate onboarding, digital workers will act like untrained hires – slow and prone to mistakes and miscommunications. If you don’t give the same amount of training and patience to your digital workers as you would to a human employee, you risk wasted budget, stalled productivity, and poor adoption because the AI agent simply doesn’t know enough to be useful.

Generic AI models will produce generic outputs unless you give them context. The more available data you have to give context to AI agents on how you want them to behave or information about your environment, the more accurate and helpful your AI agents will become. The onboarding process takes time. You may need several sessions with your AI agent to teach it your expectations, preferences, and how to do tasks. 

Moreover, it’s not just about telling the AI agent what to do, but using it as a co-collaborator to identify ways to work together for better outcomes. And the institutional knowledge you give it should be use-case specific. For example, you would ideally enable:

  • Developer‑focused AI agents to generate catalog items based on how you capture requirements
  • Admin-focused agents to perform CMDB remediation tasks with the same nuance as a seasoned IT analyst.

5. Choosing the wrong tool for the job (native vs. third-party)

CIOs often weigh whether they should rely on native AI inside their ITSM platform or bring in third‑party AI. Both have strengths and limitations.

Native AI excels at end‑user support, ticket summarization, and basic automation. However, usage‑based pricing can make the initial cost skyrocket. It also offers limited customization and fewer opportunities for true competitive gains. 

Third‑party AI platforms offer more customization, flexibility, cross‑platform orchestration, and AI model choice. The investment can be optimized for your organization’s priorities around factors such as latency, quality, computing, and security, leaving strong potential for better time-to-value and long-term business impact. The tradeoff is that they require data and governance maturity that not all organizations inherently have. On the other hand, organizations must lay this groundwork sooner or later for continued AI success. 

A balanced approach is the safest and most cost‑effective way to scale AI in ITSM. Start by fully leveraging your ITSM platform’s native AI capabilities. Push them to their intended limits so you capture the efficiencies they’re designed to deliver. In parallel, you should also experiment with third‑party AI solutions to cover the gaps the native platform can’t support. This hybrid model lets you reduce risk, avoid unnecessary spend, improve accuracy and outcomes, and free your team to focus on higher‑value, strategic work instead of fighting the limitations of any single tool.

Let the “best of times” roll 

Aligning AI agents to specific user needs, enforcing strong security and auditability, maintaining model flexibility, grounding AI in institutional knowledge, and adopting a hybrid tooling approach can help turn AI from a source of uncertainty into a controlled, strategic advantage. With strategies in place to mitigate these AI risks, your ITSM organization can let the “best of times” roll… at least until the next wave of unprecedented risk undulates. 

AI Agent Risks FAQs

What’s the most common way AI agent deployments in ITSM go wrong first?

By not defining who the AI agent is for. McKinsey’s State of AI in 2025 report found that more than two-thirds of organizations use AI in more than one function, but only half use it in three or more, which points to a real gap in specialization. Without designing AI agents around specific personas – end users, ITSM administrators, and ITSM builders – you get a generalist tool that doesn’t serve any of them well.

Do AI agents inherit the security posture of the ITSM platform they connect to?

No, and assuming they do is one of the most common misconceptions. AI agents introduce their own data and security risks that need independent governance, including the ability to access data outside their intended scope, bypass approval workflows, or take actions without human sign-off.

Why is locking into a single AI model or vendor risky for ITSM?

Because the AI landscape moves faster than any prior technology cycle, and today’s leading model won’t necessarily be the best option in 12 months. Organizations without model flexibility face three predictable risks: escalating usage-based costs, lower accuracy as competitors adopt better-performing models, and reduced agility when compliance or business needs shift.

Why do AI agents need onboarding the same way a new employee would?

Because generic models produce generic outputs without context. MIT reports that 95% of generative AI pilots fail to deliver measurable impact, largely tied to a lack of data readiness and model training. Giving AI agents institutional knowledge, naming conventions, governance rules, and environment specifics takes deliberate, use-case-specific onboarding, not a one-time setup.

Should organizations use native ITSM AI or third-party AI tools?

Both have a role. Native AI is strong for end-user support, ticket summarization, and basic automation, but usage-based pricing can escalate, and customization is limited. Third-party AI offers more flexibility and model choice but requires more data and governance maturity. The article recommends a hybrid approach: push native AI to its limits first, then bring in third-party tools to cover what it can’t do.

Richard Mendis
Richard Mendis
Chief Marketing and Strategy Officer at Bytemethod

Richard Mendis is the Chief Marketing and Strategy Officer at Bytemethod.ai, an AI-focused advisory and delivery firm designed to help companies navigate AI complexity and drive tangible business outcomes.

Want ITSM best practice and advice delivered directly to your inbox? Why not sign up for our newsletter? This way you won't miss any of the latest ITSM tips and tricks.

nl subscribe strip imgage

More Topics to Explore

Leave a Reply

Your email address will not be published. Required fields are marked *