AI Agents and thinking beyond summarising meetings notes

Do you know that feeling when the wind suddenly changes and the ship veers in a different direction? Well, I don't, since I don't sail. But if I did, it might be similar to what it currently feels like in the AI world. We had all this hype over the last year on Generative AI, and now it's suddenly AI Agents. Whilst I wrote about the significance of AI Agents previously, it's the likes of Microsoft, Accenture, Anthropic and many others that are pushing this new gust of wind.

But why? When the world is still trying to grasp and clumsily meander through what Generative AI means for them, why introduce another new thing? There are many reasons, but I wanted to share two that came to mind.

Firstly, the foundational landscape of Generative AI has begun to stabilise, converging to the large scale players of OpenAI, Anthropic, Meta and Google. This is not a huge surprise if we were to take history of large platform providers as any future indicator. As such, these players and the many smaller ones are looking for new ways to differentiate in the vie for first place.

The interesting aspect here is that the foundations of this new technology are actually stabilising. It means the big players have gotten so big that it’s getting harder for new entrants to compete. That is, even though AI will continue to get better and better, this thing that seems to exhibit some semblance of intelligence and can talk to me like my therapist is here to stay.

Secondly, I don't think Generative AI alone has emphatically convinced many that it's actually delivering on the promised productivity. Recently, LangChain indicated that 58% of businesses say the most suitable tasks for AI are still relatively basic like research and summarisation - which only make up a small subset of professionals' daily activities.

It is productivity... but is it really enough to justify the millions spent?

Interestingly, this is somewhat contrasts with Microsoft's survey results, where ROI is on average 3.7x for every $1 spent on Generative AI. Nonetheless, the goal for many business leaders today is to find direct attachment of tangible and measurable value to AI investment.

But what is the typical simple logic for the measurement of value, despite it being a crude yardstick of total impact? The reduction of time and effort to do things as a result of technology, that otherwise a human would have had to do.

Read, automation. And coupled with Generative AI, enter stage right - AI Agents.

Given the potential direct linkage of this technology to automated workflows, it’s no surprise we’re seeing businesses excited for its potential as a solution to their productivity issues. Its draw is so big that 78% of companies plan to embed AI agents in the coming years. It’s like stumbling upon water in a dry desert.

But the pragmatic reality is that there is currently a lot of talk about it, but not much meaningful change is being made. Surely, AI agents should be doing much more exciting things beyond summarising meeting notes and telling me my emails lack clarity.

To bridge this gap, I believe it's important to be specific in our framing of AI Agents. The common definition of an AI Agent is any LLM-based application that can take actions. But with a description this broad, it encapsulates nearly everything depending on how granular you define an action. This would include tasks like providing feedback, creating images, and code generation. While theoretically accurate as per that definition, it frames the use of AI Agents primarily as a question and answer machine. I ask it questions, it writes me a response.

I believe it is much more useful to think of AI agents as entities that are able to take multiple actions autonomously to achieve specific goals and outcomes. It's a system (not in the IT sense of the word) that has the self-autonomy to reflect, plan, decide, coordinate, communicate, and act alone or with others to achieve these goals.

Why does this semantic distinction matter? Because it crosses the ambiguity inflection point - where potential benefits feel understandable, but the questions suddenly become overwhelming. What can an agent do? How does it decide? If it doesn't work, does it try again in a different way? How do I integrate agents into my workflow without feeling like I'm chatting with a bot all day? These aren't questions for the future. They're questions for today - that shape the future.

If we consider AI Agents this way, we naturally evolve from being primarily concerned about whether the AI's response is complete and accurate (which is still very important) to what it can actually do. It also shifts how we think about interactions with AI beyond the "Is this a human OR an AI activity?" conversation to a "how should human AND AI do this together?" way of thinking to better achieve our goals.

I believe the simplistic draw of the OR category of substitutive thinking creates a sense of uncertainty around real AI productivity value. It forces us to always compare it to whether we can do it better ourselves.

Take for example what we do as modern professionals. Our day-to-day activities are typically more varied than they seem. There's only so many emails or blog posts we write, and so many images to create. Much of our time is filled with nuanced, exception-filled meetings and conversations that don't lend themselves to easy automation.

If we work through each of our activities and identify where AI could substitute for us, we will find hundreds of tasks that are meaningful, distinct, but likely quite small.

With that kind of variety of tasks, and given how specialised and niche AI agents currently are, how can we justify $30/month on an AI agent for a task they only do a handful of times?

To make a real difference, it also won't be just one tool - automating social media might require an AI for research, another for words, another for images, and another for posting. The result? You end up paying for tools you rarely use because your workflow as professionals is multi-dimensional and contextual.

The human OR AI thinking is a constrained and inefficient way of thinking about the opportunity of AI Agents. Of course, in a large business context where there could be many instances of these activities conducted by hundreds of people, the equation could be different.

Where does that leave us? AI agents are at a fascinating crossroads. The hype is real, but so are the potential opportunities and challenges. We need creative and innovative thinking to help us navigate the ambiguity, create better use cases, and reimagine how we can work together with the machine.

If you're thinking of creating or using an AI agent, maybe a good place to start by asking yourself a few questions:

• Can I think of new ways to achieve my goals by incorporating this magical intelligent agent in new ways, knowing that it is accessible today and will only get better?

• What are the various pieces I need to bring together to make it work? More importantly, will I be able to adapt if the picture changes or new pieces become available?

• What can I do to start small and kick-start the imaginative part of my brain to subconsciously explore new possibilities? The challenge is often in imagining what a new human-AI Agent workflow could look like, as it's probably very different from what you're doing now.

Personally, I'm leaning in. There's so much new thinking emerging about this technology's role. I'm eager to see what happens when the autonomous abilities of agents combine with financial mechanisms and new user interfaces. The future of work is being rewritten and I'm excited to be part of it.

What do you think? Are you using or planning to use AI agents? What’s working and what’s not?

If you’re reading this in your inbox, reply to this email or share your thoughts below!