- Agentic AI refers to automation systems that can make decisions and take actions with minimal human intervention
- Also known as iterative reasoning, agentic AI promises to transform automation across industries
- Achieving agentic AI at scale will require thorough integration, governance and oversight for complex workflows, analysts say
Artificial intelligence (AI) systems that can take action with minimal human intervention are poised to transform the future of automation. Also known as iterative reasoning, agentic AI is a leap toward autonomous systems that handle complex, real-world problems and decision-making. This all sounds very cool, but are enterprises ready for the shift — and what could go wrong?
At the Gartner IT Symposium this month, tech leaders said we’re on the brink of seeing these autonomous systems play a central role in how businesses operate. Gartner named agentic AI as the top tech trend for 2025, and both Google and Microsoft in recent months have talked up their work on agentic AI.
Meanwhile, analysts are optimistic but remain cautious. Agentic AI certainly holds promise, but its practical applications may initially be limited.
Ritu Jyoti, GVP of AI and Data Market Research at IDC, told Fierce Network she expects to “see significant advancements in this technology within the near future,” and said it will become “increasingly prevalent in business applications within the next few years.”
Although, it's still important to manage expectations as the full capabilities of truly autonomous, agentic AI might still be “a ways off.”
"It is early days and there are quite a few challenges and concerns,” Jyoti added.
Not so fast
During his keynote at the Gartner symposium, Nvidia CEO Jensen Huang discussed the transformative potential of agentic AI. He envisioned a future where organizations employ large populations of agentic AI workers that operate much like their human counterparts.
“The next AI layer is agentic.... That is easiest for me to imagine in the long term. And I can call long term being six months, because the world's changing and so very, very soon, it's very likely that we'll have AI employees,” Huang said.
However, many enterprises “aren’t necessarily ready for this type of transformation,” said Jason Andersen, VP and principal analyst at Moor Insights & Strategy. There are still many hurdles for organizations to get over before they’re ready to implement agentic AI systems.
That includes figuring out the well-formed workflows and services to support them. “But it’s more than that,” Andersen added. “For instance, who’s responsible when the agent does something wrong? Also, how will you know that something went wrong, and how to fix it? So, there is a readiness issue if you want to take an agent beyond a simple process.”
Andersen predicts we won’t actually see “enterprise-scale solutions” being deployed until 2026. But there will be “many stories of small-scale agentic solutions” in 2025.
For example, work is already being done for some point solutions that include human interaction. Salesforce now deploys agent-based solutions that fit the departmental model well. And Jensen noted that Nvidia has been working with SAP, ServiceNow and others to put such agents into their systems.
Agents might soon be able to conduct first interviews on behalf of recruiters, an application already being trialed in Japan, where candidate and recruiter agents are interacting through technology developed by companies like Alt.
The risk of rushing in
In reality, agentic AI is still in a phase of practical experimentation rather than wholesale adoption. Early, successful applications have had “specific, measurable processes where the ROI is clear and risks can be contained,” said Dana Daher, practice leader at HFS Research.
“The downsides of rushing into this space are significant,” Daher told Fierce, and the risk/reward ratio isn't yet there for most use cases.
“We're talking about autonomous systems making critical decisions without proper oversight, loss of human operational knowledge as processes get automated, impact on the workforce and culture, and regulatory exposure from poorly controlled AI systems," she said.
According to her, the infrastructure for true agency, which should include goal-setting, planning, self-correction, is still maturing.
Getting to real agentic AI will require some advances in architectures for multi-agent systems and causal reasoning, frameworks for defining and enforcing operational boundaries, observability, control mechanisms and major updates to risk management and governance approaches.
“There also needs to be some real consideration of what should be ‘agentified’,” Daher added.
Emerging solutions
Beyond generating hype, vendors will need to do more to help enterprises along their agentic AI journey. And some already are.
Andersen noted he’s starting to see a "different solutions space that may have the initial readiness requirements in place.” AI agents are beginning to enhance rigid robotic process automation (RPA) systems, for instance, potentially reducing human intervention when issues arise. Companies experienced in RPA are particularly well-positioned to adopt AI agents.
At the Gartner Symposium, Timothy Kim, a solution engineer at UiPath, described how agentic AI will coexist with traditional RPA systems. According to Kim, UiPath envisions agentic automation as the bridge between AI agents, bots and human workers.
“Agentic automation is about giving the agents that work for you the agency to make decisions,” Kim said.
And elsewhere, new generations of AI models are already getting closer to autonomy. IBM’s new generation of Granite models, for example, have the ability to do function holding, Maryam Ashoori, IBM’s director of product management for watsonx.ai, told Fierce at the symposium. That means when a request comes in, the AI model can break down different steps to figure out how to resolve the request on its own.
Vultr has also launched a new service that simplifies setting up agentic AI for enterprises by providing a ready-to-use vector store for retrieval augmented generation (RAG). Vultr CMO Kevin Cochrane told Fierce that companies can use the service to set up agentic AI applications in "30 minutes or less," depending on their existing data infrastructure.
Ultimately, agentic AI will bring enterprises capabilities that LLM-based solutions haven’t been able to, said Andersen. An agentic application provides the LLM with a degree of contextual guidance that is “very difficult by using prompt engineering alone,” he added, resulting in more predictability and traceability. Additionally, agents can hold previous information and prompts for the LLM so a human doesn’t have to provide the same information repeatedly.
“Regarding potential, I think the agentic trend is worth the hype,” he said.