How to Achieve Diffusion in Enterprise AI

It may not be possible without creative destruction

Not to be confused with the diffusion process in computing, this edition of the Enterprise AI newsletter is about diffusion of innovation, and more specifically wide adoption of AI in business. Unless and until the needs for sustainable adoption are fulfilled, the type of AI models employed are irrelevant outside of R&D.

E.M. Rogers developed the Diffusion of Innovation (DOI) Theory way back in 1962. Rogers was a sociologist studying the behavior of people and the influences on decision making in the adoption process of innovation and new products.

Rogers categorized adopters and made estimates of the size of each as follows:

1) Innovators: 2.5%

2) Early Adopters: 13.5%

3) Early Majority: 34%

4) Late Majority: 34%

5) Laggards: 16%

I don’t consider these categories too rigidly, but rather as general profiles. In practice, we’ve found considerable overlap. For example, within Fortune 100 companies and some large public sector organizations like the DoD, one can often find business units and individuals representing all five categories.

DOI is more behavioral economics than technology, but the tech industry lives and dies by it.

The analysis of diffusion is not the study of a rather small niche area that just considers the first use of new technology by firms in closed economy contexts, but rather encompasses many of the large, important questions underlying the international development of economic well-being, the growth of nations and the distribution of wealth. – Paul Stoneman and Giuliana Battisti, Handbook of Economics of Innovation

Adoption of EAI

Surveys on EAI are so common now they are difficult to keep track of. However, the super majority are on GenAI due to expectations for growth in consulting and software companies that produce the majority of reports. I place greater value on first-hand engagement than surveys, but I find surveys interesting and valuable, so I read a fair number and post a few .

My view is that GenAI is still in the early adopter phase. Some groups within a minority of organizations are in the early majority phase, but that’s limited to specific projects like chatbots. On average the enterprise market is still solidly in the early adopter phase. The main barriers to wider adoption are as follows:

1)? The primary beneficiaries of GenAI to date are big tech companies. If we remove market cap from the equation, which is clearly driving much of the investment and related announcements (confirming a bubble in LLMs), Nvidia is the primary beneficiary. A few other companies are beginning to report improvements in revenue and profit from GenAI, but the majority are still consulting and chip companies, not their customers. The results to date strongly suggest that big tech firms and VCs have been backing the wrong horses in LLMs . Customers need products aligned to their needs, not the tech industry.

2)??The big tech backing of LLMs was clearly a strategy intended to control AI, hoping to break through the scale ceiling facing big tech incumbents, including antitrust pressure, increasing conflicts with customers, and systemic security risks . The ensuing LLM hype-storm and historic land grab to transfer the knowledge economy to big tech predictably backfired, causing the largest backlash in the history of the industry, which badly damaged reputations and accelerated antitrust enforcement. Infrastructure needs to be controlled by independent companies free from strategic conflicts for wide diffusion to occur.

3)??We need to focus on data quality, not quantity. The environmental damage alone from the massive investment in LLMs is resulting in a massive pushback, and it’s largely unnecessary. We estimate that our KOS can reduce power consumption by about 90% off the top due to our hyper-focus on data quality, and even more reduction of water. Indeed, for a variety of reasons including security, sovereignty , and vast reduction of environmental impact, we are moving more of the functionality to devices. ?

4)??Customers need a competitive advantage. As Martin Reeves and Jay Barney point out in their recent article in HBR , LLMs don’t provide a sustainable advantage. LLMs were immediately commoditized. GenAI primarily benefits entrenched incumbents who already enjoy other types of advantages, including market power, capital and talent. It requires much more in AI system design to provide customers with a sustainable advantage.

5)??Due to the inherent conflicts of incumbents, DOI requires creative destruction. While we may be able to achieve minor productivity increases with big tech dominance, a Cambrian explosion of the type often hyped by big techs and some VCs, not to mention the dramatic economic impacts predicted by consultants, is essentially impossible without creative destruction. The reason is quite simple. Big techs must protect their pre-existing cash cows, which by definition is counter to the needs of customers, and anti-competitive in nature. New companies will be required to lead diffusion, just as occurred when current incumbents were startups.

An LLM is like an engine,” said a VP at a bank’s AI center of excellence. “No one just wants the engine of a car or a plane; they want a car or a plane. So, there are all these things you need to do to make it part of business processes, so the business can use it.” – Deloitte report, “The State of Generative AI in the Enterprise ”

Conclusion

The bottom line is broad diffusion of EAI first and foremost requires alignment of interests with customers and society. Diffusion will require diversifying the economy as the Internet did, not consolidation as big tech is attempting. ?Authentic pioneers like KYield are much better positioned to meet the needs of customers as our nearly 3 decades of R&D was intentionally free to follow the evidence for the benefits of customers and society, rather than the narrow strategic interests of big tech, conflicts in government, or anyone else. ?