Beyond Expectations: AI Agents and the Next Chapter of Work

AI agents, or autonomous agents, are in their early days. Very early – the bottom of the first inning early. The field is buzzing with innovation, from groundbreaking research to proof of concepts to practical applications – all hinting at AI's vast potential. 

There is no doubt that autonomous agents will transform every single industry, with their capabilities extending beyond mere task automation to redesigning workflows, simulating complex scenarios, and reducing the need for human intervention in various processes. We’re looking at a (near-term) future where agents can run large-scale simulations, redesign marketing campaigns, or even automate complex R&D testing processes.

Boston Consulting Group (BCG) highlights the evolutionary leap from large language models (LLMs) to autonomous agents designed to execute tasks end-to-end, monitor outcomes, adapt, and use tools autonomously to achieve goals. They represent a significant step towards true artificial intelligence, capable of independent operation without continuous human oversight. 

In terms of market size, autonomous AI and autonomous agents were valued at 4.8 billion USD in 2023 and are estimated to register a CAGR of over 43% between 2023 and 2028, reaching 28.5 billion. It's clear that we're on the cusp of a paradigm shift – a phase filled with anticipation, excitement, skepticism, and pragmatic evaluation. This shift isn't just about technological advancement; it's about redefining our very approach to work, productivity, and innovation. Nearly every investor, founder, developer, and tech enthusiast is trying to understand the impact this technology will have on how we work in our lifetime and beyond, and assess the implications for their operations and strategic goals. 

However, as of now, we lack the capability to fully comprehend the magnitude of the mass shift this will cause. All we can do is speculate. This article is just that – my speculation about the unfolding dynamics of autonomous agents and its implications for founders, investors, and the broader economy. I’ll talk about how we at Forum Ventures are thinking about and investing in the space, as well as provide a market map with the companies we believe are leading the exploration. 

Where We Are At Today

Despite the considerable advancements in research and proof of concepts, we're all still trying to make sense of and project out how to harness the full capabilities of AI agents. So far, there is a confluence of three trends:

1. Advancements in AI proficiency and efficiency, expanding the boundaries of what's possible. 
2. The decreasing cost of actioning capabilities, such as ChatGPT 4.0, for example, making the use of AI agents more accessible to more people and causing wider adoption and the overall embracing of this technology.
3. The democratization of access to AI, open source or not, enabling a wider range of entities to explore and implement AI solutions, thereby accelerating the pace of innovation.

As with any new technology, especially a transformation as big as this, there are an array of challenges that are in the process of being addressed. Here are the top two:

1. Safety & Accuracy

There's a growing focus on developing the necessary infrastructure to ensure the safe and ethical deployment of AI agents. For many industries and businesses, there is no room for error. If an LLM has a hallucination rate of even just 0.1% it could never be trusted in any critical process, and this error rate needs to be even lower for a 10 step or 100 step process. Solving this is paramount to widespread adoption, and many companies are waiting before they embrace LLMs either as part of their tech stack or as an entirely new way of operating. 

Tools for monitoring accuracy and safety through observability and user permissioning, as well as ethical frameworks, are being established to foster a responsible approach to AI integration. We’ve seen some companies doing this well, PrivateAI being one of them. They use inference to make sure companies are not training on private data so that it doesn’t leak. We’re also very excited about new companies coming to market like SafeguardAI – an autonomous AI agent that safeguards for hallucinations, allowing enterprises to deploy generative AI use faster.

Additionally, tools like automatic evaluation metrics, human evaluation frameworks, and diagnostic datasets are being developed to assist in the assessment and improvement of LLMs' accuracy. These tools help researchers and developers identify strengths and weaknesses in LLMs and guide further advancements in the field.

2. Human-AI Interaction

The challenge here is to what extent should humans interact with software that’s autonomous. There are concerns about the potential risks of AI systems operating without sufficient human control, i.e. how much autonomy is too much. But we also need to figure out how much we want humans in the loop, and what level of human interaction creates more safety whilst limiting biases and decreasing the chance for human error. We don’t have good answers to this yet, at any sort of reasonable scale.

From an opportunist perspective, I’m hopeful we can define a new paradigm for autonomous software to operate inside the control of humans in a way that it’s being monitored and observed so that humans can stop potentially “fatal” things from happening like a much bigger version of a flash crash in the economy. In my opinion, those who can build this will win and deliver transformational opportunities. 

