Agentic AI and Gen AI in ITSM Tools: Capabilities, Examples and Use Cases

Two years ago, “AI in ITSM” mostly meant a chatbot on the portal and a suggestion box in the agent console. In October 2026 almost every ITSM vendor sells Gen AI features, and a growing number say they sell agents. The marketing has moved faster than the product, and the difference matters when you are choosing a platform, renewing a contract or deciding how much to trust a feature with real tickets.

This article is a status check. We went through vendor announcements, release notes and documentation for the major ITSM tools, separated what is generally available from what is preview or announced, and added the independent evidence we could find on benefits and risks. We have kept the vendor-claimed numbers apart from the independent ones, because they are not the same kind of evidence. If you want the shorter introduction first, our earlier piece on how agentic AI is rewiring ITSM and ITOM covers the concepts. This one goes tool by tool.

Gen AI assistants and agentic AI: the practical difference

A Gen AI assistant produces content when asked. It summarizes a ticket, drafts a reply, writes a knowledge article, or answers a question from your knowledge base. A person reads the output and decides what to do. An agent takes steps toward a goal. It can look something up, call another system, change a record or trigger a workflow, and it may do several of those in sequence. The step that matters is the last one: when software starts changing records, resetting passwords or approving access, you are no longer governing a writing aid. You are governing a worker.

Gartner’s view of the timeline is useful context. In August 2025 it predicted that by the end of 2025 most enterprise applications would embed assistants, that 40 percent would include task-specific agents by the end of 2026 (from under 5 percent in 2025), that a third of agentic implementations would combine agents with different skills by 2027, and that networks of agents collaborating across applications would follow by 2028. ITSM is one of the first places this is playing out because the work is repetitive, text-heavy and already routed through a workflow engine.

What each ITSM tool offers today

We use three labels throughout. GA means the vendor says it is generally available. Preview covers preview, early access and controlled availability. Announced means we found an announcement but no availability statement. Where a vendor gives no label at all, we say so.

ServiceNow

ServiceNow is the broadest and the most complicated to summarize. Its Australia release became generally available on 5 May 2026 alongside Knowledge 2026, and the announcements included an expanded AI Control Tower for governing agents, an Autonomous Workforce with a Level 1 Service Desk AI Specialist, a conversational layer called Otto, Action Fabric, and a Context Engine. The company’s own materials describe the Level 1 specialist as handling password resets, software access provisioning and network troubleshooting, working 24 hours a day under human-equivalent permissions with an audit trail and a handoff for anything out of scope. Of everything in that release, only the MCP Server is explicitly called generally available. Reports tie the Level 1 specialist to a Q2 2026 general availability date, and we could not confirm that it shipped, so check before you plan around it.

Three other points are worth knowing. The AI Agent Orchestrator coordinates agents across IT, HR, finance and legal, and the platform is model-agnostic. The Moveworks acquisition closed on 15 December 2025 at about $2.85 billion and Moveworks is now the front door for employee requests. And licensing changed: a community post from a partner, which we treat as unofficial, describes five legacy tiers replaced on 9 April 2026 by Foundation, Advanced and Prime, with AI usage metered in assists from a tenant-level pool and legacy SKUs ending sale on 1 July 2026. Ask for assist values and overage rates in writing, because they are not published.

BMC Helix

BMC’s HelixGPT release in May 2025 introduced a no-code Agent Builder, Ops Swarmer, which opens a Teams session from an incident, Catalog Curator, Employee Navigator, Post Mortem Analyzer and Insight Finder. That release gave no availability labels or governance detail. The 26.3 release, delivered on 21 July 2026 with an update on 22 September, adds A2A and MCP support in controlled availability, centralized guardrail management, version comparison for agents, and a log analytics fetch for root cause work. BMC is a strong fit where ITSM and operations management sit close together, but the agentic features are still moving.

