The Rise of AI Agents

You're tuning in because you need to get smart fast. You want the crucial information without the fluff, and you want to understand why it matters to you.

Exactly. Time's valuable.

Well, today we're plunging straight into a technology that's really on the cusp of changing everything— AI agents.

Yeah, think of these AI agents for you as sort of digital assistants, but with, well, serious skills.

Right. Not just answering basic questions.

Yeah, no, they're software designed to act independently, like on your behalf or within a bigger system, to really tackle intricate tasks, things like, you know, understanding and responding in natural human language.

Okay. And to bring everyone the inside track on this, because it's moving so fast, we've synthesized the latest insights from McKinsey. This was compiled just last month, March 2025.

Yeah, fresh research.

So consider this your kind of express lane to becoming truly knowledgeable about Well, a really pivotal development in AI.

And our goal in this deep dive, it's pretty clear. Cut through the hype, give you a solid understanding of what these AI agents actually are, and, you know, the different types that are starting to pop up, how they operate under the hood.

Right.

And the potential impact they could have on businesses like yours.

And the hurdles too, right? It's not all smooth sailing.

Absolutely. We'll touch on the challenges in adoption and how it might just reshape the whole tech landscape. We're really trying to pull out the most vital and hopefully thought-provoking info for you.

And believe us, this isn't just like a small step up. We're talking a fundamental shift from those basic chatbots we all know to systems that can actually plan actions, collaborate with other AIs, and even learn and improve as they go. That's kind of mind-blowing.

It really is a leap.

Okay, so let's get into the nuts and bolts then. What exactly are AI agents? Can we get a bit more detail?

Sure. So at their core, like we said, software components acting autonomously. But what's really interesting, especially for you listening, is how you can organize these individual agents into interconnected systems.

Okay, so not just one agent working alone.

Right. This lets you orchestrate really complex workflows within, say, your organization. You can get coordinated efforts from multiple agents working together. They can apply logical reasoning to tricky problems, even evaluate answers to questions you might ask the system.

So, when we think about early examples, the McKinsey info mentions things like customer service chatbots, right? And maybe those generative AI models that create text or images. Are those the starting point?

Precisely. You've probably used those already. But the real jump, the big leap forward, has come from these major advancements in natural language processing, especially within modern generative AI.

Okay.

This has given these agent systems a whole new level of capability. They don't just react anymore; they can proactively plan the best way to get something done.

Plan? Like, think ahead?

Yeah, map out steps. They can work together with other specialized agents, complete complex tasks from start to finish, and— this is key— learn from their successes and failures to get better over time.

That learning aspect sounds huge. The report actually quotes Lori Hemmelinen from McKinsey, highlighting how this human-machine collaboration is leading to great quality and great productivity. That really drives home the potential benefit for, well, everyone listening.

It really does. And a big enabler for this, something the research points out, is the improvement in memory structures, both short-term and long-term, within these agents.

Ah, memory. So they can remember past interactions.

Exactly. Better memory means much more personalized interactions, whether it's with your customers or your own teams. It just makes them better at the job they're given.

So it's really moving from thought to action, as the report puts it. And we're seeing big tech players like Google, Microsoft, OpenAI investing heavily here.

Huge investments.

Tools like Microsoft Copilot, Amazon Q, Google's Project Astra, they seem to be pushing this action-oriented approach.

Definitely leading the charge. And what's really worth considering for you listening is a prediction that these AI agents could become as common as everyday as mobile apps are now.

Wow, that puts it into perspective, like just part of our daily toolkit.

Potentially, yes. That's the scale we might be looking at.

Okay, so hopefully everyone's got a clearer picture now of what these agents are and how quickly they're evolving. Let's shift gears and look at the different kinds of AI agents emerging. The McKinsey info gives a pretty useful breakdown.

Yeah, it does. They categorize them in a few ways, like by capability, Their role, the skills they have, the outcomes they're meant for. It's a helpful framework, though, you know, these things are still evolving.

Sure.

The first type they identify is individual augmentation agents. You often hear these called copilot agents.

Right. Like the name suggests, they help individuals be more productive. Things like Microsoft 365 Copilot or ChatGPT probably fit here.

Exactly right. Their main job is to boost a person's capabilities. They can help draft content, write code, find information, basically act as these tailored smart assistants for your workflow.

Makes sense.

But an important point the report makes for you to consider is that their real impact kind of depends on the individual user.

Yeah.

You know, how motivated they are to actually use the tool and integrate it.

