TECHNOLOGICAL AND ECONOMIC DRIVERS OF AI INNOVATION

The idea that a computer might be able to achieve human-like intelligence dates back to the early days of the modern computing era. In 1943, Warren McCulloch and Walter Pitts outlined a simple model based on Boolean logic which mimicked how neurons in the brain process information.Footnote ⁴ British computing and AI pioneer Alan Turing considered how a computer might be judged to have attained human-like intelligence in his 1950 paper, ‘Can machines think?’Footnote ⁵ While the Turing test is no longer considered by many researchers to be the authoritative measure of whether a computer has achieved ‘intelligence’, it has provided a useful framework for later research, both technical and theoretical. The phrase ‘artificial intelligence’ emerged from a landmark 1956 summer workshop at Dartmouth College in the United States. The questions discussed and approaches proposed at this workshop set the direction and priorities for AI research for the following several decades.

The current wave of GenAI innovations has taken the technology out of computing research laboratories and, for the first time, placed it in front of non-technical users. The enthusiastic take-up of ChatGPT, with a reported 400 million active weekly users in February 2025, has demonstrated the public appetite for being able to interact in a conversational style with a technology that can generate ‘original’ text and images.Footnote ⁶ In this sense, Turing’s proposition that a computer can be considered ‘intelligent’ when it can trick users into thinking its outputs were created by a human may have been achieved.

This generative capability of new AI systems can be seen as marking a new stage in the ongoing series of information revolutions going back to the use of movable type in the mid-15th Century. Each of these revolutions was characterised by a decreasing role for humans in the creation, reproduction and sharing of information. The first printing presses removed the need for books to be manually written and illustrated while the more recent computing-based revolutions saw increased automation of data inputting and processing. The core argument presented in this paper is that the current AI-driven revolution goes further than any previous one in taking humans out of the equation. Not only can GenAI systems create new information they can also be trained on synthetic data created by other AI technologies. In this sense, true autonomy of machines is becoming possible and presents both threats and opportunities for information workers.

AI can be seen as the latest in a long line of information technology revolutions, starting with moveable type in the 15th Century

While the resources, human, technical and financial, that have gone into creating the ‘magic’ of GenAI systems have been considerable, we are now at a turning point in the evolution of AI. The frontier AI models created by OpenAI, Anthropic, Google and Meta have required billions of dollars and massive data processing infrastructure to create but are now in danger of being commoditised through the rise of open-source releases and the ability of developers to use distillation approaches to create cheaper and smaller models out of larger ones. In January 2024, Chinese startup DeepSeek released its open-source AI reasoning model which it claimed was comparable in performance to OpenAI’s o1 offering, ran twice as fast and was developed at 10% the cost. Within weeks its AI Assistant had overtaken ChatGPT to become the top-rated free application on Apple’s app store in the USFootnote ⁷ and forced a major re-evaluation of their research and investment priorities by frontier model developers. This all points towards the commoditisation of AI models, as differentiating between competing offerings becomes more difficult and levels of performance that are acceptable to most users are the norm. While this may present an existential threat for the business models of some frontier model developers, it promises to provide a fertile environment for developers building industry and task-specific applications with those models. The combination of open-source models, distillation and reduced processing costs will encourage innovators to apply AI technologies to business processes in ways that were not technically or economically possible before.

In her influential research on technological revolutions and what she termed ‘techno-economic paradigms’, Carlota Perez described the key drivers of technology-driven change.Footnote ^8, Footnote ⁹ These drivers include a set of interrelated breakthroughs, the capacity to transform the economy, a core low-cost input and new infrastructures to channel resources. The combination of massively powerful computing infrastructure in the form of data centres and the almost daily breakthroughs in GenAI models coupled with rapidly decreasing processing costs fit Perez’s model and, I would argue, we are set to begin a new techno-economic paradigm where AI drives profound change across the global economy.

As the infrastructure powering this revolution becomes cheaper to access whilst also increasing in its capacity to process data and power AI models, we will see more developers build applications on top of those models. These will have a far greater impact than so far seen with GenAI and will mostly operate unseen in the background. As the attention on investors and developers shifts to applications for end users and businesses we can expect to see some of the froth coming off valuations of model developers such as OpenAI and Anthropic. This may be construed as the AI bubble bursting, but it should be seen as a rational shift in activity towards generating real value through AI applications.

During the dotcom bust of 2000 and 2001, many commentators claimed the internet was a flash in the pan, but in reality it was a realignment of where the real value lay in the internet economy. The massive investments during the late 1990s in building out fibre optic networks and developing the ecosystem of web technologies provided the springboard for new companies to build services which users and businesses wanted. Amazon and Google were survivors of this shakeout while Facebook and other social platforms, as well as business applications such as Salesforce and Atlassian, were built on the foundations that had been created over the previous decade. We should look on the current environment of AI innovations in the same light to anticipate what might be next.

As AI systems become more efficient, consume less energy and processing resources and, as a result, become cheaper to access, we can expect to see the Jevons Paradox play out. In the mid-19th Century British economist William Stanley Jevons observed that as technical innovations increased the efficiency of coal mining there was a resulting increase in the consumption of coal as prices fell. This is like building more and bigger roads to reduce traffic congestion, but this often results in more people using their cars and negates the intended benefits. As commoditisation of the underpinning models powering the current AI revolution makes the outputs of GenAI systems cheaper to access it is likely more developers will create applications to take advantage of this.

