Is now the right time to invest in implementing agentic AI?

Software vendors’ pitches are evolving, with  agentic AI beginning to supplant generative AI in their marketing messages. Rather than just generating code or content for human review agentic AI will, they say, follow instructions, make decisions, and take actions much as a human worker would, without human intervention.

It’s more than just a smarter RPA

Agentic AI isn’t just a better version of robotic process automation (RPA): It promises to take enterprises places RPA never could.

“Think of RPA as a train on tracks — it can only go where the tracks are laid. Agentic AI is more like a self-driving car — it can navigate different routes and situations adaptively,” said Paul Chada, co-founder of agentic AI-based software providing startup Doozer AI.

What makes agentic AI autonomous or able to take actions independently is its ability to interpret data, predict outcomes, and make decisions, learning from new data — unlike traditional RPA, which falters when encountering unexpected data, said Cameron Marsh, senior analyst at Nucleus research.

This adaptive nature of agentic AI, according to Chada, can help enterprises increase efficiency by handling complex, variable tasks that traditional RPA can’t manage, such as the roles of a claims adjuster, a loan officer, or a case worker, provided that it has access to the necessary data, workflows, and tools required to complete the task.

Software vendors are already touting agentic AI offerings with access to those resources, including the likes of Salesforce’s Agentforce, Microsoft’s Copilot-based autonomous Agents, ServiceNow’s AI Agents, Google’s Vertex AI Agent Builder, Amazon Bedrock Agents, and IBM’s watsonx Agent Builder, with more are likely to follow.

So, is it time for CIOs to invest in the technology, or is it better to wait?

The early days of a better agent

Agentic AI promises automation without human intervention that is, vendors suggest, easy to implement — but industry analysts and other experts suggest that’s far from the truth for the nascent agentic AI technologies on offer today.

 “A big gap exists between current LLM-based assistants and full-fledged AI agents,” Gartner analyst Tom Coshow wrote in a blog post in early October, noting that to close this gap enterprises will have to learn to build, govern, and trust them.

Even by 2028, Coshow forecast, agentic AI will be available in only one-third of enterprise applications, making it possible for up to “15% of day-to-day work decisions to be made autonomously.”

For Martin Bechard, principal consultant at Dev Consult Canada, “Agentic [AI] is at the early-adopter stage, with initial offerings that have flaws.”

Measuring when agentic AI will be ready for wider use is a fraught question, too, according to Greg Ceccarelli of Tola Capital, an investor in enterprise software startups. “One of the biggest barriers right now in industry is the lack of workflow-specific benchmarks” to compare the performance of agents and humans on a task, he said, and the few that do exist, such as OSWorld, are very academic in nature. “The industry is still at Day 0 right now on this topic.”

Adoption isn’t easy

While vendors portray their agentic AI tools as easy to adopt, it’s not as simple as replacing a human decision maker in a workflow with an agent.

At the simplest level, RPA workflows already designed to work with humans will most likely require significant re-engineering before they are ready for agentic AI, said Dion Hinchcliffe, Vice President of CIO Practice at research firm The Futurum Group. Taking advantage of agentic AI’s ability to process unstructured data, manage contextual decisions, and interact dynamically typically isn’t as simple as updating existing scripts or workflows, he said.

The engineering efforts necessary could include assessing and then exposing the right services, APIs, data, and controls to the agentic platform to ensure that the agent has the context and tools to complete the given task, said Jason Andersen, principal analyst at Moor Insights and Strategy.  

For Anil Clifford, founder of IT consulting firm Eden Digital, enterprises need to shift their overall approach towards automation as the probabilistic nature of agentic AI is fundamentally different to traditional, deterministic, automation.

It’s hard work making work easier

Some platform vendors are already offering low-code and no-code agent development and management platforms, but these are limited in their functionality to building simple agents or modifying templates for agents  built by the vendors themselves, analysts said.

“Creating more complex agents, specifically ones that require customized integrations and nuanced decision-making abilities still demands some technical understanding of data flows, machine learning model tuning, and API integrations,” Futurum’s Hinchcliffe said, adding that there is a learning curve on these platforms and that the migration journey could be resource intensive.

Marsh said most enterprises Nucleus Research has interviewed about experimenting with agentic AI say the learning curve is steeper than vendors claim, especially regarding the depth of customization required to implement agentic AI at scale.

Moor’s Andersen gave a concrete example: while no-code platforms offer integration tools such as connectors to work with other applications, an experienced developer or enterprise architect must first set up an entire backend workflow before an agent can be created to complete a complex task with such an application.

Enterprises still operating legacy applications, for which connectors may be unavailable or limited in functionality, have other concerns.

“These systems often present integration challenges, making it difficult to implement drastic changes to the existing technology stack. It’s like trying to fit a brand new, super smart computer into an old factory that’s still running on machines on old software,” said Shruti Dhumak, a cloud customer engineer at Google, adding that startups or firms born in the cloud might find it easier to adopt agentic AI.

If not now, then when?

Dev Consult’s Bechard sees spending on agentic AI as a bet on the technology’s potential rather than an investment at this stage. But it’s a bet in which the odds may change as agentic AI becomes more capable. “Decision makers will have to experiment to learn or establish a beachhead that can become a strategic advantage if the tech continues to improve,” he said.

Sanjeev Mohan, chief analyst at SanjMo, suggested CIOs should wait and see. He sees no need to spend on agentic AI if existing RPA is working and recommends understand the value of the use case upfront before making the decision to implement agentic AI.

Other analysts suggested a layered or phased adoption of the technology may be the most best path forward.

Eden Digital’s Clifford suggested using agentic AI as a com­ple­ment to RPA, not a re­plac­ement. “This ap­proach al­lows or­gan­i­sa­tions to main­tain their RPA in­vest­ments for struc­tured, repet­itive tasks while grad­u­al­ly in­tro­duc­ing AI agents for more com­plex, con­text-de­pen­dent process­es,” he said.

Hinchcliffe , too, recommended carefully weighing the cost — in money and time — against the benefits in enterprise agility, scalability, and operational efficiency, adding another variable to the equation: RPA vendors are likely to offer agentic AI features themselves — UiPath is already moving in this direction —  that may offer enterprises a safer and faster alternative to implementing agentic AI themselves.