AI has been around for many, many years, but it took the world by storm about two years ago when ChatGPT arrived. Since then, it’s just been evolving at a faster and faster pace. The real hot topic I hear now is the concept of AI agents. I’m hopeful, Dan, that you’re going to explain what all the fuss is about.

Why don’t we just start with the basics? The AI landscape as a whole—where are we, and how do AI agents fit into that? What are they, and why is everyone making such a big deal about them?

Agents, in the world of AI, are relatively new. We’ve had ChatGPT and Copilot for a couple of years now, but the agent side of things has only really entered the AI vocabulary for the past six months or so.

Understanding agents

The way I like to think about agents is: if you take ChatGPT and Copilot, which a lot of people are familiar with, they fit under the generalist bracket of AI. They use the knowledge of the internet or maybe your work data, and that’s how they are responding to your questions and prompts.

An agent is an offshoot of that, where it's given very specific instructions on how to behave, and also very specific information that it then uses in order to perform a task. So, in its most simplistic form, an agent is an offshoot of generalist AI.

Why do you think they're called agents? Where has that term come from?

Well, you alluded to a detective. Maybe it’s something to do with that. If you think a detective has to find a path or a line to solve a case or a crime, an agent is following a path to get to an answer or to perform a task that a generalist AI isn't.

Agentic AI and autonomy

That makes sense. There’s another term I hear banded around now in the same sentence as agents, and that's agentic AI. What does that mean, and are agents agentic?

"Agentic" is a very, very new term. I don’t think everyone’s truly coalesced on the actual meaning. I’ve seen different interpretations of what agentic is.

If you say an agent is doing a very specific task with very specific knowledge, the way you interact with that agent is important. You, as a user, are interacting with an agent—you're asking a question. You can think of an agent as something you deal with in regular life. If you have a problem with your energy bill, you go onto your power supplier’s website, and the chatbot that pops up is an agent.

Where we go into the agentic realm, though, is when we’re taking the person, the human, out of it. The human isn't interacting with it. It might be an event that is then launching this agent. So they are more autonomous and not necessarily being used by a person.

Exactly. And they might also be communicating with each other—an agent communicating with another agent to try and get to some form of answer or to perform some task.

Agents versus chatbots

Going back, you mentioned sort of virtual assistants and chatbots, which we are all familiar with when trying to speak to customer service at your bank or utility provider. What is the difference between what we’ve gotten used to there and what an AI agent would do?

Yes, those have been around for a while now. Those agents, however, are in the category of having been constructed in a way where every input is mapped to every output. A person has been responsible for figuring out, "If somebody puts this particular message, then do this or perform this."

What AI has done is stripped that out because AI is really good at understanding what a person actually wants—what the context is—and then being able to go off and do things as a result of that, maybe looping back into an agent. We’ve taken away the need, thanks to the AI, to understand the very specifics of what it is that somebody needs.

Agents in a business context

That takes us quite nicely then into a business context. What can these agents really do for businesses right now? What is the point of them?

For a very basic agent, if your company is used to using ChatGPT or Copilot in a generic way, an agent can be pointed at specific information in your business and provided with specific instructions. You can think of it as a specialist in your business. It’s got specialist knowledge that you’ve provided to it.

Then we’re starting to move out of the realms of a generalist, generic AI, and now we’re building things that are specific to your business—an AI that you’ve built that does something unique or special to your business.

If you look across a business, say, business support functions like HR or Finance, what kind of tasks within those departments lend themselves well to being done by an agent?

Anything heavily text-based, whether that's text generation or text summarising. That’s what AI is going to be the best at. But agent-wise, it's that autonomy. Taking something very textual, maybe a process that is text-based but is repeatable—something that somebody has to do every month, either writing text or summarising text or getting text-based information from other departments, and then performing some kind of repeatable activity—that is the sweet spot for it.

Common use cases in financial services

With our clients within financial services, what common use cases are you seeing at the moment for agents?

RFPs (Requests for Proposal) are a good one. Also, DDQs (Due Diligence Questionnaires). They have questions about the business when they get clients, and these are typically very repeatable—the same questions repeat over every request that comes in.

Typically, what we've seen is that clients store a block of canned answers to these questions. If you focus an agent on that knowledge and say, "This is your knowledge, and you are designed to respond to RFP questions," that’s been a really good match: something very text-based but very repeatable.

So you teach the agent to be the RFP response expert, and it builds up knowledge based on your existing information. It’s less teaching and more like you are recruiting somebody to do this task, and you have to construct a job description. It's all about the job description: what are you doing, what are you supposed to do, and how are you supposed to answer? It's that kind of job spec that really informs what the AI does and the knowledge it has access to.

Agents and the workforce

So it sounds very much like these agents are providing a supporting role for your teams. Are they replacing people, or are we starting to see what many have feared, where these agents mean you don’t necessarily need the same number of people?

This is the age-old question with AI: when is it going to start replacing people? I think we’re getting closer on the agent front, especially when it comes to this automation part, but it’s still, to use Microsoft’s word for it, it’s still in the Copilot seat. It is not the captain. It’s called that for a reason.

What I would say, though, is that it may be pushing back the milestone of when you may decide to recruit. If you're a small startup and you were thinking you would need to employ these people to start doing all of these tasks, you may actually now be delaying that because you can automate more easily with these agents and agentic automation. But I don’t think you're replacing people yet anyway.

