The hard part of AI agents is the management, not the model.

Hello, I'm Amir, and I'd like to invite you to this short session where we talk about AI agents and what they mean for executives and business leaders. By way of introduction, these are some of the companies I've worked at — AWS, Teva Pharmaceuticals, Motorola, and some startups — and alongside them, some of the customers I've worked with on innovation, on AI, or a combination of both. You're welcome to connect with me.

I want to take you through four sections, quickly and at a high level: what AI agents are, what the opportunities and use cases are, what the implications are for leaders, and then some takeaways to act on.

Let's start with AI agents, the opportunities, and the use cases. Before we get into agents, this is something I always start with: if you haven't realized it yet, intelligence is now a commodity. There are many intelligent tasks and activities you can delegate to tools that perform them as well as humans — or even better, in some cases — and it costs very little, fractions of a dollar per task. That has many implications for the way we do our work.

It's worth spending a minute on the generative AI capabilities. You've probably experienced some of them. When you've chatted with a tool like ChatGPT, Claude, or Gemini, it was probably through text. But text isn't just chatting — it's good at summarizing documents, reviewing and understanding, expressing sentiment, and extracting data. It's good at translation, even though it wasn't explicitly trained to translate. And because programming languages are languages, it's good at working with code too.

There are other capabilities. Visual assets — creating images, video, 3D rendering. Audio and voice, which can be transcribed or generated. Data augmentation — connecting structured and unstructured data sources and extracting insights from them. And an emerging category around physical assets, where AI models can either control physical things like drones, robots, and autonomous vehicles, or predict physical properties such as molecules and materials in materials science. Lots of capabilities to tap into.

I divide the business opportunities in generative AI into three pillars. The first, where leaders and companies often start, is boosting productivity in internal workflows. The second is creating value for customers and stakeholders — through communication channels, support, training, and inside your products. The third is disruption: of business models, operating models, even whole industries and economies.

On boosting productivity: this is where you perform tasks and workflows faster, at larger scale, with higher quality, and at lower cost — if you use the tools correctly. In content development, for example, where you have a workflow with tasks along the way, you can use tools as co-pilots to generate more content, and types of content that would otherwise be impossible to create, and to optimize the workflow. I use frameworks to break the opportunity down: break it into steps, identify the inputs, outputs, and data sources, and then seek out the generative opportunities.

On creating value: this is how you interact better with customers and other stakeholders with the help of AI. The capabilities come in — helping you create personalized communication, letting customers interact with your data, your products, and all kinds of help sources. It's good at identifying emotion and acting on sentiment, and even at expressing empathy — maybe scary, but that's the truth — and it's getting better. It's great at creating multimodal experiences: sound, music, visuals, video. More and more, we can let people experience our products and our brands through these tools. And because it has read countless stories, it's great at telling them, if you know how to use it. Humans connect to stories.

There are two ways to think about creating value with AI. One is enhanced services — training, support, communication — around the product rather than in it. Here's an example: a chatbot customers can talk with about an online course, asking questions about the content and how to implement it, in multiple languages; you can speak to it in Spanish or Chinese even though the content itself is in English. The other is AI-powered products. We're seeing more and more AI inside products like Miro, UX Pilot, and Canva, and inside tools like Jira, where you can speak in natural language to create things that previously required deep technical skill. I often use a canvas that separates the user-journey opportunities from the product-innovation opportunities.

Now to disruption — and this is where agents come in. This is going to disrupt value chains, business models, operating models, and more. 2025 has been the year of agents — or has it? It has definitely been the year of agents in software engineering, where a number of products took the market and grew in adoption very quickly — but not so much in other functions yet. We're starting to see that change, with more tools arriving for other functions: Claude for Excel, healthcare applications, and agents becoming part of the flow in marketing, content generation, sales, and elsewhere.

So what is an agent? How is it different from the AI tools we're used to, and how does it fit with humans and the tools we use? This is a simplified, high-level version, but it should help you understand the possibilities, risks, and implications. First, an agent has a job. It's not a tool where you ask a question and get a response, like a chat tool. You define a high-level job and ask it to achieve that job and its goals, repeatedly and continuously — much like hiring a person and giving them a job and a goal.

To execute that job, an agent needs access to tools — email, databases, communication channels, tools to create code and content. Because the job is defined at a high level, it also needs agency: it has to make decisions about what to do next, how to operate the tools, when to communicate, when the job is done, and what good quality looks like. And it needs a feedback mechanism — either inspecting its own outputs and outcomes, or getting feedback from people or other agents on what it did and whether it should continue or change course.

So what's the role of humans here? First, we define the job — what we want the agent to do, the tasks, what it should achieve, and how it fits the bigger picture. Second, we give it context — what's already been done, our brand voice, our assets, what it should and shouldn't do, that we care about security, information about our company and customers. The more context you give it, the better it performs. Third, we decide how much autonomy to give it — should it check with you at every milestone, or run end to end on your behalf with minimal control? Depending on your context — a regulated environment, how much you trust the agent, how new it is, how adoption is going — you'll make different decisions, and you may change the autonomy over time.

Fourth, you give it access to tools. If it needs to read your email, it needs credentials; if it needs to message you or your teammates on Slack, it needs Slack access. And with access comes risk, so you'll need to decide how to control permissions and credentials when it acts on your behalf. Fifth, you put in place ways to monitor both the actions and the outcomes — how it's making decisions, what judgment calls it's making — because sometimes you need to rein it in, and because you need to keep improving it with continuous feedback based on the quality of its outputs.

To make this concrete: we have workflows we've run professionally for years, with teams doing different things. Designing a user experience, for example, follows a process — written or unspoken — and before AI, it's all driven by humans. With co-pilots, humans still drive the process, but every now and then they turn to a tool to help with a specific task — designing a wireframe, checking the brand, setting up an A/B test, documenting something. That's not agents yet. Neither is the next level: an AI-powered workflow with some automation, where the human still decides the flow, the steps, when to move between them, the trigger to start, and when it's done — but the steps are automated and connected, with outputs handed from one to the next.

Agents are the next step. Here, humans define a high-level job and give the agent access to tools, and the agent decides which tool to use, which task to do when, and when to come back to the human with a result. And we're increasingly seeing agentic teams — not one agent, but several with different specialties and tools, plus an orchestrator, an agent manager that decides: this agent does research, this one builds, this one tests, and so on.

Summary and key takeaways. First, AI agents have major value potential, and you should be exploring where that potential is relevant to your context. Second, select the right tool and approach for the use case — sometimes co-pilots, sometimes automated AI-powered workflows, and sometimes agents. Don't choose blindly, and don't go all in on agents if you don't need them — especially if your organization isn't yet mature in its data, in people adopting these tools, and in safety mechanisms, or if you face strict regulations. Third, on regulations and guardrails: you need to develop the skills and capabilities as more agents arrive — guardrails, judgment, and governance — because in the end your teams are responsible and liable for the results.

It's a very exciting time to build with AI. To do so, you need to think big — and I invite you to. I also invite you to connect with me: I have a weekly newsletter on leadership, leadership principles, and business value with AI tools. I hope this was valuable. See you around.

Lightly edited from the talk for readability.