A rational conversation on where AI is actually going | Benedict Evans

Benedict Evans places AI alongside the internet and mobile as a foundational platform shift, while resisting comparisons that make it larger than either. His useful comparison is 1997: the technology is exciting, adoption and capability are uneven, many important products have not been built, and even experts cannot yet explain how the eventual system will work. That uncertainty is not evidence that the shift is small; it is part of what makes the present difficult to interpret.

The early effects are distributed unevenly. Software developers may experience AI as an immediate transformation, much as an accountant experienced the first spreadsheets, while lawyers, journalists, and other professionals may see an impressive tool without yet knowing where it fits. Evans connects this uneven adoption to a jagged frontier: users must discover where the system works, where it fails, and whether its successful outputs are reliable. The result is a wide gap between technical enthusiasts and people who use AI only occasionally or not at all.

That framing also explains why Evans rejects confident timelines for a sudden moment when everything becomes different. Software is already changing, but adoption across law firms and other organizations involves trust, workflows, hiring, and implementation. His example of professional-services firms shows that the hard part is not simply producing a model output. Companies must decide how to redesign internal processes, connect systems, train staff, and manage the risk of errors. AI can therefore create implementation work even where it eventually reduces some routine work.

The interview presents AI as both a software-development shift and an organizational problem. A developer may see an immediate change in what can be built, but a law firm still has to decide how to use the technology without submitting hallucinated work or misjudging future hiring. Evans's spreadsheet analogy captures the difference: an accountant could immediately recognize the power of a tool that compresses a week of calculations into seconds, while a lawyer or journalist might see no obvious connection to their own work.

The unexpected growth of consulting and forward-deployed engineering follows from that implementation gap. Companies do not keep spare teams waiting to redesign stores, analyze churn, build new products, or rework every internal process. They therefore hire professional-services firms, agencies, or engineers who can map existing workflows, connect systems, create new ones, and train employees. AI may automate pieces of this work, but deploying it across a real organization is itself a substantial project.