AI Agents pose a challenge for Enterprise Software

When I first heard the term "Service as a Software" (SaaS), I thought it was one of those VC puns that kept popping up every few days. But I was wrong! The entire value chain of B2B SaaS (Software as a service) rests on a shaky foundation. AI will have a strong impact on the B2B software stack.

The article by Ashu Garg and Jaya Gupta introduced this idea early last year. It argues that traditional enterprise systems—comprising Systems of Record, Systems of Engagement, and Systems of Intelligence—are outdated and inefficient. These systems, which include tools like CRMs, marketing platforms, and analytics dashboards, are described as fragmented and burdensome. The authors suggest introducing a new paradigm, Systems of Agents, which will streamline these functions and completely transform or replace the current enterprise stack. The rest of this article draws heavily from their work, so if you have not read it, this might be a good time to glance through it.

Their core idea is fantastic. Multiple AI agents, each aware of and able to communicate and collaborate with others, will surely disrupt the entire B2B SaaS space. However, it is too early to declare the demise of SaaS software.

In this article, I share my perspective on the journey to this brave new world and discuss some of the challenges that must be overcome.

The path to developing a System of Agents

There are three steps to developing a System of Agents, described below in order of increasing complexity:

Unification of the traditional workflow

Consolidating the interfaces of various enterprise tools is the initial step. For instance, Freshservice includes a Virtual Agent that can handle most employee queries and service requests. However, employees must log into a separate tool like Workday if they need to apply for leave.

We can unify this system to eliminate the need for employees to switch between different SaaS products for various tasks. Imagine a situation where both IT and HR requests are addressed by separate AI agents, each of which can communicate with an orchestrating AI agent, as depicted in Figure 1 above.

This unification impacts traditional tools in two main ways:

1. The interface no longer serves as a differentiator: This is crucial for SaaS builders to consider. If users don't have to log into different tools, the UX quality becomes irrelevant since users won't interact with it.
2. There will be demands for new APIs: SaaS products expose functionalities through APIs, but current APIs aren't designed for agent-driven workflows. SaaS products must provide well-documented, atomic APIs that function effectively in this new environment.

The basis for competition in the SaaS world will evolve. However, the existing enterprise stack won't collapse immediately. Tools like Freshservice and Workday have complex workflows that AI agents cannot yet fully handle. Agents leveraging LLMs continue to struggle with inaccuracies and hallucinations. This problem is exacerbated when organizations have complex customized workflows unique to their business context. Until these issues are completely addressed, transitioning confidently to AI agents remains challenging.

Service-as-a-Software

SaaS software traditionally includes tools to convert unstructured information into structured formats. For example, Freshservice assigns tags to tickets, like categories, for routing and reporting purposes—tasks typically handled by agents.

However, these features are becoming obsolete. SaaS tools no longer need such capabilities built-in, as AI agents can perform these tasks more accurately and consistently than humans. Even without a designated "Category" field, AI agents can efficiently route tickets and provide the CIO with insights.

This shift has two major implications:

1. Commoditization of SaaS: Over time, SaaS tools may evolve into primarily serving as databases, with AI capabilities taking over traditional functions.

2. Reduction of Seats: B2B SaaS, often sold per seat license, may see a decrease in seat sales as AI assumes responsibilities previously handled by human agents. While automation and AI have long promised to reduce seat needs, AI agents will significantly alter this dynamic, progressively handling more complex tasks.

Despite these changes, I do not foresee the end of B2B enterprise software. Instead, AI agents and SaaS tools will likely complement one another. Use cases involving security, traceability, and audits will remain within the domain of enterprise SaaS tools. These features will support AI agents, addressing ongoing security and safety concerns.

Automation

The third step is automation. AI Agents have become very good at planning. Given a task, they can develop a plan and execute it using the first two steps (above). However, as an insider, this development feels reminiscent of the self-driving car story. The 2007 DARPA challenge highlighted their potential, but even 17 years later, fully autonomous cars are not ubiquitous.

While AI agents can plan and execute tasks, complete independence remains elusive. Much like self-driving cars, where any failure carries significant risks, the stakes in enterprise workflows are similarly high. Imagine an autonomous agent incorrectly canceling an order or terminating an employee—the reputation damage and legal repercussions could be severe.

We are far from fully realizing step three. Although AI agents will take over certain domains, their adoption will be limited. Risk-averse organizations will likely continue valuing the predictability and traceability that enterprise SaaS provides.

Charting the Future

I'm optimistic about the future of AI, but I don't foresee the "collapse the enterprise stack". In the three steps mentioned, steps 1 and 2 seem imminent, which will redefine the competitive landscape for SaaS products. These changes will compel builders to create features suited for an AI-first world. However, the collapse of the enterprise stack is unlikely soon due to concerns about reliability, security, scalability, and practicality in real-world settings. While Systems of Agents could represent a paradigm shift in enterprise software, their widespread adoption will depend on overcoming these challenges and demonstrating their value in various business contexts. In short like self-driving cars, the System of Agent will have to complete years of road tests before it finds widespread adoption.