A System of Agents brings Service-as-Software to life READ MORE

A System of Agents brings Service-as-Software to life

How builders can tap into the $4.6 trillion opportunity as AI transforms software from tool to worker.

Ideas / Points of View / A System of Agents brings Service-as-Software to life

10.31.2024 | By: Joanne Chen, Jaya Gupta

Software stands at the threshold of the most profound change in its history. 

Six months ago, we wrote about AI leading a paradigm shift from software-as-a-service to service-as-software, where software is no longer simply a tool for organizing work; software becomes the worker itself, capable of understanding, executing, and improving upon traditionally human-delivered services. The shift represents a technical evolution and also a reimagining of software programs: what was once a tool for guiding human services is now the service provider. The market opportunity in this transformation is enormous—$4.6 trillion in the next five years —as AI begins to eat into in-house salaries and outsourced services.

We’ve since engaged with hundreds of startups racing to use AI to reshape traditional service industries, from sales outreach to recruiting to IT services. Their collective efforts, especially in sectors with low software penetration, like insurance and legal, have revealed the next crucial question: What does it actually mean to translate human services into AI-powered software? How does the service-as-software idea get put into action? Here is how we see software evolving from simple workflow automation to a System of Agents. 

Phase 1: The workflow-based world


Phase 1 software marked an important step in digitizing business processes but suffered from incomplete or biased data inputs, leading to frequent inaccurate predictions. 

In the SaaS era, startups built software to assist humans with workflows of varying complexity, a model that fueled enterprise growth for decades. Salesforce is the archetypal example of such a System of Record. 

The core value proposition of Salesforce is straightforward: a sales representative manually inputs activities and customer interactions, and in return, Salesforce organizes and streamlines that person’s workflow. This is a significant improvement over a paper based system, but is fundamentally limited for two core reasons: it relies on structured data, and much of that structured data must be inputted manually.

Salespeople are often so busy selling that they do not spend tedious hours updating a CRM. When they do update the CRM, their inputs are often optimistic and might not reflect reality. And the data that can be automatically captured is structured, leaving the rich context of actual sales activities—happening in emails, calls, meetings, etc.—on the cutting room floor. A sale happens outside of neat rows, columns and predefined fields. In fact, unstructured data within companies vastly outweighs structured data, with some estimates suggesting approximately 80% of corporate data is unstructured.

Because a Phase 1 CRM (i.e Salesforce) has limited, missing, or biased inputs, its predictions based on this data are oftentimes wrong. Software like this marked an important first step in digitizing business processes, but is far from the end of software’s evolution or capabilities. 

Phase 2: Enter AI 


Today’s software is more likely to capture and process both structured and unstructured data, and to generally play a more proactive role in daily work. 

Today’s systems don’t passively wait for net new input from humans, they actively capture and process both structured and unstructured data in real time. 

The key innovation that enables this are LLMs—which can process unstructured data at scale and across multiple modalities—and agents, which can understand context, make decisions, and take actions. Agents take an overall goal, break it down into a series of tasks, execute that sequence by transferring output from one step to the next, and ultimately combine outputs to achieve an outcome. Agents autonomously collect unstructured data in the right context. They independently initiate actions based on assessments of a given situation, and they make the decision to call multiple interacting systems.

In sales, a CRM takes face-value data from reps—like marking a deal “80% likely to close.” A system built with agents sees the actual signals that indicate deal momentum, monitoring real-time information: the CFO getting cc’d on email threads, technical requirements documents circulating, meeting attendees shifting from technical evaluators to financial decision-makers. The system spots deal momentum before the sales team even logs into the CRM.

In healthcare, patient intake once relied on multiple staffers inputting structured data across fragmented sources like paper faxes, form entries, phone calls, and more. But an agentic system collects a far wider variety of inputs and does it all automatically—voice transcripts, insurance APIs, and EMR system data in one place. Our portfolio company Tennr leverages LLMs trained on medical records, plus an AI agent, for workflow orchestration and real-time processing and verification systems. This produces more comprehensive outputs: integrated patient information, automated scheduling, verified insurance status, initiated authorizations, and complete digital records.

Other applications that show the transformative potential of AI agents include supply chain and logistics optimization.

Phase 3: The future, a System of Agents

A System of Agents becomes a System of Work—providing actual decision-making and task execution. Each agent creates new data that is fed back into the inputs, which then re-train and improve all agents in the system.

In Phase 3, AI agents begin working with each other, in a System of Agents. 