Atlassian Jira Service Management

Atlassian’s approach is built on Rovo. Teams build no-code agents in Rovo Studio from instructions, knowledge, skills and scenarios, and they answer on the portal and in Slack or Teams, with a confirmation step before a request is raised. Atlassian’s own customer examples show the pattern: parallel sub-agents that enrich each ticket with similar tickets and a suggested assignee, a likely root cause at incident creation, and an automatic post-incident review written to Confluence. Atlassian also announced in October 2026 that AI change risk assessment workflows are generally available in JSM. Governance detail is thin in the public material, so ask about permissions and audit.

Ivanti Neurons for ITSM

Ivanti announced persona-based agents on its Conversational AI Framework in January 2026, with customer preview from Q1 and general availability planned for later in 2026. By April the capabilities were described as knowledge search with deflection, incident creation, service requests, case status answers and live-agent handoff, introduced on a controlled basis within policy guardrails that are not described in detail. We found no confirmation of general availability.

Freshservice

Freshworks announced Freddy AI Agent Studio and an MCP Gateway for Freshservice on 14 May 2026: no-code agents for Teams, Slack and portals, connectors for Workday and Rippling, and a library of prebuilt workflows. No general availability date was given. A September 2026 update added natural-language workflow building, an email agent, file and image understanding and change recommendations, some of them in early access only.

ManageEngine ServiceDesk Plus

Zia Agents are available in the Enterprise cloud edition only. ManageEngine lists five prebuilt agents plus a no-code Agent Studio, multi-agent orchestration with a master agent over sub-agents, MCP support, cluster analysis for problem grouping, a solution generator and risk prediction. The vendor’s own roundup gives no availability labels and says little about governance. Customers can also bring their own model keys from Google or Azure.

TOPdesk

TOPdesk is a useful example of a conservative approach. AI incident categorization, knowledge generation and AI search are generally available, early access ended on 31 March 2026, and every AI feature is off by default until an administrator turns it on. TOPdesk reports that AI search reduces avoidable incidents by more than 15 percent, which is a vendor claim.

Others worth knowing

Salesforce launched Agentforce IT Service as generally available in February 2026, Slack-first with Teams support and a conversational CMDB, and claims more than 180 organizations selected it within four months. Microsoft’s Service Agent became generally available on 30 June 2026, but it is a Dynamics 365 Customer Service product rather than an ITSM tool, so do not count it as an ITSM option. Automation Anywhere acquired Aisera in November 2025. SolarWinds’ SW1 is an observability and operations agent, not a service desk agent.

The use cases that are real today

The matrix below maps ten common use cases to the tools where we found documented evidence. A blank cell means we found no vendor source in our review, not that the capability is absent, and many vendors give no availability label at all, which is why the amber circle reads “preview or status unclear”.

Reading across the rows, the pattern is clear. Employee-facing virtual agents and ticket summarization are the most mature. Triage, knowledge generation and post-incident writing are real but newer. Problem clustering and change risk are mostly preview or early access. The use cases with the least evidence are the ones that sound most impressive: autonomous major incident resolution and autonomous change approval. We found no vendor source that quantified a major incident communication or change risk result.

Virtual agent deflection and fulfillment

This is the first place most teams start, and the right place. A virtual agent answers from the knowledge base, and the better ones also act: reset a password, grant software access, check a case status, raise a request. The ceiling is set by your knowledge quality and by how many request types have a clean, automatable fulfillment path.

Incident summarization, categorization and triage

Summaries at handoff and on escalation save time without taking a decision away from anyone. Auto-categorization and routing is where the numbers show up in vendor stories, and where the quality of your categories matters most. If your category tree has 400 values and agents pick the first one, the model will learn the habit.

Knowledge generation and resolution suggestions

Drafting articles from resolved tickets is attractive because it attacks the root problem of low-quality knowledge. Treat every generated article as a draft with a named reviewer. A model that summarizes a wrong fix with confidence is a good way to scale a mistake.