Yeah. If you don't use it, you don't get the benefit. Seems obvious, but important. Okay. What's the next category?

Next up is workflow automation platforms. These are really focused on automating specific tasks. Could be a single repetitive thing or more complex multi-step workflows that might currently involve several people or systems in your company.

So taking over processes?

Kind of. Think of them as AI-powered orchestrators. They manage and execute predefined processes. The report mentions things like Microsoft's Copilot Studio, and Salesforce is apparently working on something called Agentforce.

Interesting. So instead of a person manually pushing a process along, the agent handles it. What does success depend on here for businesses thinking about this?

Well, for you to really get value, successful implementation is obviously key, but also managing the organizational change that comes with it. That's crucial. And of course, managing the agents themselves. Integration with your existing systems needs to be pretty seamless too.

Right. It's not just plug and play. Okay. Number 3 is GenAI native agents for domain solutions. Sounds very targeted.

It is. These are purpose-built for specific business areas or functions. Like, think of AI-driven customer service systems that can handle a really wide range of issues, or maybe AI-enabled software development pipelines automating parts of coding.

Okay.

The big difference here for you is that these are designed with AI right at the core. It's not just layering AI onto an existing structure. It's rethinking that whole domain with AI central to it.

That feels like a much deeper integration, reimagining entire business areas. What fall is that?

The next level up is AI-native enterprises and operating models. Now we're talking about AI agents being woven throughout the entire enterprise fabric, not just one department or process.

The whole company.

Pretty much. It involves a fundamental end-to-end redesign. How your organization interacts with customers, your core processes, your internal structure, even your basic business model, all reimagined with AI as a foundation. The report draws a parallel. for you to the big digital transformations companies went through.

Wow. Okay. That's a massive undertaking, like completely rethinking how the business runs driven by AI.

It is substantial.

Yeah.

But the 5th category offers maybe a quicker route to value for some. AI virtual workers.

Virtual workers, like digital employees?

Almost. Yeah. Agents designed to function like employees or team members within your existing structure. This could be highly disruptive because it might let companies capture AI value faster, Potentially, without needing that full, top-to-bottom operating model overhaul right away, the AI kind of slots into existing roles.

Ah, I see. So, instead of changing the whole structure, you bring in AI to do specific jobs within the current setup. That does sound like it could yield results quicker for businesses. And the report says these categories aren't exclusive; companies might mix and match.

Exactly. McKinsey expects most organizations will probably use a mix, depending on their needs and strategy. You might give everyone Copilots, automate some specific workflows, and pilot a few virtual workers in customer service, for example. It's about finding the right blend for your situation.

That gives us a great map of the emerging landscape. Really helpful. Okay, now let's get into the mechanics. How do these agents actually work? What's happening under the hood that you need to understand?

Well, what's really impressive for you to grasp is their ability to tackle complex tasks, often ones that are kind of ambiguous across lots of different industries.

Ambiguous, meaning not perfectly defined.

Right. And a key part of how they work is their flexibility with tools. They can use tools meant for humans, like web browsers, email clients, and also tools meant for computers, like APIs or databases.

Okay, so they can interact with both worlds.

Exactly. This dual capability gives them a lot of flexibility to operate across different tech setups inside your organization and outside, often without needing you to completely rebuild those existing systems.

That's a big deal, avoiding massive system overhauls. The McKinsey info mentions a typical 4-step process for how an agent tackles a task. Can you walk us through that? Sure!

It generally starts with step 1: you, the user, give the agent system a task or a goal.

Okay, the instruction.

But here's the crucial bit for you: the agent doesn't just follow rigid steps. It works autonomously to plan and figure out the best strategy to actually achieve that task. That's step 2, the planning.

Planning. That sounds key. It's not just reacting; it's strategizing. What happens after the plan is made?

So, once the task is clear and there's a plan, step 3 is the system plans, allocates, and executes the work. This often means breaking the big task into smaller subtasks.

Right.

A central manager agent might then assign these smaller jobs to other, more specialized sub-agents.

Like a team of agents?

Pretty much. These specialized agents use their existing knowledge, their domain expertise. They coordinate with each other, and they pull in data— both internal company data and external info— to do their part.

So, a network of agents collaborating. And the next step involves getting better, refining things.

Exactly. Step 4 is, the system can iteratively improve its output. It might proactively ask you, the user, for more clarification or input to make sure the results are accurate and relevant.

Uh-huh.

And even after it gives you an initial output, it might ask for feedback to help it learn and do better next time.

