Every year, most enterprises and telcos undertake a technology assessment intended to drive the budgeting for the following year. This can run between mid-September and mid-November, and most of these reviews are now complete. I’ve gotten a picture of the results of 308 enterprise reviews and 54 network operator reviews, and so I can now take a look at the issues, both with regard to specific technologies and the enterprise/operator grouping. I’m going to start this with a look at AI.
One thing that’s clear is that you can divide enterprises into two groups. The largest by far (274 of 308) is the “AI realist” group, and the other a group less influenced by active AI planning and thus tending to follow the media stories on AI in their vision of their future use. In the first group, 41 say that “generative AI” or “large-language models” are critical, and in the latter all 34 are generative-AI-focused.
Let me dispose of this last group first. Of the 34, 11 see AI dominantly in customer support or personal productivity roles, and so actually need LLM technology. The remainder have ceded AI decisions to line organizations, and are experiencing AI primarily through public services from companies like Microsoft or OpenAI, or in search applications. Only two of this group expressed any interest in self-hosting AI, and this group didn’t express any specific commitments regarding 2025 spending on AI; “stay the course” was the summary view.
In our first group, the AI realists, the 41 who see generative AI as critical are all looking at self-hosting, and of this group 29 are looking at AI in customer-facing chatbot missions. The other 233 are looking at missions they don’t see as generative AI missions at all, but rather as “AI” or “machine learning” (ML). Two missions dominate this group, business intelligence (199), and real-time/edge (78) applications linked to IoT.
The reason this critical group of 233 doesn’t toe the generative AI line is not so much the resource issues associated with LLM training and usage (they’re a factor for only the 78 who are looking at edge missions) but a lack of utility of generative AI in the missions they’re seeing. For nearly all, their focus has evolved from “generative AI service” to “self-hosted LLMs” to small language models and machine learning applications, largely due to the fact that both the missions that dominate the group are based on analysis of a very limited amount of data, where “generative AI” is usually trained on a vast amount, via the Internet or social-media interactions. We’ll get to how this might (emphasis on the qualifier is important) change in the future, at a later point in this blog.
I get the sense that SLMs are becoming the preferred AI/ML platform, though only about a quarter of my 233 make that point explicitly. The challenge for AI today may well be that just about everything being published online and promoted that’s related to AI is really related to LLMs, and just about everything that’s important to LLM evolution, like RAG, is not much of a factor to SLMs. If you’d like to read a nice piece on SLM technology, check THIS out (the full survey paper is HERE). Of the 233 companies, 188 said they were led to SLMs by a trusted strategic vendor, and of that number, 72 said they didn’t at first even realize they had shifted to an SLM focus until they were heavily involved in planning or trials.
Within the 233, there were 58 enterprises who seemed thoroughly familiar with SLM-based AI. These people say that the reason for SLM focus is simple; businesses aren’t run through some single uber-mind thinking and planning, but as a connected set of activities that are locally handled, and whose outcomes are then fed upward. Each activity, and each combinatory layer element above, are really SLM applications, they say. Focusing on SLMs means that the AI elements are now less demanding in terms of both training and running, and that keeping data in house for sovereignty reasons is facilitated. This is what AI types mean by “domain-specific applications”.
What’s particularly interesting here is the view of this group of 233 enterprises on “personal productivity” missions of AI. The point made, by 175 of that group, was that the basic generative AI support might be fine for helping with generic documents, emails, etc. this wasn’t really their goal. They point out that productivity benefits are proportional to the unit value of labor of the targets, and that those with higher unit values of labor tend to have them because of special skill requirements. Its support of those skills that are valuable, which means domain-specific knowledge.
Within this group of 175, 59 say that they believe that you could inject domain knowledge via RAG into LLMs, but the rest say that they believe SLMs would be better, or at least that an open-source LLM that would scale down to a small resource footprint and could be trained on company data would be best. However, 128 admit that it would be possible to use a pre-trained, RAG-augmented, form of generative AI for applications that involved having “typical” people generating questions.
Edge IoT-related SLM missions are getting increased attention. The 78 who had already started work on edge AI said that you could generally classify IoT applications as being “process-bound” or “independent” in terms of sensor/effector behavior. The former applications linked IoT elements to real-world processes that defined what was being done, and the latter used IoT to gather information about the real world, from which the things going on could be deduced. They point out that the former class of applications don’t require as much AI analysis; an assembly line sets the mission and coordinates the elements. The latter has to accommodate a bunch of real-world things that may be individually self-determining, and so the state of the system depends on how individual behavior and mass stimuli might impact the collection of things overall.
How about AI budgets for that AI realist group? There was a wide range of comments in that area, ranging from an 18% increase to a 300% increase over this year, but over 90% of the group said that AI spending would be the largest source of overall IT spending, and the 78 AI edge companies said that AI would increase their IoT spending by over 50% once it ramped up fully, but expected only about a 40% increase in 2025.
The sense of all of this, IMHO, is that most of what we hear about AI isn’t really moving the ball. Could online generative AI improve search, help you write emails, and so forth? Sure, but where’s the profit in that for the provider of the service or the user? According to enterprises, their 2025 AI focus is going to be on small models in domain-specific missions.
I want to emphasize that we’re still early in this SLM game, though. Remember that only a quarter of the companies whose comments identified them as SLM prospects self-classified their activities as “SLM”; most were still talking about machine learning. I think that the pace of SLM in 2025 will depend in part on the vendors (like IBM, Dell, and HPE), and on the media, who need to start giving more space to the topic. If both sources of SLM interest ramp up, then even these budget estimates may prove conservative. This, I believe, is the real AI.
