Dipping the toe into agentic AI is better than chasing the wave

It is easy to feel like your business is falling behind the adoption curve of agentic AI. The hype is as relentless as it is convincing, but there is no need to rush when the enterprise requirements for a given technology are not yet mature. Our recent research, as part of the 2025 State of AI Application Strategy Report, reveals that 37% of organizations do not have a formal approach to agentic AI and only 5% are plowing ahead with it.

The prudent approach is reading hype with a critical eye while pacing investment during the learning and experimentation phase. Today, deployment of AI with agents highly favors static AI-enabled chatbots completing pre-defined tasks versus dynamic, complex, and autonomous sets of agents that collaborate to achieve goals and improve over time.

Here are five reasons why enterprises should dip their toes into agentic AI rather than chase the wave.

Agentic AI technology is still disrupting itself faster than solutions can be developed. New capabilities, new tools, and new code libraries are emerging regularly. This is a time for prototyping, which can keep up with innovation, but full-scale solutions for the enterprise need slower churn in the stack. A few specific examples:

• Agents need strong reasoning ability to tackle the more complex problems that define goal-oriented autonomous behavior. Models are used for this, and their reasoning abilities are not yet strong enough to replace humans.
• Agents need better memories to tackle complex problems. New memory systems have emerged in response, each with their own special focus: semantic memory, episodic memory, and procedural memory to name a few. There are even examples of memory capability emerging within models.
• Significant functionality jumps are being added to APIs that developers use to build applications. On the client side, browsers are being extended to make it easier to leverage AI. On the server side, OpenAI recently announced multiple updates to their Responses API. Among them are support for Model Context Protocol (MCP) and added support for searching across multiple vector stores.

New standards continue to emerge as thinkers and builders grapple with the interoperability needs of the market. Take MCP proposed by Anthropic and the Agent2Agent (A2A) protocol proposed by Google. One addresses model-to-resources and the other agent-to-agent communication, respectively. Neither solves the complete set of interoperability requirements for agentic AI. To fill these existing gaps, it is reasonable to expect that additional standards will be proposed in the coming months. And while MCP and A2A enjoy increasing support from vendors and customers, proposed standards should be more formalized before an enterprise commits to them.

AI agents are being built and deployed. Early metrics indicate success, but they are measuring business gains in terms of revenue increases (which does not speak to costs) or portion of total requests answered by bots instead of humans (which indicates agents can perform certain tasks, but it does not indicate a full cost comparison). TCO for agentic AI needs:

• A complete bill of materials, including technology, process update/adoption, and human talent requirements such as new skills within existing roles or new roles.
• Architectural guidelines and component architecture diagrams including topology, and interoperability with existing infrastructure and services.
• Success criteria naming key metrics based on impact to the infrastructure such as performance and capacity of computing, network, and storage.

To operate effectively, AI-enabled applications require good data that is secure. Any organization lacking a successful approach for dealing with data silos, data formats, and data normalization will struggle to successfully deploy an agentic AI solution because agentic AI magnifies the data and security weaknesses of a business. It is more than just deploying numerous agents working at machine speed on a network. Security is currently being applied to agents as if they were people or devices. So, human users authorize agents without understanding the full extent to which agents will use, change, or delete their data.

AI-enabled applications require full-stack observability. If an organization has not already implemented a successful full-stack observability practice, it will not be equipped for an agentic AI solution, which critically relies on one for visibility, operations, troubleshooting, and governance.

Consider these five reasons and choose prudence over chasing the hype wave of agentic AI. If more than two present tangible challenges to your adoption plan, more time may be the answer. For a market whose technology stack and standards are still innovating, whose costs are not yet quantified, and whose prerequisites include formidable enterprise-grade capabilities such as comprehensive data management and observability by design, taking it slow with agentic AI is just fine. The first deployment of an agentic AI solution does not prove the value of agents to a business anyway. It proves the infrastructure is ready. Gradual adoption affords enterprises time to find and strengthen the weaknesses before getting in too deep.

For a deeper dive into agentic AI readiness, explore our latest insights in “Policy in Payload: Preparing for AI Agent Architectures.”