Most companies chasing an AI rollout are optimizing the wrong variable, according to Benedict Evans. Writing on his own site, Evans argues that generating software has become nearly frictionless: a model can build a tool in minutes that once took an engineer an hour, no code required. The harder problem, he says, is figuring out which task deserves that tool in the first place.
Evans’s argument rests on a distinction between two kinds of difficulty. Building the automation is now cheap. Recognizing that a workflow is broken, and that it can be redefined rather than merely accelerated, is not. He notes that most employees, a great trial lawyer or an enterprise salesperson, spend their attention on their casework or their clients, not on imagining what dedicated software could replace it. The unmet need can sit in plain view for years before anyone names it.
That is why, per Evans, Silicon Valley invented the forward deployed engineer: someone who can walk through an architecture practice or a legal office and spot the automatable task that the people doing the work never noticed. He extends this to a broader claim about company software: the giant systems like SAP and Workday sit at one end of a spectrum, hundreds of ad hoc spreadsheets and email chains sit at the other, and tasks migrate between the two poles as they become important enough to formalize or specific enough to escape formal tooling. Evans compares this cycle to the SaaS wave, which multiplied the number of applications a company ran without eliminating the freeform layer underneath it.
His central historical parallel is instructive on its own. Evans likens handing every employee a chatbot to handing every employee a PC and Lotus 1-2-3 in 1983, or a browser in 1997: necessary, but not itself the transformation. In each earlier cycle, ownership of new hardware or software did not translate into rebuilt invoice processing or a rebuilt supply chain. That took a separate, slower process of deciding what to change and rewiring the organization around it. Evans reports that roughly half of current enterprise AI pilots work, which he calls a normal pilot success rate rather than evidence the technology has failed.
The practical test this raises for any operator running a rollout is simple: is the constraint model capability, or is it the unglamorous work of choosing which task to hand over and redesigning the process around that choice. Evans’s framing suggests most companies are stuck on the second question while budgeting as if they were still solving the first. A team that has issued licenses for Claude or Copilot and stopped there has completed the easy step and skipped the one that actually changes output.
Evans expects the pattern to repeat rather than resolve. AI will expand existing vertical software and spawn new vertical applications, with the payoff determined by how each company chooses to adapt rather than by which model it licenses. For an operator, that means the next ninety days are better spent auditing three to five specific workflows for hidden, redefinable tasks than waiting for a stronger model to make the decision automatically.
Benedict Evans made this argument on his own site, ben-evans.com, on September 3, 2026.