That feedback loop seems critical for making sure they keep improving for you. And then the final step is just getting it done.

Yes. Finally, the agent takes the necessary actions to complete the task based on the plan it developed and refined. Could be sending an email, updating a database, generating a report, whatever the original request was.

The report also really stresses having good controls in place when you use these agents in a business.

Right.

What kinds of safeguards are we talking about? What should you be thinking about?

Yeah, controls are vital for responsible use. We already mentioned constructive feedback loops that helps them learn.

Right.

Agents can also be designed to try and self-correct errors, or if they hit something they can't handle, escalate it to a human.

Okay, asking for help when needed.

Exactly. Another control is collaboration between different types of agents. Like, you might have a critic agent whose job is to review the work of a creator agent to find flaws.

Interesting. Like an internal quality check?

Sort of, yeah. And organizations can even build agents specifically to automatically test and validate what other agents produce, just to ensure accuracy and reliability for you.

That multi-layered control sounds essential for building trust and making sure these things work dependably in your organization. Okay, let's clear something up. The relationship between these AI agents and large language models, LLMs— you hear LLM all the time. How do they fit together?

Good question. It's important for you to know that AI agents are designed to be flexible. They can use various types of AI models for different parts of a task. Now, when an agent needs to interact with a human using natural conversational language, that's where LLMs come in. Their strength is natural language processing, understanding, and generating human-like text that lets the agent talk to you effectively.

The report used a self-driving car analogy, which I thought was helpful. Can you explain that?

Sure. Think of a self-driving car as a system of AI agents working together, right?

Yeah.

The agent that figures out where you want to go, understanding your navigation request, that probably uses an LLM because it involves understanding language, place names.

Makes sense.

But the agent deciding if it's safe to make a left turn right now with oncoming traffic, That agent doesn't use an LLM. It uses a highly specialized model designed for processing real-time sensor data, physics, reaction times— a totally different kind of AI expertise.

Got it. Different tools for different jobs within the same system.

Exactly. So, different agents in a system use the best AI model for their specific job.

That clears it up nicely. It's not all LLMs, all the time. Okay, so given these capabilities, how might AI agents really impact business growth and efficiency for you, the listener. McKinsey had some pretty big numbers.

They do. They estimate long-term enterprise use of generative AI could unlock up to, well, a staggering $4.4 trillion in annual value globally. Wow.

$4.4 trillion.

Huge number. But, and this is the key point for you, organizations won't actually get that value, won't see the real growth in productivity gains unless they can implement AI quickly in ways that fundamentally transform how work gets done. Reimagining processes.

And the argument is that these AI agents can speed up that value capture compared to older automation tech you might already have.

Precisely. But the potential driven by these agents goes beyond just automating existing tasks faster. McKinsey also sees organizations using agents to help rethink and redesign core business processes and importantly, to modernize their IT infrastructure.

Modernize IT. How so?

Well, things like adopting easier programming languages, moving to modern tech frameworks, making systems more modular, maybe migrating more things to the cloud. Agents can potentially help accelerate all of that.

So adopting agents could actually push broader IT improvements in your company. The report also mentioned tech leaders using collaborating agents for complex stuff.

Yeah, tech leaders could use teams of specialized agents, each with different skills, to tackle really complex projects, iterating in real time, incorporating human expert feedback as they go. But the truly huge value for you likely comes from orchestrating agents to handle entire end-to-end processes, not just single tasks, but the whole flow, like, say, the entire software development lifecycle.

We're already seeing hints of this, right? Customer service bots are getting much more common, handling more complex queries than they used to.

Exactly. And McKinsey's own research found that companies using GenAI-powered customer service agents were already seeing, I think it was, a 14% increase in issues resolved per hour.

14%.

And a 9% reduction in average handling time. Real, measurable gains.

Those are significant numbers. And Jorge Amar from McKinsey apparently sees even more potential ahead— better satisfaction, even agents generating revenue.

That's right. He thinks these agents will become key in identifying chances to sell new services, addressing broader customer needs. This could lead to companies offering more diverse customer experience options, maybe even premium human interaction for those who want it.

Interesting. The report also detailed 3 ways these GenAI agents make automating complex stuff easier for you. Can you break those down?

Sure. First, they handle unpredictable situations better. Old automation often breaks if something unexpected happens, right?

Yeah, it follows rigid rules.

Exactly. But these agent systems, built on foundation models trained on massive, messy data, can adapt in real time to a much wider range of scenarios within a task. They're more flexible.

