If you’ve paid attention to conversations, headlines, and product releases in the tech world over the last year, you’ll almost certainly have seen the term ‘agentic AI’ being used pretty much everywhere.
Over the course of 2025, AI agents have become the latest stage in the ongoing AI gold rush. If successful, the technology could represent the most significant breakthrough in AI since the release of ChatGPT in 2023.
Agentic AI: Tech’s New Gold Rush
Agentic AI has clearly been all the rage this year – particularly in the ServiceNow ecosystem. But to the casual observer, it might not be immediately clear exactly what this technology involves, how it’s different, or why there has been such a drive to promote and adopt these tools in 2025.
Since 2023, we have all become very familiar with Generative AI tools (like ChatGPT and Claude) that create content based on patterns learned from data. But Agentic AI represents an alternative approach, because the models are trained for a fundamentally different goal – to make autonomous and goal-oriented decisions.
In practice, this makes AI agents much more proactive than the AI tools that most organizations have been enthusiastically adopting for the last two years. This is why tech companies – large and small – have been rushing to offer new Agentic products, tools, and features.
Unsurprisingly, ServiceNow has been no stranger to this trend. In fact, Agentic AI has been at the heart of its product strategy over the last few months:
- September 2024: As part of the Xanadu release, ServiceNow announced the first agentic AI tools for Customer Service Management (CSM) and IT Service Management (ITSM). These features saw a limited release in November 2024 and were built on in subsequent product launches.
- January 2025: At the start of this year, ServiceNow released the first of several tools to manage AI agents: AI Agent Orchestrator and AI Agent Studio. These subsequently became part of the wider AI Control Tower.
- March 2025: The Yokohama release built on the agentic functionality released in Xanadu with a new fleet of AI agents across CRM, HR, IT, and more.
- May 2025: Perhaps the most significant announcement of all: Knowledge 2025 saw the release of the ‘ServiceNow AI platform’. This consolidated Agentic AI functionality from previous releases, and included a range of new governance tools: The AI Control Tower, AI Agent Fabric, and more.
- August 2025: The latest release, Zurich, was released for early access over the last few weeks and is currently scheduled for General Availability in Q4 this year. It includes a number of improvements to Agentic AI functionality, including better LLM integration, industry-specific modules, multilingual support, and a range of new Agentic AI playbooks.
The AI ROI Problem
To understand why Agentic AI has been on everybody’s mind, we have to first consider the most important difference between it and other AI tools. Generative AI requires constant direction, but agentic AI doesn’t. This makes it fundamental to solving one of the biggest issues with AI adoption: Most organizations haven’t seen the benefits.
- According to Stanford University’s recent ‘AI Index Report’, most companies that report a financial impact of using AI estimate that the benefits are comparatively low.
- While 49% of those using AI in service operations report cost savings, these are generally lower than 10%.
- In other key business functions, the results are even lower. When it comes to AI for marketing and sales, supply chain management, or service operations, most respondents reported revenue increases of less than 5%.
- Elsewhere, we can see the same effect in play. According to Brookings, 90% of the most powerful ‘vertical’ AI use cases remain stuck in pilot mode. This refers to any example of AI fully automating a specific business process, creating the most significant potential gains.
To oversimplify, most organizations implementing AI have seen limited ROI or productivity enhancements so far. There are plenty of reasons for this, not least the various governance and compliance challenges many organizations have to overcome before AI can be rolled out at scale.
But another reason is more fundamental: Most AI adoption since 2023 has been generative AI, which, almost by definition, requires a lot of hand-holding. These models are trained to create content and answer questions – but they can’t make effective decisions on what content to create and which questions to ask.
ServiceNow’s Goal: Unlocking AI ROI for Enterprises
For AI to deliver tangible ROI and productivity improvements, it needs to do more tasks at a higher quality.
To understand this in practice, it’s helpful to consider a customer service agent at a large online retail company. So far, the organization has access to a range of ServiceNow generative AI tools that can improve the speed and efficiency of responses:
- Predictive intelligence makes it easier to work out what the problem is and identify the correct solution.
- Document intelligence helps to quickly scan and process documentation, gaining the relevant insights.
- Now Assist tools enable the agent to generate email responses, saving time with manual writing.
While these tools help the agent find information and draft responses, they can’t take over complex tasks.
With agentic AI, the situation starts to look very different because the tools can be trained to make strategic decisions based on clear parameters. Here, an agent could complete the first-level review of incoming support requests, approving the most obviously legitimate requests and triaging the rest back to a human.
If successful, the cost savings of this could be many orders of magnitude greater than most organizations have seen so far. But even then, significant challenges remain.
Governance: The Key to Deploying Effective AI Agents
According to Stanford University, the number of reports to the AI Incidents Database has risen steadily as AI adoption has grown, hitting a record 233 in 2024 – 56.4% more than in 2023.
Respondents to a survey discussed in the same report said that ‘unintended decision-making’ was one of the most common causes of incidents like this, with 51% reporting issues of this kind.
These issues aren’t specific to agentic AI. But any organization looking to bring autonomous tools to market needs to take a long, hard look at these challenges if they plan to achieve any level of success.
To help solve these problems, ServiceNow has invested heavily in technology that would enable customers to build governance guardrails around these products. The AI Agent Orchestrator, for instance, can delegate and coordinate multi-stage tasks between agents. The AI Control Tower also helps to monitor the performance, ROI, and effectiveness of AI agents from a top-down perspective.
In practice, this means there’s a ceiling on the tangible ROI and productivity that most organizations can expect to receive from their generative AI investment. These tools might help individual employees complete their tasks quickly, but until AI can take over entire tasks, or even processes, the impact on a company’s bottom line will remain limited.
This is the conclusion that many organizations have quietly come to over the last few months. And that’s exactly the problem that agentic AI aims to solve.
To understand how this works in practice, let’s look at an example of a common complex process that still struggles to be automated: Product returns. With AI agents, the situation could quickly start to look like this:
- Most product return requests can be viewed and approved by an AI agent.
- Policies defined in the AI Control Tower would decide when a task can be delegated to an agent, and when it can’t. This should involve guardrails to ensure complex tasks are handed over to human agents.
- The AI Agent Orchestrator can handle tasks between different agents, each trained to do different tasks within the process.
- Senior managers can then monitor the AI’s performance and error rate through the AI Control Tower, in order to continually assess which tasks can be covered by AI agents – and which shouldn’t.
Final Thoughts: Can Agentic ROI Really Deliver?
Together, these tools could significantly reduce the workload for customer service agents and a whole range of other employees. If successful, there’s every chance they could help organizations to bridge the divide between AI experimentation and tangible ROI.
But even then, solving these operational and implementation challenges isn’t just a case of switching on the right ServiceNow modules and calling it a day. It requires skills, perseverance, and a keen sense of when AI should and shouldn’t be used.
ServiceNow can provide the tools, but the skills and culture needed to effectively use them can’t be downloaded and installed. For all the talk of game-changing products and features, the question of whether agentic AI can solve the ROI problem remains very much with the organizations using it.
For now, therefore, the success of agentic AI remains very much an open question.