Rolls Royce has progressed from selling aircraft engines to providing airlines with engine-hours, charging them for flight hours rather than selling them the physical engines.
Hewlett Packard has shifted from selling printers to charging customers on the basis of number of pages printed.
Dr Don Perugini, an expert on Artificial Intelligence and cognitive AI agents, and founder of AI startup companies Presagen, ISD Analytics, and Socontra, tells GFiles Associate Editor Ravi Visvesvaraya Sharada Prasad ( an alumnus of Carnegie Mellon and IIT Kanpur, who was ranked by Onalytica UK as the 19th most influential thought leader in the world in Telecommunications, and by the IoT community as the second most influential thought leader in the world in IoT Internet of Things ) that India can quickly become a world leader in the Cognitive Reasoning Revolution, and pioneer Agentic Commerce and Agentic Webs.
RVSP: You are one of the world’s leading experts in Agentic Commerce and Artificial Intelligence Factories.
Last month, Sequoia Capital projected AI Factories as a ten trillion-dollar market.
Its August 2025 Agentic Commerce report spoke about business and consumer trends such as Consultative Purchasing, Predictive Shipping, and Deliveries by Autonomous Vehicles.
The Sequoia Capital report also mentioned technology breakthroughs such as Model Context Protocol Servers.
What are the opportunities for Indian high technology companies in these fields?
Dr Don Perugini (DP): There is a big opportunity right now for India to invest in new cognitive AI techniques to accelerate the Cognitive Revolution, where AI has better reasoning capabilities to automate more complex human-centric tasks, as well as to become a world leader in AI Factories and the ten trillion-dollar AI-enabled services industry.
China is an example of how India can compete successfully against USA.
Earlier this year, the USA had the lead in LLMs (Large Language Models ), which needed significant funding to create.
Well-funded US companies sat comfortably spending billions of dollars on LLMs – not innovating and doing the same expensive things over and over again.
Then China’s Deepseek, through various innovations, disrupted this model and showed that LLMs could be produced for a fraction of the cost.
RVSP: For the benefit of our readers, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Company Limited was founded in July 2023 by the Chinese hedge fund High-Flyer. In January 2025, it launched its DeepSeek chatbot, and in March 2025 it launched DeepSeek-V3-0324. In May 2025, it launched DeepSeek-R1-0528, and in August 2025 it launched DeepSeek V3.1. These were developed at a cost of only six million dollars, in contrast to the several hundred million dollars spent by OpenAI to develop GPT-4. The share prices of US AI vendors NVidia, Microsoft, Meta Platforms, Oracle, Broadcom fell sharply within a few days of DeepSeek’s chatbot being launched. USA’s OpenAI is proprietary technology, whereas China’s DeepSeek is open source and free. India’s government has banned the use of DeepSeek for governmental uses because all user data is stored in China.
DP: I foresee US companies making the same errors as we enter the Cognitive Revolution. Even though there is recognition that AI Agents need more cognitive reasoning abilities, there is very little indication that US companies will invest in new ( and cheaper ) innovative techniques based on cognitive science, as suggested by Gary Marcus, myself and others.
I am apprehensive that US companies will erroneously focus on more investment in the same expensive mainstream AI technologies – Machine Learning, Reinforcement Learning, Large Language Models – which are limited in their cognitive reasoning abilities.
US companies will thereby miss opportunities to invest more broadly in more powerful Cognitive AI techniques which will help us achieve the Cognitive Revolution faster.
Like China’s DeepSeek, India can lead by doing something different and innovative to lead the Cognitive Revolution.
Another area where India can lead the world is in Agentic Commerce and Agentic Web, to enable fully automated online shopping.
RVSP; With reference to your above recommendation – a very important and significant recommendation- that Indian companies should immediately enter the emerging field of Agentic Commerce, according to the August 2025 Sequoia Capital report on how to build the next two trillion dollar retail giant, customers have just begun using AI chat platforms such as ChatGPT, Google Gemini, Claude, Grok, Perplexity, to obtain information about products and services.
However, at present, the consumers will still need to visit the website or app of the vendors to place their purchase orders.