Major incident, problem and change

BMC’s Ops Swarmer, Atlassian’s post-incident reviews and ManageEngine’s cluster analysis show where the vendors are heading: agents that assemble context, draft the timeline and point at a likely cause. These are promising and worth piloting, but they depend on a trustworthy CMDB and on good change records, which many organizations do not have. We cover that dependency in our article on how a HAM and CMDB maturity assessment helps.

What the benefits look like, and how much to trust the numbers

The vendor numbers are impressive and mostly unverifiable. ServiceNow says its internal Autonomous Workforce handles more than 90 percent of employee IT requests and that its Level 1 specialist resolves cases 99 percent faster than people. Rolls-Royce reportedly deflected 38,000 tickets in a year with a 34 percent drop in resolution time. Atlassian says it auto-classified more than 90,000 internal tickets, saving about 200 working days. TOPdesk claims AI search cuts avoidable incidents by more than 15 percent. These are vendor statements with no stated baselines.

The independent evidence is thinner. An ITSM.tools survey sponsored by HCLSoftware, with no sample size stated, found nearly 75 percent of organizations using AI in their ITSM tools and half planning to move to agentic AI within 12 months. The top barriers were demonstrating business value (37 percent), adopting the capabilities (33 percent) and balancing cost with service quality (31 percent). A Gartner poll of 147 CIOs and IT leaders in 2025 found 24 percent had deployed a few agents and 4 percent more than a dozen. We found no independent Gartner, Forrester, IDC or HDI benchmark of deflection or resolution time for agentic ITSM. Until one exists, build your own baseline.

How it is evolving

Four shifts are worth tracking, because they will shape contracts you sign this year.

Protocols are becoming table stakes. MCP, the protocol that lets outside agents use a tool’s data and actions, is now generally available in ServiceNow, supported in ManageEngine and offered through Freshservice’s gateway, and BMC has it in controlled availability. The A2A protocol for agents talking to other agents moved into the Linux Foundation’s Agentic AI Foundation in August 2026, alongside MCP. ServiceNow’s A2A support is still announced rather than shipped.

Governance is turning into a product category. ServiceNow’s AI Control Tower, with identity-based least privilege and a kill switch for agents that go off script, and Microsoft’s Agent 365 are competing control planes, and ServiceNow has announced governance of Agent 365 agents in preview. Gartner warned in May 2026 that applying one uniform governance model to every agent will lead to failure and proposed four autonomy levels, matched to different controls.

Pricing is shifting to usage and outcomes. ServiceNow meters AI in assists and Atlassian uses credits. Zendesk charges per verified resolution, and Automation Anywhere has said Aisera pricing will move to work completed. Gartner expects 60 percent of Global 500 companies to embed AI FinOps controls at inference by 2028. Whatever the model, ask what happens to your bill when an agent retries or loops.

The market is consolidating and sorting. Gartner said in June 2025 that more than 40 percent of agentic AI projects will be canceled by the end of 2027 and that only about 130 of the thousands of vendors claiming agentic capability are real. In September 2026 it added that 70 percent of enterprises will abandon agentic AI built through vendor forward-deployed engineering by 2028. We read both as arguments for buying capabilities you can configure and govern yourself.

Risks we would plan for

Security is not theoretical. CVE-2025-12420, nicknamed BodySnatcher, carried a CVSS score of 9.3. It affected ServiceNow Now Assist AI Agents and the Virtual Agent API, and let an unauthenticated attacker impersonate a user knowing only an email address. ServiceNow patched hosted instances on 30 October 2025, the advisory followed in January 2026, and no exploitation in the wild was reported. Separately, AppOmni showed in November 2025 that a ticket description could trigger second-order prompt injection, leading a low-privilege agent to recruit a more privileged one. ServiceNow said the behavior was by design and updated its documentation. The lesson is that ticket text is untrusted input, and agents that read it need the same discipline as any other system that reads it.

Cost opacity. Assist and credit values are rarely published. Gartner names cost as one of the main drivers of cancelled agentic projects.

