Workflow is the Moat in AI

Artificial intelligence, and specifically agents, fundamentally changes the value proposition of enterprise software. It shifts the value from processes to work. Instead of following rigid, predefined processes, AI can dynamically determine the best path to achieve a given outcome. In other words, the unit of value is the work to be done.

What does this mean for building software for the enterprise customer? In order to design AI software, it must be designed and built oriented towards efficacy at “doing work”, completing the right tasks, the right way, without human intervention. For ease of discussion, we are going to focus on “agentic software” or “AI agents” and use the definition of “software systems designed to autonomously perform tasks, make decisions, and interact with their environment on behalf of users or other systems”. With this in mind, agentic software requires two primary and related capabilities: understanding of human workflows and integration with existing data technology.

• Human Workflow: Effective agents must be able to seamlessly “hand off” to and from humans in an integrated, complex process. Similarly, agents should be utilized within parts of a larger system, department, or organization in the areas most applicable for automation. Both of these requirements demand a deep understanding of what tasks humans do today, how each of them are done, and how they integrate.
• Data Integration: Valuable agents need to be able to speak the “language” of the existing technology and data. They sit on top of the data and related infrastructure, and must be effective and accurate with this data, in whatever form it is or can be organized/converted to.

Being effective at both of these requirements is challenging. That said, within the context of AI and agents and in a world where model capabilities converge or even plateau, it is also your moat.

So what does this mean for those building? Addressing this challenges requires being incredibly close to the customer. In fact, in contrast to the frequent advice of investors worrying about ability to scale, in the early days, many entrepreneurs should not being afraid of a service model, where you are hands-on with your customer doing work for them, rather than just providing software. The circumstances when such a go-to-market approach is productive is when it reveals details of your customers workflow and data. Allow your early customers to show you their workflow, where it might be similar to comparable customers, and where it might be specific to them. You are specifically interested in identifying the hand-off points between agents and/or humans and value of various tasks. Secondly, engage the customer with opening up their data to you, not the specific data but the type of data that matters, how it is used, and which software utilizes the data today. As previously described, both of these activities are core to providing a valuable solution to this customer but, more importantly, it allows you to build better and more defensible AI software for the next customer, developing a moat.

AI software changes the value metric of software in an enterprise, shifting the focus to “work to be done”. At the same time, as foundational model performance converges and is an inadequate point of differentiation, mastering human workflows and data integration becomes the competitive moat. Such a requirement makes a “service model” in the early days more appropriate, should the ROI on such received knowledge exceed the cost today.