The big business opportunity, according to Sequoia, is if customers can place their purchase orders and make their payments directly from within the AI chat itself, instead of having to separately log in to the app or website of the vendor.
According to Sequoia, the big technology opportunity is in Model Context Protocol Servers, which facilitates the AI chatbots’ interactions with external tools and services.
Vendors can also utilise these MCP servers to insert the purchase transactions of their products and services into the search recommendations of the AI Chat bots.
DP: Current solutions for Agentic Commerce are centralized and controlled by large players such as Amazon or Shopify, and are still human-driven via chat-bots.
If an online business wants to be part of Agentic Commerce, they need to join one of these large centralized aggregators, creating a range of risks and barriers for smaller businesses.
The future of Agentic Commerce will be more decentralized – like the internet is today – where any online business can set up a website and have consumers shop directly with stores.
The same will apply to Agentic Commerce – any online business / merchant can set up a Web Agent designed to commercially transact direct with consumer AI agents to completely automate the online shopping experience.
Protocols and infrastructure like Socontra.com, which is like the ‘internet for agents’ to collaborate and commercially transact, can be used by startups and developers to create apps for buyers (e.g. consumers) and sellers (businesses, merchants) to enable fully automated online shopping.
In light of this, I see the web development industry changing in the future. Startup Lovable became popular by allowing users to create websites instantly using a few prompts about what they want their website to look like.
Similarly, in the future we could see a new category of startups with apps that create Web Agents instantly, independently or as an AI agent replicate of their existing website. This will allow anyone, from large companies to small local stores, the kirana dukaans, to have a web presence to service consumers directly via their AI agent store.
RVSP: Before we move on to discussing AI Factories, a quick discussion on Smart Factories and Industry 4.0.
China, which traditionally was viewed as a source of cheap factory labour, has invested heavily in robotics, automation, and artificial intelligence. China is now third in the world in automated factories, behind only South Korea and Singapore, having leapfrogged ahead of Switzerland and Germany.
DP : This is another example of where China invested in technology to become a world leader in manufacturing, i.e. traditional factories.
Similarly, India can invest in the Cognitive Revolution and AI Factories to become a world-leader in the services industry.
Smart Factories and Industry 4.0 typically refer to factories that produce physical products – Ravi, you are included in Onalytica’s “Who’s Who in Smart Factory and Industry 4.0”.
AI Factory inputs and outputs are data and knowledge to produce or support services.
RVSP: About AI Factories using Data and Knowledge to produce or support services, there is the emerging business model of Servitization, where manufacturers of industrial products are now integrating maintenance, upgrades, training, and other services to offer complete customer solutions, with performance based outcomes.
Rolls Royce has progressed from selling aircraft engines to providing airlines with engine-hours, charging them for flight hours rather than selling them the physical engines.
Hewlett Packard has shifted from selling printers to charging customers on the basis of number of pages printed.
You have defined AI Factories as: “An AI Factory is about using Cognitive AI as a core component to create more effective and efficient services businesses in different industries, like law or accounting…Agent Communication Protocols facilitate the assembly line – used to coordinate or orchestrate the many specialist Cognitive AI agents in the production process or workflow…”.
For the benefit of our readers, could you elaborate on what ‘AI Factories’ are? You have spoken about global supply chains and AI Factories.
DP: A traditional factory comprises an assembly line with specialist machinery. Inputs into the production process are typically physical (raw) materials. You start at one end and progress through the assembly line, from one specialist machine to another, to produce the final physical product at the end, such as a car. Some of the machinery may be operated (or augmented) by a factory worker (laborer). Other machinery performs the task on its own using robotics (automation).
Similarly, an AI Factory comprises a digital assembly line of specialist AI software modules (agents) that perform specific tasks. Inputs into the production process are data and knowledge.
You start at one end and progress through the assembly line (or workflow), from one specialist AI agent to the next, to produce the final product at the other end. The product is not physical, but rather its data and knowledge (or newly trained AI models) to support knowledge services, such as legal or accounting services.