That adaptability sounds like a huge plus in today's world for businesses. What's the second way they simplify things?

Second, you can direct them using natural language. Because of the advanced NLP, users can define complex workflows much more easily than before.

So you don't need to be a coder?

Pretty much. It empowers more people in your organization to use AI tools for business outcomes without needing deep programming skills or complex IT projects. It kind of democratizes it.

Makes AI automation more accessible across the board.

Okay. And the third way?

Third, they work better with your existing tools and platforms because they're built on these foundation models. They can naturally communicate across a broader digital ecosystem. Getting different systems to talk to each other usually takes a lot of manual integration work. These agents powered by foundation models can reduce that friction significantly.

That interoperability could save organizations a ton of time and effort integrating AI with what they already have. The report also mentioned something fascinating called emergent AI.

Yeah, Aaron Bachemuth, a McKinsey partner, pointed this out. As these agents work together and interact with their environment, their combined capabilities can actually grow in unexpected ways. Grow?

How? They can start developing new behaviors, even new skills that weren't explicitly programmed in. The whole becomes greater than the sum of its parts. That's emergent AI.

Wow. So they're not just executing tasks, they're potentially evolving through collaboration. That's quite something. The report gave a real-world example. too. Lenovo, what are they doing?

Right, Lenovo, the tech company. They've been using AI agents in software engineering and customer support.

And seeing results?

Apparently so. Their CTO, Arthur Hu, reported productivity gains up to 15% in software engineering. And their COO, Linda Yao, mentioned double-digit improvements in call handling time in customer support.

So, real, tangible benefits already being measured. And their future vision sounds even bolder.

It does. Right now, they're mostly using agents as virtual assistants to humans. But Lindy Au's vision is for agents to become more like independent deputies, capable of handling complex tasks autonomously from start to finish.

Deputies? That's a big step beyond just assisting. Okay, the McKinsey info also gave some hypothetical examples for businesses. Can you share a couple to make it more concrete for people listening?

Definitely. One good one is loan underwriting. Think about assessing credit risk. It's complex, involves lots of data collaboration.

It's very time-consuming.

Right. Imagine a multi-agent system where specialized agents analyze different parts of the application, different risk factors. It could make the whole process much faster and more streamlined.

Yeah, you can easily see the efficiency gains there. Less manual work, faster decisions. What's another example?

Another one is code documentation and modernization. Lots of companies have old legacy software, right? It can be a security risk, slow down innovation.

A big headache for many IT departments.

Exactly. So you could have a legacy expert agent analyze the old code, generate documentation. Then maybe a QA expert agent reviews that documentation, refines it, ensures it meets standards. This could help tackle security risks and speed up modernization efforts.

That could be a huge help for companies stuck with technical debt. What about marketing? Was there an example there?

Yes, online marketing campaigns. Designing, testing, running these involves lots of steps, different skills. An AI agent system could potentially handle a lot of that. Develop campaign ideas based on marketing input, test them, refine them. Then maybe sub-agents handle specific tasks like writing copy, designing visuals, managing the strategy.

So, a more integrated, maybe faster way to run campaigns. Okay, it's clear the potential is massive, but the report was also realistic about the hurdles. What are the big challenges organizations face in adopting these agents? What should you be ready for?

Trust. Building trust is highlighted as a major hurdle. Nikolai von Bismarck from McKinsey notes that even Gen Z often still prefers talking to a real person for complex support issues.

That's interesting. Even the digital natives.

Yeah. He mentioned a bank that's building checks into its AI system to catch errors before the AI talks to a customer. Specifically to build that confidence.

Proactive error checking to build trust. Makes sense. Roger Roberts from McKinsey also weighed in on trust.

He did. He emphasized for you that the companies getting the most value will be those who actively build trust with customers, employees, stakeholders. People need to feel confident enough to delegate tasks to AI and trust the results.

And it has to be ethical, right?

Absolutely. Yeah. Ethical decisions rooted in the company's values and frankly, societal values too. Fairness, well-being.

Okay. Trust is huge. What other big challenges come up when scaling AI agents?

Change management. That's another big one, according to Jorge Amar.

Yeah.

He stressed that integrating agents isn't just about new software. Companies need to fundamentally rewire how functions work.

Rewire? That sounds disruptive.

It can be. It means adapting existing systems to work with AI, maybe adjusting operating models for smaller, iterative teams. And crucially, creating incentives and training so employees learn to use and trust these new tools. It's a big organizational shift.

