A few months ago, at our Change Leadership Conference, someone asked a room full of change leaders a simple question: Who here uses AI agents?

Some hands went up. Most didn’t. I personally was hesitant to raise mine.

My first instinct was to say no. Then I actually stopped and thought about it. I was already using AI in different ways, but I hadn’t really stopped to ask whether what I was doing counted as using an AI agent. When I went back and started unpacking what an AI agent actually is, I realized something important.

I was already using one. I just didn’t have the language for it.

That moment stayed with me, and it became the major inspiration for the AI Agent Lab for Change Leaders with Alex Jackson, who had joined us at the Change Leadership Conference as a speaker. Not because I believe every change leader needs to become technical, and certainly not because I think we all need to rush out and build our own AI agents from scratch. Rather, I think many of us are in exactly the position I was in: using AI in our work, hearing more and more about agents and agentic AI, and wondering whether we’re behind because of all the ‘noise’.

The language is still evolving, the technology is moving quickly, and even the organizations building these technologies don’t have one universally accepted definition of what an AI agent is.

So let’s start there.

 

What is an AI Agent?

Strip away some of the noise and the definition is simpler than the hype can make it sound.

For our AI Agent Lab, Alex used this as our working definition:

An AI agent is an AI model given a role, a goal, access to tools and knowledge, and permission to take steps toward that goal on its own, with a human deciding where the boundaries are.

One way to think about it is that the AI model is the brain, while the tools and knowledge you give it become the hands and arms that allow it to interact with the world and actually do something rather than simply talk about what could be done.

The distinction that matters most for change leaders is not necessarily the technical architecture. It is the question of who is driving the work and how much autonomy the system has.

A chat assistant generally answers what you ask it, one exchange at a time, and you decide what happens next. An agent, by contrast, can be given a goal and then work through a series of steps toward that goal, using the tools and information available to it and returning to a human at the points where review, judgement or intervention is required.

That distinction becomes increasingly important as AI moves from generating responses to actually performing work.

Nobody Has Completely Settled the Definition of an AI Agent

If you’ve ever felt as though everyone seems to mean something slightly different when they talk about an AI agent, you’re not imagining it.

During the Lab, we looked at how several major organizations in the AI and technology space describe agents. We saw different perspectives from organizations including OpenAI, Anthropic, Google Cloud, Microsoft, McKinsey and Gartner. Some emphasized autonomy, some described agents as part of a spectrum, while others focused on the ability to plan, act and interact with tools.

 

There isn’t one definition that everyone has agreed to use.

And I think that is worth acknowledging because sometimes we put unnecessary pressure on ourselves to understand terminology that is still being defined and refined by the very organizations developing the technology.

If the organizations building AI haven’t settled on one definition, you didn’t miss a memo.

What does seem to converge across the different perspectives is the idea that an agent is more than a system that simply responds to a question. It is given a goal and has some degree of autonomy to determine and execute the steps required to move toward that goal.

That gave us a useful working definition without getting lost in the terminology.

You May Already Be Using AI Without Realizing It

One of the questions we asked participants during the Lab was: Where are you on the AI agent journey?

We gave them three starting points:

1. I have no idea what an AI agent is.
2. I already use a pre-built AI agent.
3. I’ve built one from scratch.

There were people across the spectrum.

And that is important because the first step isn’t necessarily learning how to build an agent. The first step is understanding where you are.

Take for instance, many of us have been interacting with AI-powered systems for years without thinking of ourselves as people who “use AI.” Recommendation engines, predictive systems and automated decision-support have already become part of everyday digital experiences.

What is changing now is the degree of autonomy we can give these AI models and systems and the range of work they can perform.

So before you decide you need to catch up, take a moment to ask yourself a different question:

What am I already using, and what could I potentially do next?

The Six Decisions Every AI Agent Needs

Whether you use a pre-built agent, configure an existing tool or build an agent from scratch, there are six decisions that remain fundamental.

Role. Goal. Knowledge. Tools. Boundaries. Oversight.

These became the foundation of our AI Agent Lab because they provide a practical way of thinking about agent design without requiring you to begin with the technical details.