A System of Agents mirrors a human team, with agents taking on collaborative, specialized tasks and continuously learning from each other, just as human teams do. Where traditional software provides tools and workflows for humans to complete tasks, the System of Agents becomes a System of Work—providing actual decision-making and task execution and ultimately delivering services like a highly-trained team.  

The latest research on AI agents shows that multiple agents working together achieve better results than one. By training groups of agents to collaborate, compete, and train each other in pursuit of shared goals, developers create systems that dramatically exceed the capabilities of any single agent. Discrete tasks within a workflow are broken down into components and assigned to the agent best equipped to handle them. As each agent processes its part of the task and passes information to a human or the next agent, the output is progressively refined and improved. Through such specialization, the resulting agentic systems can achieve results that generalist agents struggle to match. In addition, each agent is creating new data that is fed back into the inputs that re-train and improve each agent in the ecosystem.

In the enterprise sales arena, this means an AI sales representative agent will excel at prospect engagement and qualification, an AI solution engineer agent will be able to map out products to meet technical specifications, and an AI account executive agent masters deal dynamics and closing strategies. When the agents collaborate, they form a learning ecosystem. The AI AE flags a technical question to the AI SE, which generates a solution. The AI SDR uses this interaction to refine its outbound. Each agent learns not just from its own interactions, but from checking results with human teammates and observing how other agents handle specialized tasks.  

This is the essence of service-as-software. The most sophisticated sales organizations succeed because different roles complement each other—SDRs prospect efficiently, Solutions Engineers handle technical validation, and Account Executives orchestrate complex deals. AI is evolving along a similar trajectory, with each component enhancing the others to drive greater impact.

System of Agents 50

We’re still in the early innings of building Systems of Agents. After hearing from hundreds of startups, Foundation Capital is excited to announce our list of the top 50 startups making progress building towards this vision.

Key considerations for building Systems of Agents 

​​Startups building with agents are not just adopting new technologies; they are reshaping what’s possible in their respective areas. Here are a few things to consider. 

Position at the data source, own the path

One transformative power of a System of Agents lies in its placement: right at the source of data creation. For builders, the key is finding roles and functions inside a company that touch the company’s most powerful data and sit in front of the System of Record. Using a System of Agents to synthesize that data—whether unstructured or structured—means you own the right to initiate actions with other roles and systems. When an agent owns the interface where data is born, your system then has the right to orchestrate every downstream action.

By positioning themselves at this critical juncture, Systems of Agents capture and act on information in a raw, unfiltered state. They don’t rely on existing data in a CRM or ERP system. Instead, they capture this unfiltered data at the point of origin, and use LLMs to derive insights and initiate actions. This approach transforms previously inaccessible data into usable information. Business processes move from “record first, act later” to “act first, record seamlessly.”

Move beyond software budgets, tap workforce spend

Another consideration for builders: Almost all companies have budgets for employee salaries; fewer have budgets for software. AI has traditionally competed for software budgets. But when a System of Agents can complete a whole job task, it can be categorized as a personnel cost, not a software expense. This opens up a vastly bigger market opportunity; an AI solution can now access a trillion-dollar budget landscape, rather than competing in the crowded software space. Think of it this way: a software juggernaut like Salesforce generates only $35 billion annually—a fraction of the $1.1 trillion spent globally on sales and marketing salaries.  

Sources: Turing, Statista.

As founders look for starting points, they can position their System of Agents as solutions to labor-related challenges across different sectors, specifically by alleviating labor shortages, supporting aging workforces, and meeting the demand for 24/7 operational capabilities. 

In sectors facing critical labor shortages, such as cybersecurity, the demand for qualified professionals far exceeds supply. This gap leaves companies vulnerable to the heightened risks of data breaches and sophisticated cyber threats, where AI agent systems could plug an $8.5 trillion gap in potential lost revenues. Here, AI agents can autonomously manage many of the time-consuming, high-skill tasks traditionally performed by human analysts, such as scanning logs, analyzing threat patterns, and coordinating rapid incident responses.

The need for AI is also more pressing in some sectors than others. The accounting industry faces a different kind of labor challenge: a retiring workforce, with roughly 75% of U.S. accountants nearing retirement age. This aging labor pool threatens to disrupt critical financial management functions that demand accuracy, precision, and regulatory compliance. A System of Agents offers a solution for addressing such a disruption by serving as an aid as the workforce transitions. 