It really touches everything. And then there's data security, always a concern.

Always. Erin Baucom underlined data protection as critical. Companies must have robust security operations, data controls in place. There are solutions emerging, but it's absolutely vital to get this right. for responsible adoption.

These are definitely significant challenges businesses need to plan for. Okay, let's look at the tech side again. How might AI agents change the underlying technology architecture in organizations? The report suggests a big shift.

Yes, a big shift away from the traditional application-focused model towards, potentially, a multi-agent model. Tech leaders might find themselves overseeing thousands of communicating agents.

Thousands! Wow. How would these agents actually be deployed? The report mentioned 3 possible models.

Right. First is superplatforms. Think next-gen business apps like CRM or collaboration tools with GenAI agents built right in. So, your CRM doesn't just store data; its built-in agent might analyze sales reports and talk to analytics tools automatically— stuff that needed custom programming before.

So, more intelligence baked into the platforms you already use. What's the second model?

Second is AI wrappers. These are tools that let your internal services securely talk to third-party AI services via APIs, but without exposing your proprietary data.

Ah, like a protective layer.

Exactly. The example was a bank using a wrapper around its internal credit risk AI. The wrapper could query an external LLM for risk factors using customer data, but without actually sending the sensitive customer data to the external LLM. Keeps it secure.

That sounds incredibly useful for leveraging external AI safely. And the third model?

Third is custom AI agents. These are agents developed internally, maybe by fine-tuning a pre-trained LLM with your own data or using retrieval augmented generation, RAG.

Gargar.

Yeah, RAG lets an AI model access and use your company's specific knowledge bases and data. So a customer service team could use RAG to feed an LLM with internal policies, past transcripts, customer data, etc., to create a custom agent that knows your business inside out.

So empowering teams to build their own tailored agents. Whichever model or mix of models a company uses, it sounds like IT architectures need to adapt significantly to support these agent-focused workflows and allow for constant improvement.

Absolutely. It's a move towards a more distributed intelligent network.

Okay, so for organizations listening wanting to navigate this, what are some practical steps they can take now to start implementing effectively? The McKinsey report had some advice.

Yes, 3 key things. First, closely review tech proposals with long timelines and large teams. Scrutinize them. Ask, where could GenAI reduce costs or timelines here? Be a bit skeptical if GenAI seems like just an add-on with limited value.

So, challenge existing plans through an AI lens. Good advice. What's the second step?

Second, focus on the biggest problems. Don't just tinker with small AI initiatives if you want big results. Identify your most expensive, complex, technically indebted challenges and aim generative AI squarely at solving those first. Go for high impact.

Go big where it matters most. Makes sense. And the final implementation step.

Finally, get ahead of the talent, technology, and operating model implications. Think now about the business impact of scaling multi-agent approaches. What's your talent strategy? How will you reskill people? How do operating models need to change? How do budgets need to shift? Plan proactively for these broader changes.

It's really a holistic view— tech, people, process, budget. The report finished on a pretty optimistic note about the future, didn't it?

It did. It stressed that this evolution is just beginning. There's so much more learning and innovation to come as this moves from labs to full-scale implementation everywhere. AI agents offer genuinely new ways of working that could transform organizations of all sizes. We're just scratching the surface.

Okay, so let's quickly recap this deep dive based on that March 2025 McKinsey information. We've looked at what AI agents are these software components acting for us. We've seen the different types emerging, from copilots helping individuals to potentially AI virtual workers.

Uh-huh. And how they work that process of planning, allocating tasks, executing, and iterating with feedback.

Right. We explored the huge potential impact on business growth, the significant efficiency gains possible, but also the critical hurdles like building trust and managing the necessary organizational change.

And how this is likely to reshape tech architecture moving towards these multi-agent systems.

Now, we touched on those concrete steps you can take now— reviewing proposals, focusing on big problems, and planning for the broader talent and operational shifts.

And throughout, we've highlighted just how fast this field is moving and the significant changes potentially on the horizon for how all of us get work done.

So, here's a final thought for you listening to really chew on, thinking about your own organization or even your specific role. What processes or tasks do you think could be most fundamentally changed or perhaps significantly improved by this kind of intelligent, autonomous AI agency? Definitely something worth pondering, the long-term potential here.

It really is. The evolution of AI agents is unfolding right now, and we're truly just seeing the beginning of what might be possible as this technology continues to mature.

Thank you for joining us for this deep dive.