Some of the AI Agents may be augmented by a knowledge worker, and other AI agents may perform the task on their own (fully automated).
An AI factory is underpinned by hardware infrastructure, namely the data center and networked GPUs ( Graphical Processing Units ) to train and run AI agents.
This AI factory concept is expected to create a new category of services startup companies that Sequoia Capital calculates as a $10 Trillion market. This is where AI startups transition from tech companies that sell AI products to services companies, to becoming full-fledged services companies themselves. Future AI startups will provide law, accounting or other services, and use AI – the AI Factory – as a core component of how services are produced and delivered.
As with traditional factories, global supply chains are important to source raw inputs, as well as outsource parts of the production process that other factories excel at. For example, components may be manufactured by factories in China and Germany and then assembled by factories in the USA.
The same applies to AI Factories. There may be third-party AI Factories or companies that have the necessary input data or specialist AI capability (e.g. analyzing financial data) that become part of the global supply chain into the AI Factory’s production process.
Agent Communication Protocols are critical to this process. They provide connectivity between AI agents and external tools, apps and data (email, calendars, weather services, databases, etc.); and for agent-to-agent collaboration and transactions both within the AI Factory (assembly line) and the global supply chain.
Just like how people and businesses interact to collaborate or transact to buy and sell goods and services, Agent Communication Protocols specify how AI agents should communicate and interact with other AI agents in an automated way in order to collaborate or transact to buy and sell goods and services (the HTTP for AI agents).
Popular protocols exist to support AI agents to work together within an AI Factory assembly line: Anthropic MCP and Google A2A. MCP allows AI agents to access tools and data sources, and Google A2A allows AI Agents to delegate tasks to each other.
However, when you consider AI agents collaborating between different AI Factories (or companies) in a global supply chain on the ‘open internet’, MCP and A2A protocols are insufficient. An agent in one AI Factory cannot simply delegate a task to an agent in another AI Factory. Like in the real world, getting a third-party to perform a task (including providing goods and services) is transactional (can’t be told, you must ask), and may involve payment.
Automated commercial transactions between AI agents over the open internet falls under the field of Agentic Commerce.
An example of an Agent Communication Protocol designed for Agentic Commerce on the open internet is Socontra.com. Socontra’s protocol is consistent with contract law to ensure legal compliance of automated agreements. Like the App Store, the protocol has human authorization for purchases built in to ensure user oversight and accountability to avoid disputes.
Automated global supply chains between AI Factories are not the only application for Agentic Commerce protocols. In the future, AI Agents will become ubiquitous, and every person and business will have an AI agent acting on their behalf. These agentic commerce protocols will eventually be used by everyone to facilitate fully automated online shopping – just like HTTP is used today to access web pages and online tools.
For example, my personal AI agent may transact with an AI agent representing the local Thai food store down the street to purchase and deliver lunch. My AI agent will know what my food preferences are, and will take care of the lunch order selection, purchase, and track the order through to delivery.
Although the concept of AI Factory is appealing, the biggest challenge is that current machine learning-based AI techniques, i.e. Large Language Models (LLMs) and Reinforcement Learning (RL), are limited in their ability to reason. Thus, these AI techniques are limited in automating more complex human-centric tasks.
Even a task as simple as ordering lunch may involve planning (how long will it take prepare the meal, what time should the delivery driver pick up the food, will the food arrive in time for the start of my lunchtime, etc.) and reasoning about different choices and constraints (if I choose the curry then I need to also order rice).
A recent article in the New York Times by world renowned scientist Gary Marcus discusses the limitations of LLM for reasoning, and the need for more cognitive AI techniques based on cognitive science and psychology. Gary suggests a technique known as neuro-symbolic AI as an option.
I have personally spent almost two decades using alternative cognitive AI techniques based on volition and beliefs. There are a few promising and emerging companies with their own approach to cognitive AI.
Sequoia Capital have rightly suggested that we are moving into a ‘Cognitive Revolution’. In the next 12-18 months there will be more investment and technology focused on enabling AI agents to reason and perform more complex human-centric tasks.