And there is something else I find particularly useful about this framework provided by Alex Jackson: these aren’t entirely new concepts for change leaders. We already think this way when we design and lead change.

We define the purpose and outcomes we are trying to achieve, clarify roles and responsibilities, identify the people and information needed to move the work forward, establish governance and decision rights, set boundaries around what is in and out of scope, and determine where leaders need to remain involved. For example, when we develop a change strategy, we don’t simply tell a team to “go manage the change.” We clarify what the change is intended to achieve, who needs to be involved, what information and resources are required, what decisions can be made at different levels, what needs to be escalated, and where leadership oversight is required.

The six decisions Alex introduced apply that same thinking to an AI agent. The technology is new, but the discipline of defining the work, setting expectations, establishing boundaries and maintaining accountability is not.

1. Role

What job are you asking the agent to perform? Think about this as a structured job description rather than a personality. If you were bringing a junior team member into a role, what would you want them to understand about what they are there to do, the behaviours you expect and the type of work they should be capable of performing?

The clearer the role, the easier it becomes to establish what the agent should and should not do.

2. Goal

What are you actually trying to accomplish? The goal provides direction and gives you a way to determine what “done” looks like. One of the practical ideas we explored during the Lab was to start with the outcome and work backwards. If you know what you want the agent to produce, what would need to happen to get you there?

That question helps you identify the steps, behaviours and information that need to be built into the agent.

3. Knowledge

What does the agent need to know that it cannot reasonably know already? This might include organizational documents, project information, stakeholder data, policies, processes or other context relevant to the task. An AI model may have a vast amount of general knowledge, but it does not automatically have your context. And for change leaders, context is everything.

A generic change communication is one thing. A communication that understands the organization, the change, the affected stakeholders, the current situation and what people are experiencing is something very different.

An AI model may have a vast amount of general knowledge, but it does not automatically have your context. And for change leaders, context is everything.

4. Tools

What does the agent need access to in order to perform the work? This could include files, applications, systems, connectors or other digital tools. But access should be deliberate. Just because you can give an agent access to something doesn’t mean you should.

Think about the minimum access it needs to accomplish its goal, rather than automatically giving it the keys to the castle.

5. Boundaries

What should the agent never do? What is outside its scope? When should it stop? When should it escalate? This is actually a discipline that change leaders already understand. Every project has scope. Every governance structure has decision rights. Every major transformation has communication boundaries. We regularly decide what people can determine themselves, what requires escalation and what information should not be acted upon without additional review.

The difference is that now we need to apply that same thinking to an AI system.

During our stakeholder mapping demonstration, for example, we deliberately established boundaries around what the agent should and should not infer about people. It was not asked to speculate about individuals’ motives, personal circumstances or performance, and we deliberately avoided framing people as “obstacles.”

Those decisions matter because an AI system can produce a confident-sounding answer even when the information available does not justify the conclusion.

6. Oversight

Where does the human come back into the process? That might be at every step. It might be at defined checkpoints. It might be when the agent encounters uncertainty or reaches a particular decision. The important thing is that oversight is intentional. More autonomy does not mean less accountability. The agent may be able to perform the work, but the human still needs to determine whether the work is appropriate, accurate and fit for purpose.

Your context, judgement and accountability do not disappear simply because more of the work has been automated.

 

Think of an AI Agent as a Junior Team Member

One of the mental models Alex introduced during the Lab was to think about an AI agent as a capable, eager junior team member. That analogy resonated with me.

A junior team member can be remarkably capable. They can work quickly, take on significant amounts of work and produce something that gives you a running start. But you would not normally bring someone into an organization, give them unrestricted access to your systems, provide no role definition, no boundaries and no supervision, and then expect everything to work perfectly.

The same principle applies here. The technology may be different, but the leadership discipline is familiar.

You give the agent a role. You establish the goal. You provide the knowledge it needs. You determine which tools it can access. You establish boundaries. And you decide where human oversight belongs. The only genuinely new part is the team member you’re applying those principles to.

Where Are You on the AI Agent Journey?