Hallucination and automation bias. Agents that suggest fixes will be wrong sometimes, and people stop checking when the suggestions are usually right. Plan for accuracy testing and reviewer training.

Data and knowledge quality. We could not find an independent study that quantifies how much CMDB and knowledge quality limit agent results, and the CMDB-readiness material we found came from a CMDB vendor. The reasoning is still sound. Agents act on the relationships and articles you give them, and the vendors’ own tooling, from ServiceNow’s Context Engine to Salesforce’s conversational CMDB, is aimed at exactly this gap. If your CMDB owners and knowledge owners are unnamed, fix that before you pilot anything.

What we would ask any vendor

• Which of these features are generally available in my edition today, in writing?
• What permissions does an agent run with, and can I see every action in an audit log?
• Is there a kill switch, and who can use it?
• How is usage metered, and what happens to cost on retries and loops?
• Is my data used to train or tune models, and where is it processed?
• Which actions require human approval by default, and can I change that per action?
• Which agent protocols do you support, and which are shipped rather than announced?
• How do I export agent definitions and logs if I leave?

A 90-day way in

Where a maturity assessment fits

Most failed agent pilots we hear about did not fail because of the model. They failed because the knowledge was stale, the categories were inconsistent, the CMDB was unreliable or nobody owned the process the agent was supposed to run. An ITSM maturity assessment shows you which of those problems you have before you pay for the feature. You can start with the ITSM and ITOM maturity assessment, pick a specific process from the assessment categories, or read our neutral guide to ITSM tools if a platform decision is already on the table.

Frequently asked questions

Is agentic AI in ITSM ready for production?

Parts of it are. Virtual agents, summaries and categorization are in production at many organizations. Autonomous resolution of complex incidents and changes is mostly preview or announced, and should be piloted in advise or approve-first modes.

Which ITSM tool has the best AI?

It depends on what you need and on which features are generally available in your edition. ServiceNow is broadest, Atlassian and Freshservice are strongest on no-code agent building for smaller teams, BMC is strong where operations are close to service management, and TOPdesk is a cautious option. Check status and cost for your own use cases.

Do we need a clean CMDB first?

For summaries and knowledge answers, not strictly. For impact analysis, change risk and root cause agents, yes. They act on relationships, so wrong relationships produce confident wrong answers.

Will AI agents replace service desk analysts?

Not in the near term on the evidence we have. HDI’s 2025 research found half of respondents expect AI to reduce demand for some jobs while raising it for others. The work shifts toward exceptions, knowledge ownership and agent supervision.

Sources and further reading

• ServiceNow press releases, 5 and 6 May 2026, on the Australia release, AI Control Tower and Workflow Data Fabric: newsroom.servicenow.com
• Enterprise DNA, ServiceNow Knowledge 2026 coverage: enterprisedna.co
• BMC press release, 7 May 2025, and HelixGPT 26.3 documentation: prnewswire.com, docs.helixops.ai
• Atlassian blog posts on Rovo agents and customer stories, June 2026: atlassian.com/blog/jira-service-management
• Ivanti press release, 27 January 2026, and UC Today, 27 April 2026
• Freshworks, 14 May and 15 September 2026 announcements: freshworks.com
• ManageEngine ServiceDesk Plus 2026 half-yearly releases: manageengine.com
• TOPdesk documentation, AI features: docs.topdesk.com
• Salesforce press release, 26 February 2026, and Microsoft Dynamics 365 blog, 30 June 2026
• Gartner press releases: 25 June 2025, 26 August 2025, 26 May 2026, 15 and 29 September 2026: gartner.com/en/newsroom
• ITSM.tools, State of Agentic AI in ITSM 2026 (HCLSoftware sponsored): itsm.tools
• The Hacker News, January 2026, on CVE-2025-12420, and AppOmni Labs, November 2025
• Forbes, 19 August 2026, on A2A joining the Agentic AI Foundation