I expect these cognitive AI techniques to become more mainstream and complement current AI/LLMs/RL to create more powerful AI agents needed to build AI Factories. This cognitive AI technology will enable this new category of AI-enabled services businesses, like accounting and law, to become a reality.
RVSP) India’s legal and tax and financial systems will also have to evolve and adapt to accommodate such AI Factories, and Agentic Commerce, which would involve international payments made by Artificial Intelligence Agents.
What would be the impact on India’s blue-collar and white-collar employment?
DP: The concept of AI Factories impacts white-collar employment, ie knowledge workers. It is unlikely that AI will automate all the tasks performed by humans, and therefore you would expect the nature of work to change where white-collar workers using AI will be able to achieve significantly more, and be focused on more complex tasks.
RVSP: Regarding your statement that white collar workers would use AI to be more productive. I read a news report recently that METR Model Evaluation and Threat Research found that experienced developers who were using AI coding tools, such as Cursor Pro and Claude, actually took 20 % longer compared to when they were not using such AI coding tools. Paradoxically, these developers perceived themselves as being 20% faster.
The loss of productivity was because it took the developers much longer to test the code, which had been generated by AI, for errors, than it took them to write fresh code by themselves.
Further, it took the developers considerable time and effort to learn how to create appropriate Prompts.
To move on to my next question on Internet of Things evolving into Internet of Agents.
RVSP: What are your predictions about IoT and IIoT ( Internet of Things, Industrial Internet of Things ) evolving into the Internet of Agents and Agentic Webs, spurring Agentic Commerce, and in fact, Agentic Economies and Ecosystems?
DP: Correct, I believe that IoT, IIoT, websites etc will be enhanced by AI Agents (some at the Edge) to change the internet as we know it today – which will comprise predominantly of AI Agents that interact (intelligent web), and on the other side of the AI agents will be the people, businesses, Web Agents (i.e. websites for AI agents), IoT/IIoT devices, data, etc.
All these agents will interact in intelligent ways to buy, sell, answer queries, power our homes and businesses and factories, etc.
This is why infrastructure and protocols such as MCP, A2A and Socontra.com is important to facilitate the ‘internet of agents’ for startups and developers to start to build these AI agents to create this new world.
Eventually AI (agents) will become ubiquitous, and every person and business will have an AI agent acting on their behalf (agents will also represent data, tools, apps, and ‘websites’).
Agents will perform tasks on our behalf. Importantly agents will interact and collaborate with other agents (via ‘internet of agents’, agentic web, agent protocols like Socontra.com) and help us with social tasks like helping us connect with others (friends, family, co-workers) and automate commercial transactions.
In this scenario, agents are for everyone, help bring people together, and allow us to spend more time with each other and on productive tasks rather than scheduling, managing and shopping etc. Daily life in cities such as Mumbai and Bengaluru is very inefficient because of overstretched infrastructure.
Hence this is a positive view of AI (no one will lose their jobs) versus the typical negative narrative of AI in reducing labor and jobs. It’s also more practical because we all know that getting AI to fully replace a human worker is massively challenging, but to help us with these social and mundane individual tasks is more realistic.
AI agents will become a unified interface between human and machine – allowing people to interact with computers, devices, apps etc in a more ‘human friendly’ way (text, voice, images, video, immersive environment / metaverse). GenAI / LLM is already effective in providing this interface (the challenge we have with LLMs is not necessarily the interface, it is the inability for cognitive reasoning to automate complex human-centric tasks once the AI interprets our intentions).
Web browsers and app interfaces will become obsolete, replaced by our unified AI interface – completely changing the way we interact with our devices, apps and the agentic web.
Websites (designed to service humans) will become obsolete – replaced by Web Agents designed to service our personal AI agents that will search and shop on our behalf.
E-commerce will no longer be manual and via websites, it will be replaced by fully automated agentic commerce via agent-to-agent transactions.
My prediction, which would be controversial, is that social tools like email, Whatsapp, Slack, Facebook, Twitter, LinkedIn will become obsolete. These will be replaced by our unified AI interface and social connections (personal connections, work connections, professional connections, public connections).