Another concept we explored during the Lab was the progression from working unaided to increasingly autonomous AI-supported work.

• At one end, you are working entirely unaided, as most of us did until relatively recently.
• Then you move into asking, where you give an AI assistant a question and receive an answer.
• From there comes tasking, where you give AI a task that involves multiple steps and review the result.
• Further along is delegating, where you give an agent a goal, establish the boundaries and allow it to work through the task before bringing the result back to you.
• And at the far end is orchestrating, where one agent may coordinate the work of other agents.

There is no finish line on that spectrum, and there is certainly no prize for reaching the most technically advanced end.

The better question is: Where are you today, and where do you actually need to be?

For some people, a pre-built agent will be more than enough. Others may want to configure a tool they already have access to. Someone else may eventually want to build an agent from scratch. All three are legitimate places to start.

Do Change Leaders Need to Be Technical to Use AI Agents?

No. And this was one of the things we wanted to make very clear during the Lab.

You don’t need to become an AI engineer to begin using AI agents thoughtfully. In fact, many of the disciplines required to work effectively with agents are already familiar to change leaders. We already think about roles, goals, stakeholders, governance, scope, decision rights, risk, communication, context and oversight.

The technology may be new. The leadership disciplines are not.

This is one of the reasons I believe the conversation about AI agents belongs within change leadership rather than sitting entirely within the technology function.

Why AI Agents Are a Change Leadership Issue

AI isn’t simply changing technology. It is changing how work gets done.

As organizations introduce AI into their workflows, it can change roles, responsibilities, processes, decision-making, expectations and the way people interact with information. That means organizations aren’t simply implementing another technology. They are navigating another layer of change.

And the questions change leaders need to ask go beyond: What can this technology do?

We also need to ask:

• What happens to the work?
• What happens to people’s roles?
• What decisions should remain human?
• What skills will people need?
• How will managers lead teams that are working alongside AI?
• How do we build trust while recognizing that AI can still get things wrong?
• What boundaries and governance do we need?
• And how do we help people adapt to a different way of working?

This is where the change leadership conversation becomes particularly important. AI may change the work. People still have to navigate the change.

What Can an AI Agent Actually Do? A Stakeholder Mapping Example

We didn’t want the AI Agent Lab to remain theoretical, so we built an agent live.

We deliberately chose stakeholder mapping because anyone who has worked on a large change initiative knows how much work can sit behind a stakeholder assessment. You gather information, work through spreadsheets, analyze stakeholder data, try to identify patterns and then turn all of that information into something meaningful enough to support decisions and engagement.

If you’ve ever worked on a major ERP implementation, for example, you know the volume of information that can sit behind a stakeholder impact assessment. And if you’ve ever had an Excel workbook quietly break while you’re trying to make sense of that information, you know the pain of the analysis that comes afterwards.

For the Lab, we used a fictitious organization and built a stakeholder mapping agent. The agent was given a defined role and goal, the relevant knowledge and tools, boundaries around what it should and shouldn’t infer, and clear expectations around human oversight.

The result was a structured workbook as well as an interactive visual representation of the same information.

That was the point where the potential became tangible. Not because the agent had replaced the change leader. It hadn’t.

The value was that it could provide a running start on work that might otherwise consume significant time, allowing the change leader to spend more time interpreting the information, applying context and making decisions about what it actually means for the people and organization involved.

AI can accelerate the work. It does not eliminate the need for change leadership judgement.

Start With One Real Task

If you’re wondering where to begin with AI agents, I wouldn’t start by trying to transform your entire job.

Start with one real task.

Think about the work you already do that is repetitive, time-consuming, data-heavy or simply takes more effort than the value it creates. Then work through the six decisions.

• What is the role?
• What is the goal?
• What knowledge does it need?
• What tools should it have?
• What are its boundaries?
• Where does human oversight belong?

You don’t need to build an agent from scratch to do this exercise. You can use a pre-built tool, configure something you already have or simply experiment with the principles inside the AI assistant you are already using. The important thing is to begin experimenting with a real piece of work rather than getting caught up in trying to understand every new development in AI.

