AI won't automate your way out of the software mess — Benedict Evans

Benedict Evans takes a sober look at the enterprise AI hype cycle, arguing that the dream of AI sweeping away the tangled mess of corporate software ignores how tools actually get adopted. His central claim: most people aren't tool-builders, and the bottleneck isn't writing code—it's recognizing the problem and driving organizational change.

Evans walks through the typical large company's software estate: hundreds of systems of record, vertical SaaS apps, and countless improvised workflows running on Excel, email, and shared folders. The temptation is to think AI can replace all of it with dynamic, generative software that just does the task. But he counters that most workers don't think in terms of automating their jobs—they focus on their cases, clients, and craft.

The 'forward-deployed engineer' concept gets a nod: someone who can walk into a law firm or architecture office and spot the automation opportunities that domain experts miss. But Evans argues that even for tool-builders, the hard part isn't building—it's knowing what to build. Many successful software products came after half a dozen failed attempts that misidentified the problem.

Then there's the adoption problem. A great idea for reworking accounts payable touches 500 people across five departments and three systems of record. You can't just build a tool; you need an 18-month sales cycle and executive buy-in. Evans frames software adoption as a spectrum from institutionalized (SAP, Workday) to improvised (Excel, email, chatbots). AI, he says, doesn't change this dynamic—it just adds a new freeform substrate alongside Excel and email.

He draws parallels to past technology shifts: giving everyone a PC and Lotus 123 in 1983 didn't transform invoice processing; giving everyone a browser in 1997 didn't rebuild supply chains. The same pattern is playing out with enterprise AI—Copilot licenses are deployed, but usage is concentrated in a small group, and the rest of the company isn't finding ways to integrate it into their workflows.

The piece concludes with a critique of the pilot-based approach to AI adoption. Pilots are fine—about half work, as expected—but CEOs are asking: we have hundreds of workflows, and we've done five pilots. That doesn't scale. Evans suggests the answer isn't more pilots but a deeper understanding of how software becomes embedded in organizations over time.