Software Eats Services

The thesis

"AI will enable enterprise software to eat services." Robert F. Smith states this twice in the same interview, the closest thing in it to an explicit thesis statement.¹ The claim is a deliberate sequel to the older idea that software eats the world, and the substitution matters: software ate the tooling budget, replacing the tools a worker used, while agents come for the labor budget, replacing the underlying task itself. Smith is precise about what services means in this context, not a software company's own professional-services revenue line, but the people and dynamics performing a task that agents could carry out more efficiently, at a higher rate and higher frequency.

The scale of the opportunity follows from the substitution. Enterprise software has traditionally sold against a budget an order of magnitude smaller than the payroll of the function it serves. A vendor whose agents can perform the tasks inside a workflow it already owns is no longer selling against the software line item at all, it is effectively selling against the headcount line, which is why Smith expects "hyperaccelerated growth in these markets."¹

Why the vendor, and not the services firm

The load-bearing assumption is positional rather than technical. Performing a services task with an agent requires three things at once: the workflow the task happens inside, the data the task operates on, and write access to actually act. An enterprise software vendor that already owns a customer's workflow has all three by construction. A services firm has the people but neither the workflow nor the data infrastructure. A frontier AI lab has raw intelligence but neither of the other two. So, in Smith's framing, the vendor already sitting inside the workflow "can deliver agentic solutions to an industry that no one else can."¹ This is also why the same interview can hold both a bullish and a bearish view of software without contradiction: vendors that genuinely own a workflow are positioned to eat the services layer built around it, while vendors that do not own anything distinctive risk being eaten themselves.

Three categories of enterprise software

Smith sorts enterprise software into three categories to make the argument concrete. The first is software that can become agentic, where agents perform workflow tasks at a higher frequency and higher precision than a human worker using a tool, which is where the eating of services actually happens. The second follows what he calls a new margin standard, in which applying generative AI across product development, sales, service delivery, and back-office functions can lift the achievable combination of growth and margin well beyond the older rule of thumb used to judge software companies, in some cases nearly doubling margins. The third category is software that, in his words, has no right to exist: products that simply repackage data and information already available in the public marketplace and resell it. On this third category Smith concedes the bear case entirely and without hedging: it will be eaten, no question about it. The dividing line he repeats twice between the first two categories and the third is a single phrase, whether a company has sovereignty and dominion over its own workflows and data sets.¹

Why it matters

The claim functions as the constructive half of a broader argument about AI and existing enterprise software business models, which is more often framed defensively, asking which vendors survive being replaced, than offensively, asking where vendors that survive can actually go on the attack. It also sharpens what claims of a larger addressable market actually require: the generic version of the pitch, that a company can now address the budget currently spent on outside services, is closer to a slogan, while Smith's version names the precondition explicitly, ownership of the underlying workflow, without which the same claim is closer to a wish.

Tensions

Services margins are not simply transferred to the software vendor once agents make a task cheap, they are competed away, and whether the vendor captures the freed-up budget or simply passes savings back to the customer depends entirely on whether the vendor remains the only party who can perform the task, which returns the whole argument to the workflow-ownership precondition. Regulated professions also resist for reasons that have little to do with capability, since licensure and professional liability determine who is legally allowed to sign off on a piece of work regardless of who or what actually produced it. And the claim is made by the owner of the vendors it describes, in a fundraising-adjacent setting, without a disclosed revenue figure attached to any specific portfolio company.