Microsoft’s agent product

You mentioned Microsoft. Our financial services clients use a lot of Microsoft technology. Tell us a bit about Microsoft’s AI agent product and how it fits into the wider Microsoft product set.

The Microsoft world is all about Copilot; that's their main AI, and agents are an offshoot of that. An agent sits within the Microsoft 365 Copilot, so it is a licensable product. It’s tied closely to Copilot Studio, which is where you actually create these agents.

Agents in the Microsoft world are very easy to create, especially if you’re already in the Microsoft Copilot world. The agentic side of it—the automation side, where it's chatting with other agents and going off and doing things—sits within their automation product as well, which is a bit trickier to do. But very much, it is just the same kind of underlying tech as Copilot.

What do you need to get started? It sounds like it’s all there and quite easy to access and use.

The starting point is a license that gets you access to the studio. Then, really, it is about that job description I mentioned. You have to be able to write really good descriptive prompts about who your agent is and what it's supposed to do. That is a bit of an art form, and more is more in that world.

Also, you're going to want your data to be in Microsoft 365 so you can then point your agent to it, and the agent then uses that as its knowledge. Not everyone has moved all of their data into Microsoft 365, so that can be a challenge. But at its basics, it's just the license and the prompt for creating an agent.

Technical skill and risk

What level of technical skill do you need? Do you need to be in the IT department, or are we talking about regular business users being able to create these things?

You don’t need to be a programmer to create these agents. The interface is very simple. It's about that prompt, which is really important.

However, when we move into the agentic automation side of things—doing things—that's where we start to get out of the realms of a standard person in a business. We’re then talking around more development power users or the IT department.

It's very easy to create a very basic one that you can interact with. But when it actually gets more agentic and becomes more autonomous and can go off and do things, that’s when you need some of the more technical skills, which is arguably where the most value is.

This starts to make me feel a little bit nervous. We work in IT and know about Shadow IT, where users can buy or develop applications without the IT or security department knowing. So let’s talk about some of the risks here. How do Microsoft address the security and privacy concerns that come alongside this?

You get the same benefits and strong terms and conditions with Copilot Agent that you do with regular Copilot around privacy. They won't store any of your data, or at least they won't use it for future models. Copilot sits within Microsoft’s "trust boundary," the same place where you’re storing your emails. That's also where Copilot Agent sits.

You are right, though. There is something really key: the Copilot system honours existing permissions. It will only respond based on things that I have access to, that people have given me access to.

Copilot Agent works slightly differently in that it can surface under some circumstances information that you don’t necessarily want people to have access to because it’s using the identity of the person who created the Copilot Agent.

Let me get this straight. Let’s say Bob creates an agent. He’s in the HR department and has access to sensitive HR information. He points the agent at that information. Then, Sharon gets access to that agent, but she shouldn’t have access to the knowledge and information that the agent has access to through Bob. That’s a real risk.

It can be. It's slightly more nuanced, but there is the potential for that to happen. So, it's not that you need to not use it because that risk has a high likelihood. It's just something to be aware of that you don't necessarily have to be aware of from the regular Copilot. There's just a little bit more attention that needs to be paid to what the agent has access to, and whether the person who's going to be interacting with that should have it. The creator of the agent—the agent inherits their permissions for certain data sources—that’s where the nuance comes.

The future of software

Let’s move on and talk a little bit about the future. What does this mean for software vendors more broadly? We all use lots of third-party software in our business. Are these software vendors now creating agents to go sit alongside their software? What’s happening to software vendors, and what does this mean for the future of software?

There are two things I see happening. One is that all the software vendors are building agents. Every bit of software is going to have AI plus a bot or an agent to interact with it. We have use cases at Lanware as an IT company where we see that we want agents for, so it's going to be prevalent across all software.

The other track is very interesting: the introduction of MCP (Model Context Protocol), which is now an open-source way for software vendors to make their software available for agents to interact with.

It's like a common standard, a bit like open banking, so you can open up to allow it.

Very much so. Lots of software vendors are now opening up their software so that instead of the human interface going and doing things, an AI agent will be able to go in and manipulate or perform tasks. That’s very interesting.

So, is this more for those who create their own agents—maybe in Copilot—so they can hook them into or talk to some of these third-party software vendors and their products? The agent can not just go off to information sources to build up knowledge and access large language models; it can actually talk to these third-party software applications, right?

And all using natural language, right? At this point, we’re talking about not even building the automation—you just describe the automation, and it’ll go off and do it. That’s everything talking to everything.

Microsoft's next steps

So, Microsoft: what’s coming next? Where are their Copilot agents going to go?

We're already seeing from Microsoft that they are building MCPs for all of their software, so they are on that track, which makes it a lot easier from the Copilot perspective to start building those automations.

Microsoft has released "Actions," which is their first iteration of this. It is different to agents; it's different to the agentic. But they are trying to come up with their own way so that you don't have to build an agent or the agentic automation side of things—they’re going to provide it to you. However, I think that’s going to be superseded by everything else I’ve just mentioned around MCP. That is the direction of travel.

Is the hype deserved?

Do you think the hype around these agents is deserved?

I think it is deserved. I’ve said for a long time that AI plus automation is where it gets really exciting and really interesting. That’s where we are. This is a significant milestone where AI is meeting automation, and this is a little bit what we’ve been waiting for: where AI, the intelligence of it, can now go and actually do things. The models will keep getting better, they’ll be able to provide better answers, they’ll be able to summarise better, but mixing that with it, then being able to go and do something—that’s what we’ve been looking for.