Why? We use these social tools because we (humans) need the interface. Our unified AI interface removes that need, and my AI agent can directly connect with the AI agent(s) representing all my social connections directly over the agentic web.
For example, I typically post about AI, AI agents and startups. When I do post, I publish the exact same post on both LinkedIn and X (Twitter). If my personal AI agent has a list of all my connections from LinkedIn and X, then I could post my article via my AI agent, which could directly send this post to all my social connections via their personal AI agent.
Regarding Artificial General Intelligence, at a high level, I’m not a fan of the concept of ‘AGI’, which is why I always put quotes around it to indicate ‘whatever that means’. People imply AGI means AI that will become super-intelligent, gain consciousness and destroy the world, which is negative, an unlikely scenario, and only instills fear and distracts from the benefits that AI could have on society and the economy.
I view ‘AGI’ as AI that can do many types of tasks well, including the ability to automate ‘more complex’ human-centric tasks, but not necessarily all human tasks.
I have two positions on how we get the next step-change in AI to get us closer to ‘AGI’:
1. Cognitive AI to support more complex reasoning: As discussed above, we need cognitive AI to supplement current ML-based AI techniques (LLM/RL) to enable reasoning.
2. Scaling AI Agents – the future agent economy: The current theory is that AI will get smarter by scaling – scaling data and compute. Even if AI could get smarter by ingesting more and more data, the amount of data on the internet that is publicly accessible, like websites, news, social media, etc., this is only a tiny fraction of the total data that exists in the world.
This decentralized data source can be accessed using decentralized AI Agents, and the AI agent becomes a local specialist based on the data or resources (goods and services) it has access to.
So my view is: ‘AGI’ is unlikely to be achieved from scaling data and GPUs to create a single centralized super-intelligent AI.
‘AGI’ will likely emerge from scaling AI agents – billions of decentralized specialized agents collaborating to solve any query or task on behalf of users.
This is how collective intelligence works in the real world. We don’t have one super-intelligent person that’s a deep specialist in everything, however many deep specialists collectively sent a man to the moon.
Regarding Generative AI, I am very optimistic about Generative AI (GenAI) and LLMs. As discussed above, I think it will become central to how we interface with computers, devices, apps and the internet, and will replace current user interfaces and browsers that we use today.
GenAI allows us to communicate with computers in a more ‘human-friendly’ way via text, voice, images, video. For example, it allows us to communicate with our device by voice, e.g. ‘what is the weather tomorrow’. Rather than return the results via a browser, GenAI could dynamically generate the output based on the search results, e.g. a mix of text and images containing a description of the weather, plus a video of the expected rain map over the 24-hour period.
GenAI could also be the driver of a consumer shift from phones to VR/AR glasses. Right now, we use phones as our primary device for apps and communication. We typically use our fingers to type in text or interact with apps, and phones have very small screens.
In the future we are likely to replace our phone for VR/AR glasses with sight and sound. This will give us an immersive and more natural environment to interact with the digital world, with much larger screen. With Augmented Reality capability, we will be able to do more things with these glasses than is possible with today’s phones.
Dr Don Perugini is an AI scientist and entrepreneur. He has a PhD from the University of Melbourne and spent 25 years in field of AI agents including agent protocols and cognitive reasoning. He was the founder of multiple AI agent startups (Presagen, ISD Analytics, Socontra) that operated across San Francisco Bay Area and Australia, with two exits via acquisitions. Don invented and led the development of novel patented AI algorithms, has numerous publications, is a recipient of many innovation awards, and was also a mentor for AI start-ups.
Ravi Visvesvaraya Sharada Prasad is an internationally renowned high tech consultant, with four post graduate degrees in different fields of engineering from Carnegie Mellon University and IIT Kanpur.
Since his family members have been prominent in public life for centuries, he has a deep knowledge of India’s political history.
He is Associate Editor of gfiles
Ravi Visvesvaraya Sharada Prasad is a computer scientist and author. He writes on technology and historical events in post-independent India. He is Associate Editor at gfiles.
