OK, yesterday I took the markets and hedge funds to task because of the story that AI was somehow going to eat software. If I’m right (which, of course, I think I am) then is there no value to AI? Robert Heinlein’s line from “The Moon is a Harsh Mistress” comes to mind, “There ain’t no such thing as a free lunch”, or the famous acronym TANSTAAFL. The question is not, has never been, and will never be whether AI can do this thing or that, it’s whether anyone will pay to make it do it. Tech advances on what people and businesses will invest in.

My contention is that consumers don’t want to pay for anything, and businesses want an ROI that makes a business case. I don’t think that hosted-chatbot AI, or any form of hosted AI, can build the maximum value AI needs to achieve. To get to that, I think you need a model of AI agent that enterprises have favored from the first, one that embeds AI into current application workflows at appropriate (makes-a-business-case) points. But supplementing workflows won’t generate a boom. To get to that we need to look to applications that AI could unlock, ones that don’t exist without AI because they can’t make the business case. The question for this year is whether we can focus on both these things.

Will 2026 be the year agent-AI gets real, or the year it gets hammered? That debate has vocal advocates on both sides, particularly in the tech media, but the outcome of the debate won’t answer the question. Only the unfolding developments in AI can do that, and the mechanism for action is simple—return on investment. AI gets real if there’s a real direction, at least one but preferably many, that generates significant ROI. The AI players will flock to such a direction because real benefits alone drive real investment in the long term.

The question, of course is not only whether there is such a “real direction” but also what that direction is. Can we know it today? The answer, I think, has to come by tracing backward from the notion of ROI to how it could be achieved.

Right now, enterprises tell me that they’re looking for ROIs in the 25-35% range. Since ROI is the ratio between benefit and cost, this means that the capital cost of deploying AI would have to be no more than three and four times the annual benefit. Right now, enterprises are looking at two-and-a-half to three times, due to a combination of their uncertainty about sizing benefits correctly, and concerns that the technology write-down for AI could be faster than for traditional data-center gear.

One of the unsung beauties of AI agents is that enterprises find they require only modest-sized models, which means they can be hosted on two GPU-equipped servers, one being primarily for hot-standby backup. We’re very early in agent self-hosting but so far, enterprises say that there’s not much value in or need for the creation of a large AI cluster for resource-sharing. AI agents tend to be more latency-sensitive, since they often fit into a workflow like a software component of an application, or are used in operations roles and other event-handling, where latency is clearly something to minimize. Having a pool of GPU servers makes it hard to guarantee response time. Thus, you would normally justify agent hosting costs based on a single, or a few related, applications.

The key recommendation of the enterprises with deployment experience for AI agents, in fact, could be paraphrased from an old portion-control saying, as “Don’t let your eyes get bigger than your wallet!” Successful justification depends in large part on not going crazy with model-hosting resources, particularly early on. About half of successful enterprises think that starting very small, with perhaps two servers with GPUs or a single high-availability one, is the best approach.

Proving out, and sizing, the benefits and costs is the first critical step. You need to know what the payoff of the first trial is, but also the extent to which both cost and benefits scale. One thing about projects aimed at productivity enhancement have tended to show is that the best targets are almost always taken first, and that the cost/benefit ratio is sure to change as you eat up the pool of targets.

You may wonder at this point what the last paragraphs have to do with AI agent progress in 2026, so let’s get to that. Answer: It’s key. This sort of process doesn’t happen quickly, even for established technologies. It’s rare to see any significant project hit its stride in the first year, and for at least 80% of enterprises, AI agents aren’t really even in the early stages.

As I’ve said in many blogs, the real transformational opportunity in IT is real-time applications that escape from their current on-a-premises confines. Of almost 600 enterprises I’ve heard from, only 181 had any real-time missions in place, and so could hope to expand. From what I hear, we can expect to see the number of potential successes roughly triple this year, but that doesn’t add up to enough to move the needle of IT spending significantly. Remember, we only had 48 real candidates for success in real-time applications that could drive significant IT investment, and none had any immediate plans to extend the applications, meaning that few projects are likely to get beyond the dreaming phase in 2026.

What this means is that the likely most we can expect from enterprise-hosted AI agents in 2026 is better ink created by momentum in the space. They’re not likely to take up much of the slack in spending if the big AI bubble bursts.

It doesn’t have to. I’m working through the details for a blog later this week, but so far it looks pretty clear that proper targeting of centrally hosted AI in 2026 can sustain the market. There plenty of business case available to give us enough time to get the real AI agent opportunity going, providing of course that we can properly target to address it. And I’d submit that what we need is so obvious that even AI knows it. It’s private 5G/6G networks.

Yes, I know that I’ve been constantly saying that private 5G was overhyped, and I still believe that. The problem is that it’s being promoted in a vacuum, a classic “Field of Dreams” model. There has to be a driver, so I asked AI what it would be (Gemini Pro 3) and here’s the result:

Private 6G networks will provide dedicated, high-security, ultra-low latency, and high-capacity connectivity, building on 5G capabilities for specialized industrial, manufacturing, and enterprise applications. Expected to emerge post-commercialization, they will integrate AI-driven automation, terahertz communications, and digital twins to enable advanced IoT, autonomous drones, and immersive, 3D holographic experiences.

The bold text is mine, and the reference to “digital twins” was from AI with no suggestions of mine to drive it. The truth is that you can’t do real-time systems at scale without digital twin technology, because you can’t expect distributed real-time systems to be non-contextual in behavior, so you need to model them to understand, control them, and optimize them. And who reinforces this but Nvidia, announcing a partner deal with Dessault Systems on what they call “visual twins” for industrial AI?

Technology presents new options, but it doesn’t guarantee new value, new benefits, new business cases. We have to fit AI into a big picture because the size of the picture we can fit it in determines how much value it will deliver. We may, just may, be in the early stages of this now, and how fast we progress will determine whether AI is a revolution this year, or just another tech element battered around by hedge fund traders and hype.