AI explained: GPTs, ChatGPT Operator, AI agents and Agentic AI

ChatGPT is what is called a large language model or LLM. This means that it takes human queries, runs them against an internally trained model and generates a reply. The internally trained model is important because it means that if the model does not contain the data, it cannot answer a query. This can result in the dreaded hallucinations where it just invents an incorrect answer.

To get over this limitation, several solutions have been implemented:

• Fine-tuning: You can fine tune the model and show it new things it did not know before.
• RAG: You can have an external database with data to look up answers, e.g. find how much reward points a customer has by looking up their account. For more information about RAG see my other blog post.
• Code integration: You can create code integrations that call other systems. Solutions like LangChain, LangGraph,… fall into this category. See my post on LangChain here.

However there are new ways to extend the functionality of LLMs. So let’s look at them.

ChatGPT and GPTs

On the left menu of ChatGPT, you can find “Explore GPTs”. Here you can even create your own GPT:

You can upload files with new information your GPT can use, e.g. research or industry specific data. Via actions you can fetch data from outside systems via APIs:

This last one is very similar to LangChain, allowing you to programme actions to get information to private systems. Above you can see a simple example of getting the weather. So by adding external APIs, you can provide ChatGPT with additional knowledge and functionality.

ChatGPT Operator

Operator is a new research product that is currently only available in the USA. Users can together with ChatGPT browse the Internet and let ChatGPT do most of the work. Below you can see how the user asked Operator to find the highest rated one-day tour in Rome on Tripadvisor and book it.

Operator uses a new technology called Computer-Using Agent or CUA. CUA allows Operator to see a website and interact with it like a human. So in the above example, Operator will look on Tripadvisor through the search field for the best one day tour in Rome and use the website to make a reservation. It is called Operator because the human is still in control and “supervising” what the AI does. This is a more advanced version of Robot Process Automation which a lot of larger enterprises use to automate repetitive tasks often by mimicking what users do on legacy systems.

AI Agents

AI Agents take things a step further and perform a task autonomously. So instead of supervising how Operator would go to Tripadvisor to find a one-day tour and book it, an AI Agent would use the Tripadvisor API and search for the highest rated one-day tour and book it through the API. So AI Agents are similar to GPTs in the sense that they can interact with APIs but often they can also learn and improve over time. So if they suggest three tours and most people always select the third tour which is less walking, the AI agent could learn and improve suggestions over time. This is because AI agents often have memory and reinforced learning capabilities.

Agentic AI

Agentic AI is when multiple AI Agents start working together. So instead of having a TripAdvisor Agent, let’s assume we have a Business Travel Agentic AI. What would be different is that the first AI Agent you interact with would ask you what you wanted to do and delegate tasks to other AI Agents. So if you were going to Rome for a business trip and you had one free day in which you could do a tour, you wouldn’t just ask the Agentic AI to find and book a tour but instead handle your complete travel needs. This could involve calling several AI agents:

• to get the latest company travel policies
• find the best flights and hotels
• make restaurant reservations
• book transport to and from the airport
• get your travel approved
• book a one-day tour
• …

Basically all the tasks a business travel agent would do.

Agentic and the enterprise

Now that you understand what Agentic AI does, let’s broaden the impact it can have for your business. Most websites are built for humans to interact with. Most business platforms are built for deep one time integration. Agentic AI enables business processes both towards employees but also customers and partners to be brought to a new level. Imagine if your company is a travel insurer and instead of offering a website or a complex API, it offers a solution AI Agents can integrate with. One such standard is the Model Concept Protocol or MCP. MCP allows LLMs to easily interact with an external system without knowing anything about the website or API. Extensions can be made and installed which allow the LLM to talk to more and more systems. Below is a screenshot from Goose, an AI Desktop which enables local, enterprise or remote LLMs to be integrated with lots of extensions.

Back to our travel insurer who now via MCP can offer other AI Agents the capability to add travel insurance to a trip. You can easily see that if most companies move away from using human travel agents, the world of Agentic Travel Insurance is likely to grow substantially. Not only travel but basically any tasks employees, partners and customers need to do repetitively benefits from having Agentic protocols. Think getting quotes, taking orders, fulfilling orders, emitting invoices, paying invoices, signing contracts like NDAs, onboarding new customers, partners, agents and employees,…

The #AgenticRevolution

Agentic AI is one of the most disruptive innovations that is coming to the business world. Digital transformation was about allowing humans to use a website or mobile app or machines to use an API to automate repetitive processes. Agentic AI is about rethinking how your next customer is not a human or a partner/customer backend system but instead an AI Agent. With AI Agents being the customer, a fully automated business is able to grow at unseen levels. Sam Altman, OpenAi’s CEO, is predicting One-Person Unicorns. What he means by that is that any person who is able to make an AI Agent business that lots of Agentic AI can use, e.g. think about a fully automated AI travel agent, could get to a $1B valuation without hiring employees and as such turn into a one person business that is valued a Unicorn.

Now most business will still require humans and agents to work together, however you don’t want your business to be late to the #AgenticRevolution. If your business could be fully automated and AI agents can continuously order from it, then revenue per employee could go to unseen levels.

If you are unsure about your AI strategy or want to see a quick prototype before making any further commitments to Agentic AI, why don’t we talk.