TAKING HUMANS OUT OF THE LOOP

One of the core drivers of autonomous AI systems will be the rise of agentic AI applications that will process data and make decisions on our behalf. At the moment these are primarily used in customer facing operations to power chatbot interfaces that can answer customer queries. Technology analyst firm Gartner predicts that by 2029 AI agents will resolve 80% of common customer service issues, lowering operational costs by 30%.Footnote ¹⁰ While attractive to businesses from a cost-saving perspective, these agentic AI systems are relatively basic and are low-risk in terms of their ability to cause organisational harm. More significant in terms of driving the AI revolution forward will be the development of more autonomous systems that can make decisions on behalf of an organisation without the need for human inputs. According to Deloitte, more than $2bn was invested in agentic AI startups between 2022 and 2024, with 50% of organisations using GenAI predicted to deploy agentic systems by 2027.Footnote ¹¹

However, a lack of interoperability between agentic AI systems is a significant roadblock to more widespread adoption. By their nature agentic systems need to communicate with other applications and pull data in from a variety of sources, often from external locations. Competing and proprietary protocols powering agentic systems makes the sharing of data and processing capabilities difficult and often impossible. A new open-source agentic standard, Model Context Protocol (MCP), promises to change that.

It took the open SMTP email and HTTP web protocols to drive internet adoption in the 1990s at the expense of the previously dominant proprietary and closed networks offered by AOL, Compuserve and Prodigy. MCP has the potential to do the same for agentic AI. Released by Anthropic in late 2024 it already has support from industry giants such as Microsoft, Cloudflare and OpenAI. Thousands of developers are building agentic systems using MCP and it is likely a tipping point will be reached soon whereby it becomes a key enabler of autonomous agentic innovations.

Alongside this we are seeing rapid releases of reasoning models from major developers including OpenAI and Google. These break down complex problems into smaller steps utilising explicit logical reasoning. These are a major stepping stone to more reliable and trusted GenAI applications. The problem of hallucinations, whereby LLMs make up answers that look correct based on the data they were trained on, but which are wrong, can be reduced through a reasoning approach. This provides a more transparent overview of how an LLM arrived at an answer and with better prompting alongside supervised finetuning and reinforcement learning we can expect more reliable outputs.

COMPETING IN THE AGE OF AI

The ability of increasingly intelligent and faster AI systems to perform research, analysis, and create outputs will change the nature of much information work. While the current AI research offerings from Anthropic, OpenAI and Google are not 100% reliable in their outputs, they are good enough for many organisations. Google’s NotebookLM Plus product, for example, provides a more personalised experience for users that allows research projects to be created and documents and other digital resources to be uploaded for analysis. The ability to remember previous prompts and store files and answers for later reference is a significant step forward for anyone using GenAI as a research tool. Issues of security and reliability aside, it seems likely that any organisation engaged in research and analysis of any sort will be utilising such products over the coming several years. It is important to remember that these applications are the worst they will ever be and will only get better. It would be foolish for anyone working with information to not at least be seriously experimenting with such tools and exploring their potential.

So how will workers and organisations compete in this emerging age of autonomous agents, low-cost research products and AI-powered data analysis? This is the big question troubling employees, employers, politicians and policy makers. Predicting the future impacts of technology is an impossible task as unforeseeable innovations and seemingly random economic reactions will throw off even the most reasoned forecasts. However, we can learn some lessons from the past.

A key lesson is that while some existing occupations and roles stand to be significantly changed or even made obsolete by AI, new ones will emerge. The demise of many administrative roles was predicted in the 1980s and 1990s from the rise of personal computers. However, according to US labour statistics, while admin jobs peaked in 1980, they then only fell back to the level of the pre-PC 1950s.Footnote ¹² Similarly, the rapid adoption of desktop publishing software in the 1990s was seen by many as a threat to graphic designers but their numbers continued to rise. Clearly the generative nature of current AI systems is a step change from more traditional PC applications and the need for human inputs in many creative and business processes will decline as a result. Perhaps the most immediate threat is to routine administrative tasks. The Institute for Public Policy Research estimated in a 2024 assessment that of 22,000 tasks currently performed within the UK economy, 11 % were immediately exposed to AI but that figure could rise to nearly 60% if AI systems were more deeply integrated within organisational processes.Footnote ¹³ Based on the current ways AI is being deployed, this does not seem unrealistic.

Successful businesses and workers will be those that build new skills and utilise the technologies to create new and better outputs than was previously possible. This will be particularly true for those managing information resources. In a world where every worker and business has access to the same low-cost tools, as is increasingly the case, success will go to those best able to manipulate those tools while having access to unique data which can be fed into them. Access to and control of proprietary data will become a key differentiator of success and organisations should be focusing on identifying what data assets they currently have and what they might need in the future. Data cooperatives may become more significant as businesses pool their assets in closed communities for mutual benefit.

Hoping that all this AI hype will blow over and we will all return to business as usual is not an option. We should bear in mind Hemingway’s character from the novel The Sun Also Rises. When asked how he went bankrupt he said: “Two ways: gradually then suddenly”.Footnote ¹⁴ The momentum behind AI investments and innovations and the rapidly increasing number of applications being developed to solve real business problems will only accelerate. For anyone working with information these are challenging but exciting times.