The systems of record are not going anywhere. SAP spent thirty years installing itself into the Fortune 500, Salesforce did the same to revenue operations, and ServiceNow and NetSuite hold the IT and service-management layer and the upper end of the mid-market ERP space between them. These are the systems that contain the actual data of the business, and they are slow to install, expensive to migrate, and almost impossible to replace. The a16z piece "Why the World Still Runs on SAP" puts the cost of an ECC-to-S/4HANA migration at $700M, three years, and a fifty-person Accenture team. Lidl scrapped its SAP transition after spending half a billion dollars, and roughly seventy percent of digital transformations fail their objectives, by McKinsey's own estimate.
The records persist. What changes is what sits on top of them.
A new layer is forming above the systems of record
A new layer is forming, and it does not have a settled name yet. a16z calls it the system of action: the layer that reads from systems of record, does work, writes back, and stays accountable for the result. Foundation Capital calls it the system of agents, plural, because the work usually requires several specialized agents working together. Bessemer calls it AI-enabled services, distinguishing it from copilots (which augment a human worker) and pure agents (which replace one). Whatever the name settles into, the layer is real, and it is where the next decade of enterprise software is being built.
The layer is not a copilot. A copilot lives inside an existing application and helps the user of that application do their job faster, while the system of action lives between applications and does the job in place of the user. The two sit at different layers of the stack and are addressed at different parts of the customer's budget: copilots come out of the SaaS budget, and systems of action come out of the labor budget, which is roughly an order of magnitude larger.
Why the system-of-action layer is forming now
The technology to build this layer has been around in pieces for a long time: process mining vendors (Celonis, Apromore) have been mapping enterprise workflows since the early 2010s, and RPA vendors (UiPath, Automation Anywhere) have been automating those workflows for nearly as long. What neither could do, and what has been the structural bottleneck for a decade, is handle the unstructured-input judgment work that constitutes most of the actual work in a business. Foundation models can handle it. Wornow, Narayan, Ré and colleagues at Stanford put the productivity opportunity at four trillion dollars per year in a 2024 paper that named foundation models as the missing piece process mining and RPA were always waiting for, and the pipeline (process mining, then workflow understanding, then execution) is now coherent in a way it has never been.
The vendors of the previous era have noticed. ServiceNow partnered with Celonis in 2021, acquired the task-mining company UltimateSuite in late 2023, and now embeds process mining and agentic automation directly into the platform with its Zurich release. The category convergence is happening at the platform level, and Constellation Research's coverage names the shift directly: you cannot deploy agents reliably without process intelligence underneath them, which is why the formerly separate diagnostic discipline of process mining is being absorbed into the deployment substrate of the new agentic platforms.
What the winning companies will look like
The companies that win the system-of-action layer will look more like services firms staffed by engineers than like the SaaS companies that won the previous one. Because the work requires understanding the specific stack and the specific workflows, delivery teams sit embedded in the customer's environment; because the deliverable is measurable and the billing runs against the labor budget, pricing follows outcomes. Vertical depth compounds as agents trained on one customer's workflows carry what they learned to the next, and outcome data aggregated across customers in the same operational pattern becomes a network effect.
Bessemer's vertical AI thesis predicts vertical AI's market capitalization will reach at least ten times that of legacy vertical SaaS, whose top twenty US public companies are worth roughly $300 billion combined. The math behind the multiple is mechanical: vertical SaaS extracted a margin from a software budget, vertical AI extracts one from a labor budget, and the labor budget is an order of magnitude larger.
Why this is not a repeat of the RPA hype cycle
A reasonable counter is that all of this resembles the RPA hype cycle of 2018, which plateaued for exactly the reasons most agent skeptics raise now: edge cases multiply, integrations break, governance is hard, the demos work and the production deployments do not. The structural difference is what foundation models can do that RPA never could. RPA could only handle deterministic rule-following on structured inputs, which meant it could only automate the parts of the workflow where the rules were already explicit and the data was already clean. Foundation models can handle unstructured inputs with judgment, which is most of the actual work. The Wornow paper is the formal version of this argument, and it is worth reading in full if you are inclined to treat the agent wave as another RPA wave.
The platform layer has shifted before: the operating system anchored the 1980s and 1990s, and the system of record anchored the two decades that followed. The system of action is the shift now underway, and it is underway because the bottleneck that stalled the last automation wave is gone.
The system of action does not require migration
The system of action does not require the customer to migrate; that is the point of the layer. It runs on top of the existing records, the existing CRM, the existing ERP, and the customer keeps everything they have spent thirty years installing. The new layer does the work that used to require a person sitting between the systems: the manual reconciliation, the data entry, the exception handling, the reporting.
a16z's line for what this looks like operationally is the cleanest: "the bridge becomes the highway." The integration that used to be a temporary connection between two systems becomes the place where the work happens, with the system of record holding the data and the system of action moving it.