Garbage In, Garbage Out Still Applies

One of the oldest principles in technology still matters here: garbage in, garbage out.

If the information, instructions or context you provide are poor, the output can be poor too. The difference with AI is that the system may still produce an answer, even when it doesn’t have enough information to support one.

That is one reason boundaries are so important. A boundary can be as straightforward as: If you don’t know the answer, tell me you don’t know.

The human in the loop also remains important because AI does not eliminate the need for validation. The model can process information quickly, identify patterns and produce outputs, but it does not have your lived experience of the organization, your understanding of the relationships involved or your accountability for what happens next.

That is where your judgement still matters.

More Powerful Isn’t Always the Right Tool

Another point that came up during the Lab was the temptation to assume that the most powerful model must always be the best model. It isn’t.

Alex described this as a kind of curve. At one point, you have enough capability and context for the model to do what you need. Beyond that point, more capability can simply introduce unnecessary complexity.

If you ask a highly capable model to solve a simple problem, you may end up with a technically impressive answer to a question you never actually asked. The goal isn’t to use the most powerful AI available. The goal is to use the right level of AI for the work.

For many change leaders, that may mean starting with the lower-value, repetitive activities that consume time and then gradually experimenting with more complex use cases as confidence and capability develop.

Treat Data Privacy as a First Decision

Data privacy cannot be an afterthought when you’re working with AI agents. If you’re working with confidential organizational information, personal information or sensitive stakeholder data, you need to understand what tools and environments your organization has approved for that purpose and what the relevant licensing and data controls actually allow.

This is particularly important for change leaders because so much of our work involves people and information about people.

The fact that an AI tool can access information does not automatically mean it should. Data access is part of the boundary you establish when designing the agent.

What We Wanted the AI Agent Lab to Do

When Alex and I started putting the Lab together, our goal wasn’t to turn change leaders into AI engineers. We wanted to demystify AI agents.

There is a lot of noise around agentic AI right now. New models, new tools, new terminology and increasingly ambitious claims about what these systems can do. It can be difficult to know what matters and where to begin. So we wanted to bring the conversation back to something practical.

• What is an AI agent?
• What decisions do you need to make?
• What tools can you use?
• What information does it need?
• Where are the boundaries?
• Where does the human remain accountable?

And perhaps most importantly: How could you use this in the work you are already doing?

The Lab was designed to give people a starting point, not to suggest that there is one right way to use AI agents. The technology will continue to evolve. The tools we use today will change. New capabilities will emerge, and some of the terminology we are using now will probably change with them. That makes the ability to experiment and learn even more important.

The Journey Has Already Started

AI agents are not something we are waiting for.

They are already here.

And whether you realize it or not, you may already be using them.

You don’t have to become an AI engineer. You don’t have to build an agent from scratch. And you don’t have to understand every new model, tool or piece of terminology that emerges.

But you do need to start paying attention.

Learn what these tools can do. Experiment with them. Understand where they can genuinely make your work easier and where human judgement, context and accountability still need to remain firmly in the picture.

And most importantly, learn how to lead the change they are creating.

Because AI agents may take on more tasks. They may automate more processes. They may fundamentally change how work gets done.

But they don’t take away the need for people who can lead, navigate and adapt through change.

That part is still ours.

And perhaps the most useful place to begin is simply to ask yourself:

Where am I on the journey, and what is one thing I can experiment with next?

 

Continue Exploring AI Agents for Change Leadership

If you attended the AI Agent Lab, use the resources as a starting point for your continued experimentation.

If you missed the session, you can still explore the Change Agent Lab to work through the six decisions and begin applying them to a real task of your own. The Change Agent Lab was developed exclusively for this workshop by Alex Jackson, Experience Innovation.

You don’t need to know everything about AI before you begin.

The journey doesn’t begin when you know everything about AI. It begins when you start experimenting.

 

➜ If you are looking to strengthen your change leadership capabilities within your team, explore our Change Leadership Training programs. You can also learn more about our upcoming Change Leadership Conference, where change leaders gather to discuss the future of leading change.