At the end of the year, we often take stock of the past 365 days. Turns out, the events of the past week mirror much of what happened in 2025. Where is generative AI after three years? What is it being used for? Are we in a bubble? Here are some end-of-year reflections. I have more posts planned over the next week for a few of my own reflections.

The Consultant View

BCG released the AI Adoption Puzzle: Why Usage Is Up but Impact Is Not. This is a good piece focusing on adoption in the enterprise and how it’s not as simple as “do you use AI?” BCG posits there are five stages in applying the technology:

1. Information assistance
2. Task assistance
3. Delegation
4. Semiautonomous collaboration
5. Fully autonomous orchestration (which has not yet been realized)

They make the case that most of the world is defining adoption as reaching stage 1. But the real value and returns occur at stage 4, much later in the progression.

AI emerges as a true collaborator when AI agents plan and execute work with human oversight, shaping workflows rather than simply responding to requests.

This observation helps to explain why we see diverging stories (such as the MIT report about massive use but little returns). It’s also provides a great lens for viewing the report from OpenAI, since (as a model provider) their primary metric for adoption is usage (the number of messages with a chatbot) which provides NO insight into which of the five stages of adoption.

The Model Provider View

OpenAI published the State of Enterprise AI. It’s a survey of workers across 100 companies, mixed with enterprise usage data of ChatGPT. Their report agrees with other recent reports indicating that not only is adoption still growing, but the companies who are at the frontier are pulling ahead. As mentioned, they’re primarily measuring usage, so “frontier” companies are those that are using genAI more, without accounting for the different kinds of uses.

Anthropic published a survey of 500 tech leaders in their 2026 State of AI Agents report. My takeaways:

• More than half of companies surveyed are deploying agents with multi-step workflows; they’re progressing quickly from the ask-answer dynamic to more complex interactions
• 80% claim AI is “delivering measurable financial impact” today (take with a grain of salt – that’s very amorphous and likely inflated)
• Looking to the future, efficiency is still the #1 expected outcome, but other outcomes (improved quality and/or customer satisfaction, cost savings) are not far behind, indicating a shift from productivity to more valuable impacts

It’s worth reading pages 43-44 for the conclusions of the report. One of the takeaways is that context is the real bottleneck. By that they mean two things: one, that giving the models more context provides better results (but with some questionable metrics) and two, that the models need to have the right knowledge to work in the enterprise. That’s been a longtime theme of this blog – RAG is an essential component to useful enterprise AI. They don’t mention RAG specifically but acknowledge data silos will hamper AI’s value.

The Investor View

Menlo Ventures published 2025: The State of Generative AI in the Enterprise, seeking to explain the state of the market and buying patterns…and if we’re in a bubble. Their conclusion is: not yet. Despite the massive inflows of money, enterprises are adopting AI and seeing benefit, and the current trends show no signs of slowing.

This year’s findings make clear that the shift is no longer speculative. Enterprise AI is now a $37 billion market—the fastest-scaling category in software history. Across industries, AI has become core to how work gets done. Enterprises, seeing real returns, are doubling down.

This is more of a “what’s going on in the market” for investment and returns, but the other point I found interesting is that the market has shifted from building to buying. That indicates a maturation of the providers, from raw tech (an LLM chatbot) to more packaged, more productized offerings that don’t require custom scaffolding.

Evidence from Middleware

A report from a middleware company (software that operates between the software you use and the software that powers it) is also worth mentioning: OpenRouter’s State of AI report. Not because it’s important for business, but because it’s the biggest analysis of generative AI usage ever: 100 trillion tokens.

That’s big. Roughly equivalent to 20x the entire Library of Congress! Unfortunately, we don’t know their users or their customers, so the report has little value for companies. But one thing is worth mentioning. If you exclude commercial models and look only at open models, what do you think is their #1 use?

If you guessed programming (the #1 use of commercial models), you’re wrong. If you guessed roleplay, you’re right.

What is that you ask? Yes, it’s a euphemism for using AI for companionship…and sexual dialogue. Which makes sense; open models typically cost less than commercial models, and long conversations use a lot of tokens…and they are likely to have less robust safety features.

