*Januray 10, 2025* There's a growing tendency in the market to transfer all the historical conceptual baggage of automation into a new framework for thinking about agents. This isn't merely superficial or due to loose terminology; there are efforts to derive this equation from first principles. Looking at the two a16z articles shared by Stefano last week: - About RPAs: > *By automating tasks traditionally handled by labor, they can now tap into markets and opportunities that were previously too small or too difficult to pursue.* > - About BPOs: ![[Pasted image 20250426162558.png]] The diagram explicitly says that *Human Labour is Error Prone* whilst *AI Agents are Accurate -* a **statement that clearly can **only be taken as wishful thinking. It is obviously dangerous to start deriving conclusions with false assumptions. - McKinsey articles: > *The value that agents can unlock comes from their potential to automate a long tail of complex use cases characterized by highly variable inputs and outputs—use cases that have historically been difficult to address in a cost- or time-efficient manner.* > In the case of BPOs, the articles primarily address high-volume, mundane, and repetitive work—customer support, outsourced IT, financial claims processing. These are relatively simple tasks where successful applications might be built. But are we truly automating the process? For these use cases, it's possible to achieve good enough results because the processes themselves tolerate high error margins, trading errors for scalability. This represents a softer form of "automation." The same is not true for knowledge work, such as creating credit memos. Knowledge work is fundamentally about quality. When organizations consider automation within the context of knowledge work, they employ entirely different assessment criteria. Nearly all financial institutions in developed countries are engaging with consultancies to explore the technology's potential. The more progressive ones are already considering if agentic AI can transform their operations by focusing directly on core revenue-generating processes. All large organizations have established relationships with RPA providers, and their mindset is irremediably coupled with automation frameworks. Transformation programs carrying these old intuitions will inevitably fail because they measure the wrong metrics. To start with, there are structural difficulties for the “Agents will automate labour” hypothesis. Our legal infrastructure is built around human rights, property, and liability. Societies would not tolerate "It was the AI"-style explanations in courtrooms. Labor represents over 50% of global GDP. Agentic labor automation would increase return on capital at the expense of return on labor—something societies worldwide would struggle to accept. As Stefano says, "Being served by a human in a restaurant will be a luxury." The way we value human labor will need to fundamentally shift before such a transition occurs. ### **The technological reality** Despite the hype, we're operating with flawed technology. These models are essentially interpolative databases. The misconception is thinking that "memorization" means storing answers to certain inputs. Instead, the memorization is of "program templates"—programs that transform inputs in specific ways, like *write_like_shakespeare*. From a technical point of view there’s still a long tail of problems whose resolution might be exponentially harder. Even if it is possible to make it work with the current technology type, we’ll still need to wait for scientists to create conceptual frameworks to think about these “program templates” - can we group them into different classes? what does their internal structure look like? how much composable are they? how are they represented across the network? etc. Once the foundational conceptual baggage is in place, we will still need to engineer mechanism to better control these programs and provide us some levels of guarantees, or at least clear trade off interfaces. And lastly these engineering methods will need to be able to run at scale. Thus, it might very well be the case that we’ll never reach a point where AIs can run forever, uninterrupted, without human guidance. Therefore a better metric to measure our progress in a particular domain is productivity per unit of human input. When we think about our customers problems, we need to be pushing the boundary of how much they can get done for what they give. ### Implications for product strategy The previous discussion does not mean Reveal shouldn’t try to focus on attempting to fully automate a mission critical process as a starting use case. Maybe that’s the best thing we can do - we might be able to create a great product by doing a lot of traditional software engineering coupled with LLM-based augmentations. But in that case we need to dial down the prevalent presence of the “agentic” gene in our DNA in favour of a more targeted focus on the problem - relegating the tech to a background plane. We're having productive discussions about "problems of the future," but the present challenge might be how organizations conceptualize and scope agentic adoption. From a product perspective, we need to discover new patterns for knowledge work. If I do a thought experiment about investing in pre-seed companies, I'd choose those focusing on discovering new knowledge work patterns for humans—viewing agents as intelligence amplifiers, emphasizing the "assistive" component rather than the "automation" angle. Our current concept of a Workspace is a good initial interface for the agent to publish its work. But maybe even ourselves our falling into the old “standardisation” trap that stems from automation initiates. The Plan is only one - the source of truth for the process. Why not let every user have an individualized Plan? Wouldn't a product built on this premise reinforce human accountability and position the agent in a more assistive role? I maintain the same worldview I had months ago. From a product vision perspective, the winning structure is the meta-cursor approach—a factory for cursor-like experiences tailored to specific business processes within an organization. The major challenges lie in managing collaboration between workers, segregating the Master Plan from Individual Plans, and implementing knowledge management. That's a tool that truly empowers knowledge workers while providing automation-like benefits. Essentially, the workspace would be an environment where a digital version of you begins drafting work in your historical style, and you take it from there. Instead of agents we should call them clerks or shadows.