Staffing agencies frequently step up hiring of blue-collar workers during peak seasons. Instead of going through the time- and labor-intensive process of bringing on new  recruiters to handle the increased workload, staffing companies can use an AI agent like Converz, which streamlines outreach to potential workers at scale, automating processes like phone screens, texts, and back-and-forth emails. The more efficient approach lets agencies screen candidates much faster than human recruiters could, in the end placing workers much more quickly.

24/7 AI scalability creates new workstreams

Sectors like medicine present yet another labor challenge: the need for uninterrupted, 24/7 operations. In healthcare settings, constant patient monitoring and timely emergency responses are critical (and often required by law). Unlike human employees, AI agents can operate 24/7, providing reliable support and an around-the-clock presence, filling critical roles like patient observation, data integration, and record management. This opens the door to quicker response times, improved patient outcomes, and better care.

Some will consider using AI to replace work that humans once did; founders should think about using AI to do work that humans never did, due to time constraints or other limitations. 

For example, large companies usually conduct annual sales team alignment exercises, mapping salespeople to specific territories. The effort ensures salespeople are working efficiently, but the process can be labor-intensive and expensive, especially if it involves an outside consultant. An AI agent handling sales team alignment could run the process much more frequently–say, quarterly or even monthly–and pull in many more data sources, including market trends, customer feedback, or real-time reports from a CRM. This means an AI agent could redirect a sales team to not only pivot faster to pursue more strategic sales targets–the numbers may show that California is a richer territory than Texas–but also could change their incentives dynamically and in real time, to go after certain types of accounts. 

Take another sector, asset management. Wealth managers often send personalized gifts to ultra-high-net-worth individuals, but they don’t send gifts to all clients because they simply don’t have the time. Now imagine an AI agent built for client profiling that could build rich profiles based on analyzing client preferences, past interactions, and even subtle cues from social media. The agent could identify and send personalized, appropriate gifts to every client on the list. Similarly, a marketing team might use an AI agent to scan, summarize, and synthesize calls, listening for signals about clients’ personal lives or interests. When the agent discovers that a client has recently had a child, it orders up a teddy bear; for a wine lover, a 2015 Château Margaux. 

Stay ahead of pricing and business model shifts

In our April essay, we touched on the ways in which service-as-software shifts the software business model—mainly from seat-based pricing to outcome-based pricing. Rather than charging customers per seat based on the number of sales-development representatives (SDRs) and account executives (AEs) on a platform, software vendors instead charge based on the number of qualified opportunities or signed customers their software delivers.

The switch to outcome-based AI pricing is already happening. Intercom made a splash with pricing for their AI support chatbot, Fin. The product costs $0.99 per successful resolution—directly aligning the price paid with the value received. Intercom counts a resolution as either (a) the customer confirms the answer is satisfactory or (b) the customer exits the conversation without escalating to a human. The AI chatbot is an add-on to the company’s core subscription offerings, which are priced on a per-seat basis.

The shift has other business model implications too. Take the observability sector. The industry still grapples with data overload, skyrocketing costs, and a critical shortage of skilled personnel. Incumbent company pricing models (think Datadog or Splunk) are wildly mismatched to customer needs, charging per GB for log management, for example, or ratcheting up costs with infrastructure size and data volume. 

An AI Site Reliability Engineer (SRE) working on an outcome-based model would not only interact via natural language, autonomously resolving issues and preventing incidents, but would also slash Mean Time to Resolution (MTTR), freeing up human SREs for more strategic work. The AI SRE vendor would be paid for improvements in MTTR and system uptime, shifting value to the tangible improvements that AI delivers.

The leap from automation to autonomous intelligence

The AI-led transformation of services into software marks a radical change in how we think about work. Repetitive, rule-based tasks were the low-hanging fruit; these business processes have long been ripe for disruption. Going forward, what makes a System of Agents so effective at automating services is its ability to grasp context, apply reasoning, and generally function as a team of experts, each contributing unique knowledge and abilities. Software is no longer merely assisting humans. It is acting as an autonomous worker, capable of understanding and evolving beyond human limitations. Still, the most transformative applications of AI won’t be those that simply replace human labor; they will collaborate seamlessly with human teams, unlocking new, previously unimaginable categories of work. 

If you’re building with agents, we’d love to hear from you. Email: jchen@foundationcap.com and jgupta@foundationcap.com


Published on October 31, 2024.
Written by Foundation Capital

Related Stories