Broadly, across the open source ecosystem…roleplay and creative dialogue [are] the top category, likely because open models can be uncensored or more easily customized for fictional persona and story tasks.

Yet another model

OpenAI announced GPT-5.2. I was a little surprised – I thought xAI would launch a new Grok model first. There was a lot of talk last week about Sam Altman declaring a “code red” at OpenAI because Google’s latest model was clearly the best in the market, and there are rumors that some people at OpenAI felt that rolling out 5.2 now was a bit rushed. Who knows?

There are three versions: GPT-5.2 Instant, GPT-5.2 Thinking, and GPT-5.2 Pro. Although the number changing from 5.1 to 5.2 implies it’s an incremental update, it’s a very solid model that performs significantly better than 5.1 on a number of benchmarks. Artificial Analysis puts it (in extra-high thinking mode…which means lots of tokens at inference) in second place, only very slightly behind Gemini 3 Pro. One area where it’s improved a lot is working with Excel and PowerPoint.

But how excited should we be about a model that is still wrong 6% of the time?

Open Models: Meta Out, NVIDIA In

China has completely dominated the space for so-called “open” models (open models means the weights are published for free, so if you have the hardware you can download and run the models yourself). The only open models from the US were from Meta (Llama) and one from OpenAI…and Meta’s model was not competitive. Meta has hinted lately that they’re going to save their best models for use on the Meta platform, calling into question the future of the Llama open models.

NVIDIA appears to be stepping in to fill the void, releasing the Nemotron-3 family of models. Most interesting is that these models are much more open than Meta’s – not only did they release weights, but they also provided training datasets and reinforcement learning libraries to make it easy for people to train custom models based on Nemotron.

As of December 17, Artificial Analysis ranked Nemotron 3 Nano as the 10^th best open source model, which sounds low until you realize it’s the smallest and least capable of the three. So it will be interesting to see where Super and Ultra appear when they are tested, they may give the open Chinese models some competition.

---

---

Adoption

We have seen a steady stream of analyses and reports about AI adoption and value in the enterprise that say the same thing. That same thing is that:

• Everyone’s doing it
• Most aren’t getting a big impact or broad returns
• But the few that are, are seeing high value
• From specific, focused use cases
• And, most importantly, those few are seeing accelerating value

BCG’s insights about adoption helps to provide clarity in interpreting those results. Adoption is not a simple yes or no, this is a complex technology that can be applied in many ways, and its proper application yields outsized returns. Most companies haven’t figured out or deployed the right applications yet (no criticism to them – it takes time).

Value in the Enterprise

Since OpenAI measures the use of their models, they don’t have insight into the context in which those models are deployed, so their conclusions provide only a partial picture, and are self-serving in that convincing companies that the solution is “more use” rather than “use for the right things” drives more business. ChatGPT for Enterprise has extremely limited RAG capabilities, so most of this use is static, and does not leverage enterprise knowledge in real time.

We know that the high-value use cases come from RAG. You can’t get to BCG’s stage 4 (at least not at scale) without sophisticated RAG.

I also noticed something very interesting. OpenAI surveyed 9,000 people and 75% of them “report that using AI at work has improved either the speed or the quality of their output.” That means that one-quarter of the people who are using AI don’t see any benefit! Again, this confirms that just using AI (stage 1 adoption) is not what matters. The value comes from using AI for the right things.

Yes, We are(n’t) in an AI Bubble

Oh, and on whether or not we’re in a bubble? Kind of like the debate over whether or not we’re on the path to AGI, there are many people much smarter than me who hold diametrically opposing views. Alberto Romero explores this topic in detail in Why Industry Leaders are Betting on Mutually Exclusive Futures. Here’s his conclusion:

Which one is the true portrayal? How would I know? Sutskever and Karpathy don’t. LeCun and Altman don’t. Sequoia, a16z, and Goldman Sachs don’t. Trump and Xi don’t. Industry, academia, investors, world leaders: no one has a clue.

And Scott Galloway remarked (about investing in the market in general, but much of the market’s gains have been driven by AI):

I have access to the world’s brightest people, and I still don’t know what to do. I literally walk around most of the time going, I should buy — no I should sell. 

My point is, if you’re out there and you’re not sure what to do, welcome to the club.

What a fun year it’s been!