The notion of AI agents is just the latest in the ongoing wash-with-AI process, but as usual there’s a grain of truth in all the nonsense. There are, to be sure, negative views on AI agents, but there are also some thoughtful pieces that, while they are likely victims of the normal survey bias problems, show a lot of enterprise interest. Some comments correspond to what enterprises tell me, but I have some different impressions of the space overall after my own chats.
Let me start by saying that I’m not including enterprise comments on the use of public generative AI, because those who comment to me say they aren’t involved in this part of AI. Even when the services are paid for, enterprises say they’re not coordinated by IT (about half think they should be), and thus don’t see what’s happening with them. They do offer some dismissive remarks on the value, but I don’t think I can rely on their being objective. Agents are a different story.
Enterprises seem to take a rather broad and diffuse view of what an “agent” is, which is hardly surprising given the market seems to have its own problems nailing down a definition. No significant number of enterprises offer a specific definition of their own, but the view I draw from their comments is that agentic AI is a limited-in-many-dimensions form of or alternative to generative AI. Rather than being a SaaS-like service trained on the general online knowledge pool, agents are specialized and contained in nature, not designed to offer broad assistance but rather to perform a very specific task.
This definition allows agents to divide into three general groups. One is the workflow agent, which performs its function in the context of current workflows, and thus is an augmented form of a software component. The second is the integrated tool, added to an existing application (often operations automation) to provide interpretation and recommendations, and the third an expert companion that can address more general questions but within a very specific context. I think that the flexibility in the way enterprises use the term is the explanation of how enterprise adoption rates on agentic AI seem high and growing faster than expected. Many, in fact, use “agent” to include LLMs similar to the online generative AI type, but trained and run only on company data. That may be due in part to the flexible definition, and in part to AI vendors pushing the AI term that gets the most ink.
The common definitions of an AI agent stress autonomy, which is interesting to me because enterprises don’t seem to think much about that, and in fact seem to fear it as much or more as seeking it. In the workflow missions, they expect AI to do its thing independently, as they would with any software component. In the expert companion missions, they obviously want it to answer them, and in the integrated tool missions, they want to get advice and perhaps approve actions, but rarely want an agent to do something on its own. They do believe that might change with their experience with the agent, though.
There does seem to be some difference in the way the three agent groups are developing. Some enterprises said they had applications in multiple groups. Workflow agents are cited by more enterprises than other types (about 40% gave this as their first group mention), but a significant contribution to this group is the accounts where IBM is the dominant strategic interest. Expert companion gets the least mentions overall (22%) but most of the enterprises who say they have their own AI cluster listed it, and integrated tool gets a 35% mention rate, with almost two-thirds citing an operations mission.
In terms of interest, what enterprises say is that they like the notion of agents based on foundation models, but draw on their own data. They don’t like having any form of RAG or MCP or A2A used with the public AI models because they don’t trust them. A third say that they would be interested in a cloud-hosted agent that fit into any of the groups, but they’d want it from a trusted company (IBM and Microsoft are the two regularly mentioned to me, Oracle and HPE are occasionally mentioned) to provide the agent and hosting, for security/governance reasons.
The notion of self-hosting is also shifting more to focus on agents. While only about 15% of enterprises either do or are interested in doing their own generalized AI cluster, over three-quarters either already host or plan to host their own agents (remember the enterprise definition, though), and all admit they’re looking at the notion. Of those who are at least planning agent hosting, about a third think that might evolve into a general cluster deployment, but even that group isn’t currently doing anything to insure that agent hosting relies on common tools and technology so as to facilitate this generalization.
I think there are some interesting things to learn here, both in terms of what we think/know about AI and in terms of what enterprises are likely to value. Let’s summarize some.
First, enterprise professionals are far from convinced that the popular online generative AI models are worth paying for. The problem, according to those who have cited a cause, is their populism. To make something useful to all makes it trivial, they say. How much can a company save, or make, by reducing the time that people spend writing an email, or even by improving the quality of writing? Not enough to make it worthwhile to invest, my contacts say. If AI is to do something truly helpful, it has to impact the workers who do critical things, and it has to be specificized to those worker’s needs. Artificial general intelligence (AGI) is a case in point. Why spend a zillion dollars and enough power to drive a small city just to create something that can think like a human, when there are plenty of humans you can hire? Enterprises who understand the cost of AI understand that to be beneficial, it has to do something that humans can’t do well and target those who perform critical functions for the business. In most cases, that means narrow expertise, specialization. That, to enterprises, is what agents should be about.