The Shift from Task-Oriented to Goal-Oriented Processes

There isn’t going to be any sector or field of work that will remain untouched by AI agents, and a lot of the change that happens will be in the near future. In my opinion, one of the most profound impacts that AI agents will have is the shift from task-oriented to goal-oriented processes. Today, you input something into a computer, such as “write me an op-ed about AI Agents”, and the computer gives something back to you, which you then action. This is a very task-oriented prompt, and still requires the user to train the agent according to the goals and tone of voice of the person. However, it is limited to this, and therefore the output is largely determined by the quality of the training input, plus the pre-determined (and possibly limited) goals of the user, which is still heavily reliant on human actions. 

The underutilized power of AI agents is in the power of goal-oriented work. The future will no longer be one of rote step by step process description or complicated prompt engineering for processes. Companies and leaders should shift their thinking of how they build and use autonomous rules-based processes, whereby goals are prescribed and agents determine the best path forward to achieve that outcome (with appropriate human interventions). An example of this could be, “book me an event in New York City with 100 professionals that want to learn about how AI is penetrating the U.S. healthcare market from one of our speakers”. In a case like this, AI will be utilized to operationalize strategic thinking beyond the limited scope of possibility that a simple task could accomplish.

This is a whole new way of thinking and working. There are almost no set of goals we are currently pursuing with a computer that won’t be pursued wildly differently. This will be a fundamental change in how we orient ourselves, and how work is conceived and executed. 

Monetization and Market Dynamics

As AI becomes more integral to business models, traditional monetization strategies are being re-evaluated. For example, right now in enterprise software, generally, customers buy seats and usage. On the consumer side, people make in-app purchases. Our hypothesis is that this will shift such that increasingly, software companies will be able to sell outcomes, rather than tools. Will people and businesses pay for results? For their goals to be reached? We’re not sure yet. But we see this as a reflection of the broader trend towards value-based engagements. However, there are challenges in predicting profitability and managing costs, especially given the computationally intensive nature of AI technologies. 

Deciding Who And What To Invest In At The Earliest Stage

Whenever we’re investing at this early stage, the founder is one of the biggest bets we make  – looking at both founder-market fit and founder personality. With AI Agents, this lens becomes even more important because with so many unknowns, the solution being built today will likely not be what’s being built tomorrow, but the founder will stay the same. So, we look at not only founder-market fit, but also their attachment to the problem, how they look at the problem set differently than the existing paradigm, that they are willing to embrace the unknown, and that they have plasticity and flexibility to keep pace with a market that has this much flux. 

After the founder, we look at the market and if there is a large total addressable market and a credible path to a $1B revenue opportunity. We are open to both legacy markets like proptech and supply chain, and more forward thinking, flexible markets like fintech and eCommerce, as long as the startup solution / tool will deliver a step function improvement over the old way.

Our third focus when evaluating an AI agent solution is if the tool will be compatible within an AI-centric software future. In other words, will the proposed solution seamlessly integrate with and enhance how we see the future software landscape and stack within that market.

We can’t make proper cost-based predictions yet. Right now, AI businesses are fundamentally less profitable than SaaS businesses. The costs associated with processing and analyzing data in AI systems can quickly accumulate. There will need to be near-term progress that enhances AI efficiency and reduces operational costs before we can do this type of evaluation. Ideally, there are advancements that mirror Moore's Law in the AI sector, and both power and chip costs are reduced due to increased investments. If we can find a balance where AI is not only innovative but also economically sustainable, then we’re golden. But there are still so many unknowns, and most of us are guessing (making informed speculations, to put it nicely).

A ‘Brave New World' of Possibilities

Most people consider the introduction of ChatGPT to be AI’s”iPhone moment”. However, I don’t think we’re there…yet. To date, these chat interfaces haven’t done much more than streamline our current workflows. While these tools have undoubtedly made tasks easier to manage, our approach remains fundamentally task-oriented. The broader vision is to transform this dynamic entirely, where AI will be able to operationalize strategic thinking and perform complex output, with even less input from humans. The true iPhone moment, therefore, might be the unveiling of AI Agents as the default B2B application set, which will in turn have an outsized impact on the future of work. 

A decade from now, there is no doubt that we’re going to look back and marvel at the idea that we used to operate based on to-do lists rather than setting strategic goals and allowing AI to help us iterate and refine those objectives. This shift toward a goal-oriented work environment represents not just an evolution in technology but a transformation in how we conceptualize and approach our work. 

The path forward is filled with uncertainties, but the potential for AI to revolutionize industries, amplify human potential, drive meaningful progress, and deliver lasting value is undeniable. Our commitment is to navigate these uncertainties, and identify, bet on, and support early-stage AI initiatives and the brilliant minds that are bringing their visions to life.