AI, tools and transformation
Benedict Evans argues the popular hope that AI will sweep away the sprawling mess of enterprise software misunderstands where software comes from and how companies change. The typical big company has hundreds or thousands of pieces of software — SAP and Workday at one end, a 10-megabyte spreadsheet running a department at the other — and yet is full of repetitive tasks. The tempting vision is that with AI you can now build the tool in five minutes, or just have the model do the task itself, and software becomes dynamic, generative, and spontaneous.
Evans' counter has three moves. First, most people aren't tool-builders and don't see their own job as a workflow to be automated; the task sits in plain sight and its owner doesn't notice it (which is the whole rationale for the "forward-deployed engineer"). Second, the hard part was never writing code — it's knowing you need a tool and what it should do; most successful software companies had half a dozen failed predecessors that didn't find quite the right problem. Third, even a working tool has to be adopted by 50–500 people across departments, systems of record, and regulatory regimes, which is a purchase and an 18-month sales process, not a five-minute build.
The essay's central frame is a spectrum from institutionalised to improvised. Tasks done in dedicated tools (SAP, Carta, Rippling) are institutionalised; edge cases live in improvised freeform substrates (Excel, email, Tableau, screenshots, conference calls). When a task becomes repetitive, cross-departmental, and risky, the company institutionalises it — "you pave the desire path and pay someone to set it in stone." This is why the company has hundreds of apps, and the SaaS era was one long cycle of bundling and unbundling (Carta is a $4bn company that manages one spreadsheet for your CFO).
AI rolls across this without changing the question. It expands the existing apps and adds the chatbot as a new freeform substrate that takes tasks from apps and loses tasks to them. It moves the thresholds — the small firm stays on Google Sheets longer; a PwC team still improvises around Workday. The enterprise evidence bears this out: give everyone Copilot and a few people use it heavily, some use it twice a week, most not at all — the same problem you'd have had giving everyone Lotus 123 in 1983 or a web browser in 1997. Pilots don't scale: the CIO runs five or ten pilots while the CEO asks why hundreds of workflows aren't transformed.
Evans closes with the three questions every company must answer about a generational technology — buy/build/deploy, how far operations change, and whether it's an existential threat — none of them answered by "giving everyone Claude for X" — and notes the simpler framing: every technology is first used to do existing work faster, then to make things that weren't possible before, and the latter is what actually mattered.