If you are like me, your use of AI is limited to asking a question on a chat box, potentially refining it a few times and then getting an answer. If you end up going to an online support AI agent, the last action you do is typically click on the “talk to a real human agent” button. In summary, they don’t do much, other than answer your questions, which you will copy and paste or that you will follow the instructions it provided for you to accomplish the task.
For example, if you are having issues with the laptop, you may reach out to ChatGPT to guide you to fix the issue, or if you have questions about your tax filling, it will let you know what boxes to fill out and with what. In business settings, LLMs are learning how your company works—its processes, tools, and systems—so they can help you in your work. As such the answers will be procedures such as how to properly file a claim, approve an invoice, open up a work order, and so on.
Autonomous AI Agents go beyond – they will actually do the task whenever possible. It’ll handle whatever multi-step process is needed: filling out forms, launching apps, submitting files—whatever it takes to get it done for you (with or without supervision).
For me, I find this amazingly useful when, for example, going through my emails and we just ask the agent to write a reply. It is not as simple as just generating some response based on the one email but read the threads, understand my position from previous replies potentially, and compose an answer and then reply. Additionally, I enable this agent to do that going forward.
Admittedly, this is a simple example but it could potentially save me hours of work at the end of the month. It could also cause me a lot of issues depending on how well it did its job.
Here are some more examples of what they could do:
- Support: Chatbots with natural language processing (NLP) provide instant responses while launching AI agents that integrate with the CRM/Finance systems to retrieve customer history, etc. They will provide the answers as the agents retrieve the information and only escalate to human agents only when necessary, reducing response times and costs.
- Finances: An agent is tasked with analyzing financial transactions for anomalies (fraud prevention) and decline transactions, payments and send alerts/emails as needed. Or agents that are tasked with monitoring an email inbox for emails containing invoices and process them by OCR, identifying the required elements, input them into accounting, and matching them with the bank transactions.
- Logistics: AI monitors and forecasts inventory needs based on past data and current demand fluctuations and launches autonomous agents to negotiate with suppliers and manage orders.
- Sales: Autonomous agents are tasked to handle lead qualification, emailing, and follow-ups and optimizes frequency, email content and personalization as it learns.
- Citizen Support: Autonomous agents tasked with processing drivers’ licenses, social benefits, and tax filings.
- Public Safety: AI analyzes real-time surveillance feeds to detect suspicious activities and triggers alerts with police.
- Healthcare: AI agents assist emergency dispatch centers by prioritizing 911 calls or provide mental health support and medical advice on the call.
Isn’t this just (software) automation?
Reading the above use cases, we may think at first glance that this is just some level of automation but these are based on general needs. The idea of an AI Agent is not to answer a question but to DO A TASK.
In all cases, we are not asking the agent to answer ONE question but rather do ONE task be it provide support, call for help, file a document, send an email, optimize delivery routes, renew an official document, provide medical advice, eliminate humans due to being inefficient and self-destructive. Okay, maybe I went too far but everything else is valid.
What is more interesting is that these Agents that will do some tasks may be remote agents hired or paid to do that task. In essence, there could be the concept of a marketplace of agents that can do certain tasks and be hired to perform a task by an orchestrator.
Software apps can replace their API with agents that perform a specific task within their software. The agents can be called upon by third parties (with proper security one would hope) and perform the task on their account.
If you were concerned about the rise of AI in taking over jobs well then this Autonomous Agents will add a few more concerns to your list. But not only jobs, Microsoft CEO Satya Nadella has highlighted that AI agents can reshape software applications and business processes. He envisions a future where AI agents automate various backend business functions, leading to a new tier of multi-agent orchestration.
These agents will further operate across multiple apps or databases, independent of the back-end systems, this is what we might call the ‘AI tier’, said Nadella. Once this takes place, companies may start replacing back-ends with AI agents.
He continues to say that he plans to collapse back-ends and make SaaS applications less important such as his own application – Microsoft’s Dynamics.
In my opinion, Microsoft Dynamics does not need additional help from the leadership or AI to collapse, its doing a great job on its own mostly because it is a 30 year old mix of acquisition products that never really got the focus it needed to become a true SaaS player. But that is just my opinion and not really relevant to this article 😉
Reactive versus Cognitive Agents
Up until now, we’ve been talking about agents that simply react—they’re given a specific task within a defined context, and they perform it based on what they were trained on.
Cognitive agents take things a step further. Instead of just reacting to one isolated request, they remember what they’ve done before and learn from it. They use that history to improve how they handle future tasks.
Take my earlier email example: a basic agent could just respond to a single email by generating a reply using GPT. But a cognitive agent does more. It can look at how I’ve replied to similar emails in the past, consider its own previous responses, and use all of that context to write better replies going forward. So if I get an email from the same person again, it might start adding phrases like “As we discussed last week” or “Following up on our previous conversation.”
These kinds of agents already exist in real-world applications—think self-driving cars that learn from previous trips, healthcare systems that adapt to patient history, or fraud detection tools that refine their models over time.
Another example of where GPT is reaching outside of its own application is already live in CoPilot and Chat GPT.
In the example above, ChatGPT—after I gave it permission to access my operating system—recognized which apps I had open. It noticed I was writing this blog in Notion and suggested working directly with it. Now, it pulls in the content I have open in Notion to better understand the context and improve its responses.
⚠️ At this point ChatGPT does not have access to all my files or even Notion pages but that is only due to licensing issues not capabilities. Notion (and other apps) has its own economic model with its own AI so naturally it is not in their interest to let any GPT into their vault of knowledge.
Ultimately these same AI Agents can be invoked not only from software but from IoT devices such as your home thermostat (temperature too hot → open windows ), someone at the door (answer and provide instructions or simply open the door), and so on.
Impact on Humans…
There will be an impact on the workforce for sure. Our older generations just got used to the idea of Facebook and Google and now they have this thing to deal with. I will be looking forward to the hours of conversation my Mom will have on the phone with an AI agent. But more seriously, we are already seeing questions on the impact of AI on Data Analysts , Developers. There is an impact on many many roles from sales teams to educators. The list does not stop there and runs already in the near million from 2024 to 2025, all claiming an AI revolution or transformation.
There is also an important aspect which is where you can (sometimes) reason with human beings, it will be very difficult to reason with AI. Potentially the most well know case is when United Healthcare CEO Brian Thomson was murdered by a person that most likely was denied an insurance health claim. United Healthcare was using an algorithm known as nH Predict, allegedly had a 90 percent error rate — and according to the families of the two deceased men who filed the suit, the company knew it.
Finally, we get to the question of bias in the training data based on an overload of potentially biased data already available here and here. This presents a high risk for all of us. It is also the reason why today you are unable to get ChatGPT/DALL-E to show you a clock at any time other than 10:10.
Final thoughts
If you were not scared previously by the advent of GPT, maybe now its the time to be scared. As more and more companies adopting AI, many times incorrectly and without any significant long term testing and others making AI agents running in unknown locations processing your data and now also making decisions and performing actions.
AI Agents now transpose the confines of a chat bot that makes Google look like Yahoo, that have the potential to save you time performing repetitive tasks that require contextual intelligence and can’t be automated using a rules based system.
With that, I leave you with this television camera AI agent story.
